Research

Cost-optimal planning is at the heart of planning research, with many existing planners that produce provably optimal solutions. While some applications pose additional restric- tions, such as producing shortest (in the number of actions) among the cost-optimal plans, standard cost-optimal planning does not provide such a guarantee. We discuss two possible approaches to produce provably the shortest among the cost- optimal plans, one corresponding to an instantiation of cost- algebraic A∗ , the other based on a cost transformation. We formally prove that the new cost-transformation method in- deed produces the shortest among the cost-optimal plans and empirically compare the performance of the approaches in different configurations.

The aim of the AIPlan4EU project is to make state-of-the-art planning technology easily accessible for real-world applications and to integrate planning as a service on the European AI-on-demand platform. At the very heart of the project is the development of a unified planning framework that permits to programmatically specify planning tasks and to interact with different planning engines through a common interface. The demo presents the AIPlan4EU vision for the unified planning framework and the current state of the development.

Correlation complexity is a measure of a planning task indicating how hard it is. The introducing work provides sufficient criteria to detect a correlation complexity of 2 on a planning task. It also introduced an example of a planning task with correlation complexity 3. In our work, we introduce a criterion to detect an arbitrary correlation complexity and extend the mentioned example to show with the new criterion that planning tasks with arbitrary correlation complexity exist.

We introduce transition-counting constraints as a principled tool to formalize constraints that must hold in every solution of a transition system. We then show how to obtain transition landmark constraints from abstraction cuts. Transition landmarks dominate operator landmarks in theory but require solving a linear program that is prohibitively large in practice. We compare different constraints that project away transition-counting variables and then further relax the constraint. For one important special case, we provide a lossless projection. We finally discuss efficient data structures to derive cuts from abstractions and store them in a way that avoids repeated computation in every state. We compare the resulting heuristics both theoretically and on benchmarks from the international planning competition.

The merge-and-shrink framework is a powerful tool for computing abstraction heuristics for optimal classical planning. Merging is one of its name-giving transformations. It entails computing the product of two factors of a factored transition system. To decide which two factors to merge, the framework uses a merge strategy. While there exist many merge strategies, it is generally unclear what constitutes a strong merge strategy, and a previous analysis shows that there is still lots of room for improvement with existing merge strategies. In this paper, we devise a new scoring function for score-based merge strategies based on answering the question whether merging two factors has any benefits over computing saturated cost partitioning heuristics over the factors instead. Our experimental evaluation shows that our new merge strategy achieves state-of-the-art performance on IPC benchmarks.

Classical planning considers a given task and searches for a plan to solve it. Some tasks are harder to solve than others. We can measure the “hardness” of a task with the novelty width and the correlation complexity. In this work, we compare these measures. Additionally, we introduce the river measure, a new measure that is based on potential heuristics and therefore similar to the correlation complexity but also comparable to the novelty width. We show that the river measure is upper bounded by the correlation complexity and by the novelty width +1. Furthermore, we show that we can convert a planning task with a polynomial blowup of the task size to ensure that a heuristic of dimension 2 exists that gives rise to backtrack-free search.

The merge-and-shrink framework is a powerful tool to construct state space abstractions based on factored representations. One of its core applications in classical planning is the construction of admissible abstraction heuristics. In this paper, we develop a compositional theory of merge-and-shrink in the context of probabilistic planning, focusing on stochastic shortest path problems (SSPs). As the basis for this development, we contribute a novel factored state space model for SSPs. We show how general transformations, including abstractions, can be formulated on this model to derive admissible and/or perfect heuristics. To formalize the merge-and-shrink framework for SSPs, we transfer the fundamental merge-and-shrink transformations from the classical setting: shrinking, merging, and label reduction. We analyze the formal properties of these transformations in detail and show how the conditions under which shrinking and label reduction lead to perfect heuristics can be extended to the SSP setting.

Abstraction heuristics are a state-of-the-art technique to solve classical planning problems optimally. A common approach is to precompute many small abstractions and combine them admissibly using cost partitioning. Recent work has shown that this approach does not work out well when using such heuristics for decoupled state space search, where search nodes represent potentially large sets of states. This is due to the fact that admissibly combining the estimates of several heuristics without sacrificing accuracy is NP-hard for decoupled states. In this work we propose to use a single large abstraction instead. We focus on merge-and-shrink and symbolic pattern database heuristics, which are designed to produce such abstractions. For these heuristics, we prove that the evaluation of decoupled states is NP-hard in general, but we also identify conditions under which it is polynomial. We introduce algorithms for both the general and the polynomial case. Our experimental evaluation shows that single large abstraction heuristics lead to strong performance when the heuristic evaluation is polynomial.

Abstraction heuristics are the state of the art in optimal classical planning as heuristic search. A popular method for computing abstractions is the counterexample-guided abstraction refinement (CEGAR) principle, which has successfully been used for projections, which are the abstractions underlying pattern databases, and Cartesian abstractions. While projections are simple and fast to compute, Cartesian abstractions subsume projections and hence allow more finegrained abstractions, however at the expense of efficiency. Domain abstractions are a third class of abstractions between projections and Cartesian abstractions in terms of generality. Yet, to the best of our knowledge, they are only briefly considered in the planning literature but have not been used for computing heuristics yet. We aim to close this gap and compute domain abstractions by using the CEGAR principle. Our empirical results show that domain abstractions compare favorably against projections and Cartesian abstractions.

The computation of high-quality landmarks and orderings for heuristic state-space search is often prohibitively expensive to be performed in every generated state. Computing information only for the initial state and progressing it from every state to its successors is a successful alternative, exploited for example in classical planning by the LAMA planner. We propose a general framework for using landmarks in any kind of best-first search. Its core component, the progression function, uses orderings and search history to determine which landmarks must still be achieved. We show that the progression function that is used in LAMA infers invalid information in the presence of reasonable orderings. We define a sound progression function that allows to exploit reasonable orderings in cost-optimal planning and show empirically that our new progression function is beneficial both in satisficing and optimal planning.

Previous work introduced the concept of progress states. Once greedy best-first search expanded a progress state, it expands only states with lower heuristic values. Current methods identify progress states for a single task and after a solution for the task is found. We introduce a novel approach which learns for a classical planning domain a description logic formula to characterize all its progress states. This formula can be used during future searches. We present a first use case which exploits the learned formulas: During search we break ties in favor of progress states. This often significantly reduces the search effort.

Neural networks (NN) are increasingly investigated in AI Planning, and are used successfully to learn heuristic functions. NNs commonly not only predict a value, but also output a confidence in this prediction. From the perspective of heuristic search with NN heuristics, it is a natural idea to take this into account, e.g. falling back to a standard heuristic where confidence is low. We contribute an empirical study of this idea. We design search methods which prune nodes, or switch between search queues, based on the confidence of NNs. We furthermore explore the possibility of out-of-distribution (OOD) training, which tries to reduce the overconfidence of NNs on inputs different to the training distribution. In experiments on IPC benchmarks, we find that our search methods improve coverage over standard methods, and that OOD training has the desired effect in terms of prediction accuracy and confidence, though its impact on search seems marginal.

Abstraction heuristics are the state of the art in optimal classical planning as heuristic search. Despite their success for explicit-state search, though, abstraction heuristics are not available for decoupled state-space search, an orthogonal reduction technique that can lead to exponential savings by decomposing planning tasks. In this paper, we show how to compute pattern database (PDB) heuristics for decoupled states. The main challenge lies in how to additively employ multiple patterns, which is crucial for strong search guidance of the heuristics. We show that in the general case, for arbitrary collections of PDBs, computing the heuristic for a decoupled state is exponential in the number of leaf components of decoupled search. We derive several variants of decoupled PDB heuristics that allow to additively combine PDBs avoiding this blow-up and evaluate them empirically.

Since its invention in 1974, the Rubik’s Cube puzzle fascinates people of all ages. Its rules are simple: the player gets a scrambled cube and rotates the six faces until each face contains only stickers of one color. Nevertheless, finding a short sequence of rotations to solve the cube is hard. We present the first model of Rubik’s Cube for general problem solvers. To obtain a concise model, we require conditional effects. Furthermore, we extend counterexample-guided Cartesian abstraction refinement (CEGAR) to support factored effect tasks, a class of planning tasks with a specific kind of conditional effects which includes Rubik’s Cube. Finally, we evaluate how newer types of abstraction heuristics compare against pattern database (PDB) heuristics, the state-of-the-art for solving Rubik’s Cube. We find that PDBs still outperform the more general Cartesian and merge-and-shrink abstractions. However, in contrast to PDBs, Cartesian abstractions yield perfect heuristics up to a certain problem difficulty. These findings raise interesting questions for future research.

In classical planning, cost partitioning is a powerful method which allows to combine multiple admissible heuristics while retaining an admissible bound. In this paper, we extend the theory of cost partitioning to probabilistic planning by generalizing from deterministic transition systems to stochastic shortest path problems (SSPs). We show that fundamental results related to cost partitioning still hold in our extended theory. We also investigate how to optimally partition costs for a large class of abstraction heuristics for SSPs. Lastly, we analyze occupation measure heuristics for SSPs as well as the theory of approximate linear programming for reward-oriented Markov decision processes. All of these fit our framework and can be seen as cost-partitioned heuristics.

While the unsolvability IPC sparked a multitude of planners proficient in detecting unsolvable planning tasks, there are gaps where concise unsolvability arguments are known but no existing planner can capture them without prohibitive computational effort. One such example is the sliding tiles puzzle, where solvability can be decided in polynomial time with a parity argument. We introduce separating functions, which can prove that one state is unreachable from another, and show under what conditions a potential function over any nonzero ring is a separating function. We prove that we can compactly encode these conditions for potential functions over features that are pairs, and show in which cases we can efficiently synthesize functions satisfying these conditions. We experimentally evaluate a domain-independent algorithm that successfully synthesizes such separating functions from PDDL representations of the sliding tiles puzzle, the Lights Out puzzle, and Peg Solitaire.

Potential functions are a general class of heuristics for classical planning. For satisficing planning, previous work suggested the use of descending and dead-end avoiding (DDA) potential heuristics, which solve planning tasks by backtrack-free search. In this work we study the complexity of devising DDA potential heuristics for classical planning tasks. We show that verifying or synthesizing DDA potential heuristics is PSPACE-complete, but suitable modifications of the DDA properties reduce the complexity of these problems to the first and second level of the polynomial hierarchy. We also discuss the implications of our results for other forms of heuristic synthesis in classical planning.

Abstraction heuristics are the state of the art in optimal classical planning as heuristic search. Despite their success for explicit-state search, though, abstraction heuristics are not available for decoupled state-space search, an orthogonal reduction technique that can lead to exponential savings by decomposing planning tasks. In this paper, we show how to compute pattern database (PDB) heuristics for decoupled states. The main challenge lies in how to additively employ multiple patterns, which is crucial for strong search guidance of the heuristics. We show that in the general case, for arbitrary collections of PDBs, computing the heuristic for a decoupled state is exponential in the number of leaf components of decoupled search. We derive several variants of decoupled PDB heuristics that allow to additively combine PDBs avoiding this blow-up and evaluate them empirically.

Stubborn sets are a pruning technique for state-space search which is well established in optimal classical planning. In this paper, we show that weak stubborn sets introduced in recent work in planning are actually not weak stubborn sets in Valmari's original sense. Based on this finding, we introduce weak stubborn sets in the original sense for planning by providing a generalized definition analogously to generalized strong stubborn sets in previous work. We discuss the relationship of strong, weak and the previously called weak stubborn sets, thus providing a further step in getting an overall picture of the stubborn set approach in planning.

The merge-and-shrink framework has been introduced as a general approach for defining abstractions of large state spaces arising in domain-independent planning and related areas. The distinguishing characteristic of the merge-and-shrink approach is that it operates directly on the factored representation of state spaces, repeatedly modifying this representation through transformations such as shrinking (abstracting a factor of the representation), merging (combining two factors), label reduction (abstracting the way in which different factors interact), and pruning (removing states or transitions of a factor).

We provide a novel view of the merge-and-shrink framework as a "toolbox" or "algebra" of transformations on factored transition systems, with the construction of abstractions as only one possible application. For each transformation, we study desirable properties such as conservativeness (overapproximating the original transition system), inducedness (absence of spurious states and transitions), and refinability (reconstruction of paths in the original transition system from the transformed one). We provide the first complete characterizations of the conditions under which these desirable properties can be achieved. We also provide the first full formal account of factored mappings, the mechanism used within the merge-and-shrink framework to establish the relationship between states in the original and transformed factored transition system.

Unlike earlier attempts to develop a theory for merge-and-shrink, our approach is fully compositional: the properties of a sequence of transformations can be entirely understood by the properties of the individual transformations involved. This aspect is key to the use of merge-and-shrink as a general toolbox for transforming factored transition systems. New transformations can easily be added to our theory, with compositionality taking care of the seamless integration with the existing components. Similarly, new properties of transformations can be integrated into the theory by showing their compositionality and studying under which conditions they are satisfied by the building blocks of merge-and-shrink.

Cost partitioning admissibly combines the information from multiple heuristics for optimal state-space search. One of the strongest cost partitioning algorithms is saturated cost partitioning. It considers the heuristics in sequence and assigns to each heuristic the minimal fraction of the remaining costs that are needed for preserving all heuristic estimates. Saturated cost partitioning has recently been generalized in two directions: first, by allowing to use different costs for the transitions induced by the same operator, and second, by preserving the heuristic estimates for only a subset of states. In this work, we unify these two generalizations and show that the resulting subset-saturated transition cost partitioning algorithm usually yields stronger heuristics than the two generalizations by themselves.

Cost partitioning is a general method for admissibly summing up heuristic estimates for optimal state-space search. Most cost partitioning algorithms can optimize the resulting cost-partitioned heuristic for a specific state. Since computing a new cost-partitioned heuristic for each evaluated state is usually too expensive in practice, the strongest planners based on cost partitioning over abstraction heuristics precompute a set of cost-partitioned heuristics before the search and maximize over their estimates during the search. This makes state evaluations very fast, but since there is no better termination criterion than a time limit, it requires a long precomputation phase, even for the simplest planning tasks. A prototypical example for this is the Scorpion planner which computes saturated cost partitionings over abstraction heuristics offline before the search. Using Scorpion as a case study, we show that by incrementally extending the set of cost-partitioned heuristics online during the search, we drastically speed up the planning process and even often solve more tasks.

Landmarks of a planning task denote properties that must be satisfied by all plans. Existing landmark heuristics exploit that each landmark must be achieved at least once. However, if the orderings between the landmarks induce cyclic dependencies, one of the landmarks in each cycle must be achieved an additional time. We propose two novel heuristics for cost-optimal planning that consider cyclic dependencies between landmarks in addition to the cost for achieving all landmarks once. We show that our heuristics dominate the minimum hitting set solution over any set of landmarks as well as h + if all delete-relaxation landmarks are considered. An experimental evaluation on benchmarks from the International Planning Competition shows that exploiting cyclic dependencies can lead to improved heuristics.

Optimal cost partitioning can produce high quality heuristic estimates even from small abstractions. It can be computed with a linear program (LP) but the size of this LP often makes this impractical. Recent work used Lagrangian decomposition to speed up the computation. Here we use a different decomposition technique called Dantzig-Wolfe decomposition to tackle the problem. This gives new insights into optimal cost partitioning and has several advantages over Lagrangian decomposition: our method detects when a cost partition is optimal; it can deal with general cost functions; and it does not consider abstractions in the linear program that do not contribute to the heuristic value. We also show the advantage of the method empirically and investigate several improvements that are useful for all cost partitioning methods.

It is usually prohibitively expensive to recompute landmarks in every state of a heuristic search algorithm that bases its heuristic estimate on landmarks. Successful planning systems like LAMA compute landmarks just for the initial state instead and only keep track for each state encountered during search which landmarks of the initial state are accepted or required again. The LM-A* algorithm describes a similar idea for cost-optimal heuristic search. However, it lacks support for reasonable orderings, which have recently been shown to be informative when cyclic dependencies between landmarks are exploited in the heuristic computation. We propose the novel LM-BFS framework for best-first heuristic search with landmark-based heuristics which allows to describe a variety of algorithms in terms of five components: landmark generation, landmark graph progression, landmark graph merging, landmark graph extension, and exploration strategy. We use LM-BFS to show that considering reasonable orderings in the landmark graph progression component leads to flawed landmark graphs in LAMA , and propose a novel landmark graph extension that captures the same information flawlessly. We observe empirically that considering reasonable orderings in the extension rather than the progression component has a positive effect both for LAMA and for cost-optimal planning.

Saturated cost partitioning and post-hoc optimization are two powerful cost partitioning algorithms for optimal classical planning. The main idea of saturated cost partitioning is to give each considered heuristic only the fraction of remaining operator costs that it needs to prove its estimates. We show how to apply this idea to post-hoc optimization and obtain a heuristic that dominates the original both in theory and on the IPC benchmarks.

Optimal cost partitioning of classical planning heuristics has been shown to lead to excellent heuristic values but is often prohibitively expensive to compute. We analyze the application of Lagrangian decomposition, a classical tool in mathematical programming, to cost partitioning of operator-counting heuristics. This allows us to view the computation as an iterative process that can be seeded with any cost partitioning and that improves over time. In the case of non-negative cost partitioning of abstraction heuristics the computation reduces to independent shortest path problems and does not require an LP solver.

Cost partitioning is a method for admissibly combining admissible heuristics. In this work, we extend this concept to merge-and-shrink (M&S) abstractions that may use labels that do not directly correspond to operators. We investigate how optimal and saturated cost partitioning (SCP) interact with M&S transformations and develop a method to compute SCPs during the computation of M&S. Experiments show that SCP significantly improves M&S on standard planning benchmarks.

Counterexample-guided Cartesian abstraction refinement has been shown to yield informative heuristics for optimal classical planning. The algorithm iteratively finds an abstract solution and uses it to decide how to refine the abstraction. Since the abstraction grows in each step, finding solutions is the main bottleneck of the refinement loop. We cast the refinements as an incremental search problem and show that this drastically reduces the time for computing abstractions.

Saturated cost partitioning is a general method for admissibly adding heuristic estimates for optimal state-space search. The algorithm strongly depends on the order in which it considers the heuristics. The strongest previous approach precomputes a set of diverse orders and the corresponding saturated cost partitionings before the search. This makes evaluating the overall heuristic very fast, but requires a long precomputation phase. By diversifying the set of orders online during the search we drastically speed up the planning process and even solve slightly more tasks.

The benchmarks from previous International Planning Competitions are commonly used to evaluate new planning algorithms. Since this set has grown organically over the years, it has several flaws: it contains duplicate tasks, unsolvable tasks, trivially solvable domains, and domains with modelling errors. Also, diverse domain sizes complicate aggregating results. Most importantly, however, the range of task difficulty is very small in many domains. We propose an automated method for creating benchmarks that solves these issues. To find a good scaling in difficulty, we automatically configure the parameters of benchmark domains. We show that the resulting benchmark set improves empirical comparisons by allowing to differentiate between planners more easily.

MDPs with factored action spaces, i.e. where actions are described as assignments to a set of action variables, allow reasoning over action variables instead of action states, yet most algorithms only consider a grounded action representation. This includes algorithms that are instantiations of the trial-based heuristic tree search (THTS) framework, such as AO* or UCT.

To be able to reason over factored action spaces, we propose a generalisation of THTS where nodes that branch over all applicable actions are replaced with subtrees that consist of nodes that represent the decision for a single action variable. We show that many THTS algorithms retain their theoretical properties under the generalised framework, and show how to approximate any state-action heuristic to a heuristic for partial action assignments. This allows to guide a UCT variant that is able to create exponentially fewer nodes than the same algorithm that considers ground actions. An empirical evaluation on the benchmark set of the probabilistic track of the latest International Planning Competition validates the benefits of the approach.

Stubborn sets are an optimality-preserving pruning technique for factored state-space search, for example in classical planning. Their applicability is limited by their computational overhead. We describe a new algorithm for computing stubborn sets that is based on the state variables of the state space, while previous algorithms are based on its actions. Typical factored state spaces tend to have far fewer state variables than actions, and therefore our new algorithm is much more efficient than the previous state of the art, making stubborn sets a viable technique in many cases where they previously were not.

Cost partitioning is a method for admissibly combining a set of admissible heuristic estimators by distributing operator costs among the heuristics. Computing an optimal cost partitioning, i.e., the operator cost distribution that maximizes the heuristic value, is often prohibitively expensive to compute. Saturated cost partitioning is an alternative that is much faster to compute and has been shown to yield high-quality heuristics. However, its greedy nature makes it highly susceptible to the order in which the heuristics are considered. We propose a greedy algorithm to generate orders and show how to use hill-climbing search to optimize a given order. Combining both techniques leads to significantly better heuristic estimates than using the best random order that is generated in the same time. Since there is often no single order that gives good guidance on the whole state space, we use the maximum of multiple orders as a heuristic that is significantly better informed than any single-order heuristic, especially when we actively search for a set of diverse orders.

The performance of domain-independent planning systems heavily depends on how the planning task has been modeled. This makes task reformulation an important tool to get rid of unnecessary complexity and increase the robustness of planners with respect to the model chosen by the user. In this paper, we represent tasks as factored transition systems (FTS), and use the merge-and-shrink (M&S) framework for task reformulation for optimal and satisficing planning. We prove that the flexibility of the underlying representation makes the M&S reformulation methods more powerful than the counterparts based on the more popular finite-domain representation. We adapt delete-relaxation and M&S heuristics to work on the FTS representation and evaluate the impact of our reformulation.

We tackle two long-standing problems related to re-expansions in heuristic search algorithms. For graph search, A* can require Ω(2^n) expansions, where n is the number of states within the final f bound. Existing algorithms that address this problem like B and B' improve this bound to Ω(n^2). For tree search, IDA* can also require Ω(n^2) expansions. We describe a new algorithmic framework that iteratively controls an expansion budget and solution cost limit, giving rise to new graph and tree search algorithms for which the number of expansions is O(n log C*), where C* is the optimal solution cost. Our experiments show that the new algorithms are robust in scenarios where existing algorithms fail. In the case of tree search, our new algorithms have no overhead over IDA* in scenarios to which IDA* is well suited and can therefore be recommended as a general replacement for IDA*.

Pattern databases are the foundation of some of the strongest admissible heuristics for optimal classical planning. Experiments showed that the most informative way of combining information from multiple pattern databases is to use saturated cost partitioning. Previous work selected patterns and computed saturated cost partitionings over the resulting pattern database heuristics in two separate steps. We introduce a new method that uses saturated cost partitioning to select patterns and show that it outperforms all existing pattern selection algorithms.

This paper investigates online stochastic planning for problems with large factored state and action spaces. One promising approach in recent work estimates the quality of applicable actions in the current state through aggregate simulation from the states they reach. This leads to significant speedup, compared to search over concrete states and actions, and suffices to guide decision making in cases where the performance of a random policy is informative of the quality of a state. The paper makes two significant improvements to this approach. The first, taking inspiration from lifted belief propagation, exploits the structure of the problem to derive a more compact computation graph for aggregate simulation. The second improvement replaces the random policy embedded in the computation graph with symbolic variables that are optimized simultaneously with the search for high quality actions. This expands the scope of the approach to problems that require deep search and where information is lost quickly with random steps. An empirical evaluation shows that these ideas significantly improve performance, leading to state of the art performance on hard planning problems.

Optimal cost partitioning of classical planning heuristics has been shown to lead to excellent heuristic values but is often prohibitively expensive to compute. Lagrangian decomposition and Lagrangian relaxation are classical tools in mathematical programming that apply to optimization problems with a special block structure. We analyze the application of Lagrangian decomposition to cost partitioning in the context of operator-counting heuristics and interpret Lagrangian multipliers as cost functions for the combined heuristics. This allows us to view the computation of an optimal cost partitioning as an iterative process that can be seeded with any cost partitioning and improves over time. We derive an any-time algorithm to compute an optimal non-negative cost partitioning of abstraction heuristics without involving an LP solver. In each iteration, the computation reduces to independent shortest path problems in all abstractions. Finally, we discuss the extension to general cost functions.

Potential heuristics are weighted functions over state features of a planning task. A recent study defines the complexity of a task as the minimum required feature complexity for a potential heuristic that makes a search backtrack-free. This gives an indication of how complex potential heuristics need to be to achieve good results in satisficing planning. However, these results do not directly transfer to optimal planning.

In this paper, we empirically study how complex potential heuristics must be to represent the perfect heuristic and how close to perfect heuristics can get with a limited number of features. We aim to identify the practical trade-offs between size, complexity and time for the quality of potential heuristics. Our results show that, even for simple planning tasks, finding perfect potential heuristics might be harder than expected.

The aim of classical planning is to minimize the summed cost of operators among those plans that achieve a fixed set of goals. Oversubscription planning (OSP), on the other hand, seeks to maximize the utility of the set of facts achieved by a plan, while keeping the cost of the plan at or below some specified bound. Here, we investigate the use of reformulations that yield planning problems with two separate cost functions, but no utilities, for solving OSP tasks. Such re-formulations have also been proposed in the context of net-benefit planning, where the planner tries to maximize the difference between the utility achieved and the cost of the plan. One of our reformulations is adapted directly from that setting, while the other is novel. In both cases, they allow for easy adaptation of existing classical planning heuristics to the OSP problem within a simple branch and bound search. We validate our approach using state of the art admissible heuristics in this framework, and report our results.

Cost partitioning is a method for admissibly adding multiple heuristics for state-space search. Saturated cost partitioning considers the given heuristics in sequence, assigning to each heuristic the minimum fraction of remaining costs that it needs to preserve its estimates for all states. We generalize saturated cost partitioning by allowing to preserve the heuristic values of only a subset of states and show that this often leads to stronger heuristics.

We transfer the notion of structural symmetries to lifted planning task representations, based on abstract structures which we define to model planning tasks. We show that symmetries are preserved by common grounding methods and we shed some light on the relation to previous symmetry concepts used in planning. Using a suitable graph representation of lifted tasks, our experimental analysis of common planning benchmarks reveals that symmetries occur in the lifted representation of many domains. Our work establishes the theoretical ground for exploiting symmetries beyond their previous scope, such as for faster grounding and mutex generation, as well as for state space transformations and reductions.

The International Planning Competition 2018 consisted of several tracks on classical, temporal and probabilistic plan- ning. In this paper, we focus on the discrete MDP track of the probabilistic portion of the competition.

We discuss the changes to the input language RDDL, which give rise to new challenges for planning systems, and ana- lyze each of the eight competition domains separately and highlight unique properties. We demonstrate flaws of the used evaluation criterion, the IPC score, and discuss the need for optimal upper bounds. An evaluation of the three top- performers, including their post-competition versions, and a brief analysis of their performance highlights the strengths and weaknesses of the individual approaches.

We argue that planner evaluation metrics should satisfy the independence of irrelevant alternatives criterion, i.e., the decision whether planner A is ranked higher or lower than planner B should be independent of planner C. We show that three metrics used in classical planning competitions do not necessarily satisfy this criterion and highlight alternative metrics that do so.

Pattern databases are the foundation of some of the strongest admissible heuristics for optimal classical planning. Experiments showed that the most informative way of combining information from multiple pattern databases is to use saturated cost partitioning. Previous work selected patterns and computed saturated cost partitionings over the resulting pattern database heuristics in two separate steps. We introduce a new method that uses saturated cost partitioning to select patterns and show that it outperforms all existing pattern selection algorithms.

The performance of domain-independent planning systems heavily depends on how the planning task has been modeled. This makes task reformulation an important tool to get rid of unnecessary complexity and increase the robustness of planners with respect to the model chosen by the user. In this paper, we represent tasks as factored transition systems (FTS), and use the merge-and-shrink (M&S) framework for task reformulation for optimal and satisficing planning. We prove that the flexibility of the underlying representation makes the M&S reformulation methods more powerful than the counterparts based on the more popular finite-domain representation. We adapt delete-relaxation and M&S heuristics to work on the FTS representation and evaluate the impact of our reformulation.

Landmarks are a core component of LAMA, a state-of-the-art satisficing planning system based on heuristic search. It uses landmarks to estimate the goal distance by summing up the costs of their cheapest achievers. This procedure ignores synergies between different landmarks: The cost of an action is counted multiple times if it is the cheapest achiever of several landmarks. Common admissible landmark heuristics tackle this problem by underapproximating the cost of a minimum hitting set of the landmark achievers. We suggest to overapproximate it by computing suboptimal hitting sets instead if admissibility is not a requirement. As our heuristics consider synergies between landmarks, we further propose to relax certain restrictions LAMA imposes on the number of landmarks and synergies between them. Our experimental evaluation shows a reasonable increase in the number of landmarks that leads to better guidance when used with our new heuristics.

Classical planning problems are defined using some specification language, such as PDDL. The domain expert defines action schemas, objects, the initial state, and the goal. One key aspect of PDDL is that the set of objects cannot be modified during plan execution. While this is fine in many domains, sometimes it makes modeling more complicated. This may impact the performance of planners, and it requires the domain expert to bound the number of required objects beforehand, which can be a challenge. We introduce an extension to the classical planning formalism, where action effects can create and remove objects. This problem is semi-decidable, but it becomes decidable if we can bound the number of objects in any given state, even though the state space is still infinite. On the practical side, we extend the Powerlifted planning system to support this PDDL extension. Our results show that this extension improves the performance of Powerlifted while supporting more natural PDDL models.

Correlation complexity is a measure of a planning task indicating how hard it is. The introducing work provides sufficient criteria to detect a correlation complexity of 2 on a planning task. It also introduced an example of a planning task with correlation complexity 3. In our work, we introduce a criterion to detect an arbitrary correlation complexity and extend the mentioned example to show with the new criterion that planning tasks with arbitrary correlation complexity exist.

Planning domains are an important notion, e.g. when it comes to restricting the input for generalized planning or learning approaches. However, domains as specified in PDDL cannot fully capture the intuitive understanding of a planning domain. We close this semantic gap and propose using PDDL axioms to characterize the (typically infinite) set of legal tasks of a domain. A minor extension makes it possible to express all properties that can be determined in polynomial time. We demonstrate the suitability of the approach on established domains from the International Planning Competition.

We introduce transition-counting constraints as a principled tool to formalize constraints that must hold in every solution of a transition system. We then show how to obtain transition landmark constraints from abstraction cuts. Transition landmarks dominate operator landmarks in theory but require solving a linear program that is prohibitively large in practice. We compare different constraints that project away transition-counting variables and then further relax the constraint. For one important special case, we provide a lossless projection. We finally discuss efficient data structures to derive cuts from abstractions and store them in a way that avoids repeated computation in every state. We compare the resulting heuristics both theoretically and on benchmarks from the international planning competition.

One of the strongest approaches for optimal classical planning is A* search with heuristics based on abstractions of the planning task. Abstraction heuristics are well studied in planning formalisms without conditional effects such as SAS+. However, conditional effects are crucial to model many planning tasks compactly. In this paper, we focus on factored tasks which allow a specific form of conditional effect, where effects on variable x can only depend on the value of x. We generalize projections, domain abstractions, Cartesian abstractions and the counterexample-guided abstraction refinement method to this formalism. While merge-and-shrink already covers factored task in theory, we provide an implementation that does so. In our experiments, we compare these abstraction-based heuristics to other heuristics supporting conditional effects, as well as symbolic search. On our new benchmark set of factored tasks, pattern database heuristics solve the most problems, followed by symbolic approaches on par with domain abstractions. The more general Cartesian and merge-and-shrink abstractions fall behind.

Classical planning considers a given task and searches for a plan to solve it. Some tasks are harder to solve than others. We can measure the “hardness” of a task with the novelty width and the correlation complexity. In this work, we compare these measures. Additionally, we introduce the river measure, a new measure that is based on potential heuristics and therefore similar to the correlation complexity but also comparable to the novelty width. We show that the river measure is upper bounded by the correlation complexity and by the novelty width +1. Furthermore, we show that we can convert a planning task with a polynomial blowup of the task size to ensure that a heuristic of dimension 2 exists that gives rise to backtrack-free search.

The merge-and-shrink framework is a powerful tool to construct state space abstractions based on factored representations. One of its core applications in classical planning is the construction of admissible abstraction heuristics. In this paper, we develop a compositional theory of merge-and-shrink in the context of probabilistic planning, focusing on stochastic shortest path problems (SSPs). As the basis for this development, we contribute a novel factored state space model for SSPs. We show how general transformations, including abstractions, can be formulated on this model to derive admissible and/or perfect heuristics. To formalize the merge-and-shrink framework for SSPs, we transfer the fundamental merge-and-shrink transformations from the classical setting: shrinking, merging, and label reduction. We analyze the formal properties of these transformations in detail and show how the conditions under which shrinking and label reduction lead to perfect heuristics can be extended to the SSP setting.

Abstraction heuristics are a state-of-the-art technique to solve classical planning problems optimally. A common approach is to precompute many small abstractions and combine them admissibly using cost partitioning. Recent work has shown that this approach does not work out well when using such heuristics for decoupled state space search, where search nodes represent potentially large sets of states. This is due to the fact that admissibly combining the estimates of several heuristics without sacrificing accuracy is NP-hard for decoupled states. In this work we propose to use a single large abstraction instead. We focus on merge-and-shrink and symbolic pattern database heuristics, which are designed to produce such abstractions. For these heuristics, we prove that the evaluation of decoupled states is NP-hard in general, but we also identify conditions under which it is polynomial. We introduce algorithms for both the general and the polynomial case. Our experimental evaluation shows that single large abstraction heuristics lead to strong performance when the heuristic evaluation is polynomial.

Restricted Tasks (RT) are a special case of numeric planning characterized by numeric conditions that involve one numeric variable per formula and numeric effects that allow only the addition of constants. Despite this, RTs form an expressive class whose planning problem is undecidable. The restricted nature of RTs often makes problem modeling awkward and unnecessarily complicated. We show that this can be alleviated by compiling mathematical operations that are not natively supported into RTs using macro-like action sequences. With that, we can encode many features found in general numeric planning such as constant multiplication, addition of linear formulas, and integer division and residue. We demonstrate how our compilations can be used to capture challenging mathematical problems such as the (in)famous Collatz conjecture. Our approach additionally gives a simple undecidability proof for RTs, and the proof shows that the number of variables needed to construct an undecidable class of RTs is surprisingly low: two numeric and one propositional variable.

Abstraction heuristics are the state of the art in optimal classical planning as heuristic search. A popular method for computing abstractions is the counterexample-guided abstraction refinement (CEGAR) principle, which has successfully been used for projections, which are the abstractions underlying pattern databases, and Cartesian abstractions. While projections are simple and fast to compute, Cartesian abstractions subsume projections and hence allow more finegrained abstractions, however at the expense of efficiency. Domain abstractions are a third class of abstractions between projections and Cartesian abstractions in terms of generality. Yet, to the best of our knowledge, they are only briefly considered in the planning literature but have not been used for computing heuristics yet. We aim to close this gap and compute domain abstractions by using the CEGAR principle. Our empirical results show that domain abstractions compare favorably against projections and Cartesian abstractions.

Algorithms are usually shown to be correct on paper, but bugs in their implementations can still lead to incorrect results. In the case of classical planning, it is fortunately straightforward to check whether a computed plan is correct. For optimal planning, however, plans are additionally required to have minimal cost, which is significantly more difficult to verify. While some domain-specific approaches exists, we lack a general tool to verify optimality for arbitrary problems. We bridge this gap and introduce two approaches based on the principle of certifying algorithms, which provide a computer-verifiable certificate of correctness alongside their answer. We show that both approaches are sound and complete, analyze whether they can be generated and verified efficiently, and show how to apply them to concrete planning algorithms. The experimental evaluation shows that verifying optimality comes with a cost, but is still practically feasible. Furthermore, it confirms that the tested planner configurations provide optimal plans on the given instances, as all certificates were verified successfully.

The computation of high-quality landmarks and orderings for heuristic state-space search is often prohibitively expensive to be performed in every generated state. Computing information only for the initial state and progressing it from every state to its successors is a successful alternative, exploited for example in classical planning by the LAMA planner. We propose a general framework for using landmarks in any kind of best-first search. Its core component, the progression function, uses orderings and search history to determine which landmarks must still be achieved. We show that the progression function that is used in LAMA infers invalid information in the presence of reasonable orderings. We define a sound progression function that allows to exploit reasonable orderings in cost-optimal planning and show empirically that our new progression function is beneficial both in satisficing and optimal planning.

Classical planning tasks are commonly described in a first-order language. However, most classical planners translate tasks by grounding them and then rewriting them into a propositional language. In recent years, the grounding step has become a larger bottleneck. In this work, we study how to improve it. We build on top of the most common grounder for planning tasks which uses Datalog to find all reachable atoms and actions. Inspired by recent progress in lifted planning, database theory, and algorithmics, we develop a new method to ground these Datalog programs. Our algorithm can ground more instances than the baseline, and most tasks it cannot ground are out of reach from any ground planner.

In classical planning, the aim is to find a sequence of deterministic actions leading from the initial to a goal state. In this work, we consider the scenario where a party who knows the solution to a planning task, called the prover, wants to convince a second party, the verifier, that it has the solution without revealing any information about the solution itself. This is relevant in domains where privacy is important, for example when plans contain sensitive information or when the solution should not be revealed upfront. We achieve this by introducing a zero-knowledge protocol for plan existence. By restricting ourselves to tasks with polynomially-bounded plan length, we are able to construct a protocol that can be run efficiently by both prover and verifier. The resulting protocol does not rely on any reduction, has a constant number of rounds, and runs in time polynomial in the size of the task.

The goal of generalized planning is to find a solution that works for all tasks of a specific planning domain. Ideally, this solution is also efficient (i.e., polynomial) in all tasks. One possible approach is to learn such a solution from training examples and then prove that this generalizes for any given task. However, such proofs are usually pen-and-paper proofs written by a human. In our paper, we aim at automating these proofs so we can use a theorem prover to show that a solution generalizes for any task. Furthermore, we want to prove that this generalization works while still preserving efficiency. Our focus is on generalized potential heuristics encoding tiered measures of progress, which can be proven to lead to a find in a polynomial number of steps in all tasks of a domain. We show our ongoing work in this direction using the interactive theorem prover Isabelle/HOL. We illustrate the key aspects of our implementation using the Miconic domain and then discuss possible obstacles and challenges to fully automating this pipeline.

Previous work introduced the concept of progress states. Once greedy best-first search expanded a progress state, it expands only states with lower heuristic values. Current methods identify progress states for a single task and after a solution for the task is found. We introduce a novel approach which learns for a classical planning domain a description logic formula to characterize all its progress states. This formula can be used during future searches. We present a first use case which exploits the learned formulas: During search we break ties in favor of progress states. This often significantly reduces the search effort.

Neural networks (NN) are increasingly investigated in AI Planning, and are used successfully to learn heuristic functions. NNs commonly not only predict a value, but also output a confidence in this prediction. From the perspective of heuristic search with NN heuristics, it is a natural idea to take this into account, e.g. falling back to a standard heuristic where confidence is low. We contribute an empirical study of this idea. We design search methods which prune nodes, or switch between search queues, based on the confidence of NNs. We furthermore explore the possibility of out-of-distribution (OOD) training, which tries to reduce the overconfidence of NNs on inputs different to the training distribution. In experiments on IPC benchmarks, we find that our search methods improve coverage over standard methods, and that OOD training has the desired effect in terms of prediction accuracy and confidence, though its impact on search seems marginal.

Abstraction heuristics are the state of the art in optimal classical planning as heuristic search. Despite their success for explicit-state search, though, abstraction heuristics are not available for decoupled state-space search, an orthogonal reduction technique that can lead to exponential savings by decomposing planning tasks. In this paper, we show how to compute pattern database (PDB) heuristics for decoupled states. The main challenge lies in how to additively employ multiple patterns, which is crucial for strong search guidance of the heuristics. We show that in the general case, for arbitrary collections of PDBs, computing the heuristic for a decoupled state is exponential in the number of leaf components of decoupled search. We derive several variants of decoupled PDB heuristics that allow to additively combine PDBs avoiding this blow-up and evaluate them empirically.

Debugging can be a painful task, especially when bugs only occur for large input files. We present Machetli, a tool to help with debugging in such situations. It takes a large input file and cuts away parts of it, while still provoking the bug. The resulting file is much smaller than the original, making the bug easier to find and fix. In our experience, Machetli was able to reduce planning tasks with thousands of actions to trivial tasks that could even be solved by hand. Machetli is an open-source project and it can be extended to other use cases such as debugging SAT solvers or LaTeX compilation bugs.

The goal of generalized planning is to find a solution that works for all tasks of a specific planning domain. Ideally, this solution is also efficient (i.e., polynomial) in all tasks. One possible approach is to learn such a solution from training examples and then prove that this generalizes for any given task. However, such proofs are usually pen-and-paper proofs written by a human. In our paper, we aim at automating these proofs so we can use a theorem prover to show that a solution generalizes for any task. Furthermore, we want to prove that this generalization works while still preserving efficiency. Our focus is on generalized potential heuristics encoding tiered measures of progress, which can be proven to lead to a find in a polynomial number of steps in all tasks of a domain. We show our ongoing work in this direction using the interactive theorem prover Isabelle/HOL. We illustrate the key aspects of our implementation using the Miconic domain and then discuss possible obstacles and challenges to fully automating this pipeline.

Since its invention in 1974, the Rubik’s Cube puzzle fascinates people of all ages. Its rules are simple: the player gets a scrambled cube and rotates the six faces until each face contains only stickers of one color. Nevertheless, finding a short sequence of rotations to solve the cube is hard. We present the first model of Rubik’s Cube for general problem solvers. To obtain a concise model, we require conditional effects. Furthermore, we extend counterexample-guided Cartesian abstraction refinement (CEGAR) to support factored effect tasks, a class of planning tasks with a specific kind of conditional effects which includes Rubik’s Cube. Finally, we evaluate how newer types of abstraction heuristics compare against pattern database (PDB) heuristics, the state-of-the-art for solving Rubik’s Cube. We find that PDBs still outperform the more general Cartesian and merge-and-shrink abstractions. However, in contrast to PDBs, Cartesian abstractions yield perfect heuristics up to a certain problem difficulty. These findings raise interesting questions for future research.

In classical planning, cost partitioning is a powerful method which allows to combine multiple admissible heuristics while retaining an admissible bound. In this paper, we extend the theory of cost partitioning to probabilistic planning by generalizing from deterministic transition systems to stochastic shortest path problems (SSPs). We show that fundamental results related to cost partitioning still hold in our extended theory. We also investigate how to optimally partition costs for a large class of abstraction heuristics for SSPs. Lastly, we analyze occupation measure heuristics for SSPs as well as the theory of approximate linear programming for reward-oriented Markov decision processes. All of these fit our framework and can be seen as cost-partitioned heuristics.

How can we train neural network (NN) heuristic functions for classical planning, using only states as the NN input? Prior work addressed this question by (a) per-instance imitation learning and/or (b) per-domain learning. The former limits the approach to instances small enough for training data generation, the latter to domains where the necessary knowledge generalizes across instances. Here we explore three methods for (a) that make training data generation scalable through bootstrapping and approximate value iteration. In particular, we introduce a new bootstrapping variant that estimates search effort instead of goal distance, which as we show converges to the perfect heuristic under idealized circumstances. We empirically compare these methods to (a) and (b), aligning three different NN heuristic function learning architectures for cross-comparison in an experiment of unprecedented breadth in this context. Key lessons are that our methods and imitation learning are highly complementary; that per-instance learning often yields stronger heuristics than per-domain learning; and the LAMA planner is still dominant but our methods outperform it in one benchmark domain.

Testing is a promising way to gain trust in neural action policies π. Previous work on policy testing in sequential decision making targeted environment behavior leading to failure conditions. But if the failure is unavoidable given that behavior, then π is not actually to blame. For a situation to qualify as a “bug” in π, there must be an alternative policy π0 that does better. We introduce a generic policy testing framework based on that intuition. This raises the bug confirmation problem, deciding whether or not a state is a bug. We analyze the use of optimistic and pessimistic bounds for the design of test oracles approximating that problem. We contribute an implementation of our framework in classical planning, experimenting with several test oracles and with random-walk methods generating test states biased to poor policy performance and/or state novelty. We evaluate these techniques on policies π learned with ASNets. We find that they are able to effectively identify bugs in these π, and that our random-walk biases improve over uninformed baselines.

While the unsolvability IPC sparked a multitude of planners proficient in detecting unsolvable planning tasks, there are gaps where concise unsolvability arguments are known but no existing planner can capture them without prohibitive computational effort. One such example is the sliding tiles puzzle, where solvability can be decided in polynomial time with a parity argument. We introduce separating functions, which can prove that one state is unreachable from another, and show under what conditions a potential function over any nonzero ring is a separating function. We prove that we can compactly encode these conditions for potential functions over features that are pairs, and show in which cases we can efficiently synthesize functions satisfying these conditions. We experimentally evaluate a domain-independent algorithm that successfully synthesizes such separating functions from PDDL representations of the sliding tiles puzzle, the Lights Out puzzle, and Peg Solitaire.

Potential functions are a general class of heuristics for classical planning. For satisficing planning, previous work suggested the use of descending and dead-end avoiding (DDA) potential heuristics, which solve planning tasks by backtrack-free search. In this work we study the complexity of devising DDA potential heuristics for classical planning tasks. We show that verifying or synthesizing DDA potential heuristics is PSPACE-complete, but suitable modifications of the DDA properties reduce the complexity of these problems to the first and second level of the polynomial hierarchy. We also discuss the implications of our results for other forms of heuristic synthesis in classical planning.

Lifted planners are useful to solve tasks that are too hard to ground. Still, computing informative lifted heuristics is difficult: directly adapting ground heuristics to the lifted setting is often too expensive, and extracting heuristics from the lifted representation can be uninformative. A natural alternative for lifted planners is to use width-based search. These algorithms are among the strongest for ground planning, even the variants that do not access the action model. In this work, we adapt best-first width search to the lifted setting and show that this yields state-of-the-art performance for hard-to-ground planning tasks.

Abstraction heuristics are the state of the art in optimal classical planning as heuristic search. Despite their success for explicit-state search, though, abstraction heuristics are not available for decoupled state-space search, an orthogonal reduction technique that can lead to exponential savings by decomposing planning tasks. In this paper, we show how to compute pattern database (PDB) heuristics for decoupled states. The main challenge lies in how to additively employ multiple patterns, which is crucial for strong search guidance of the heuristics. We show that in the general case, for arbitrary collections of PDBs, computing the heuristic for a decoupled state is exponential in the number of leaf components of decoupled search. We derive several variants of decoupled PDB heuristics that allow to additively combine PDBs avoiding this blow-up and evaluate them empirically.

Heuristics for lifted planning are not yet as informed as the best heuristics for ground planning. Recent work introduced the idea of using Datalog programs to compute the additive heuristic over lifted tasks. Based on this work, we show how to compute the more informed FF heuristic in a lifted manner. We extend the Datalog program with executable annotations that can also be used to define other delete-relaxation heuristics. In our experiments, we show that a planner using the lifted FF implementation produces state-of-the-art results for lifted planners. It also reduces the gap to state-of-the-art ground planners in domains where grounding is feasible.

Since no classical planner consistently outperforms all others, it is important to select a planner that works well for a given classical planning task. The two strongest approaches for planner selection use image and graph convolutional neural networks. They have the drawback that the learned models are complicated and uninterpretable. To obtain explainable models, we identify a small set of simple task features and show that elementary and interpretable machine learning techniques can use these features to solve roughly as many tasks as the complex approaches based on neural networks.

Stubborn sets are a pruning technique for state-space search which is well established in optimal classical planning. In this paper, we show that weak stubborn sets introduced in recent work in planning are actually not weak stubborn sets in Valmari's original sense. Based on this finding, we introduce weak stubborn sets in the original sense for planning by providing a generalized definition analogously to generalized strong stubborn sets in previous work. We discuss the relationship of strong, weak and the previously called weak stubborn sets, thus providing a further step in getting an overall picture of the stubborn set approach in planning.

The merge-and-shrink framework has been introduced as a general approach for defining abstractions of large state spaces arising in domain-independent planning and related areas. The distinguishing characteristic of the merge-and-shrink approach is that it operates directly on the factored representation of state spaces, repeatedly modifying this representation through transformations such as shrinking (abstracting a factor of the representation), merging (combining two factors), label reduction (abstracting the way in which different factors interact), and pruning (removing states or transitions of a factor).

We provide a novel view of the merge-and-shrink framework as a "toolbox" or "algebra" of transformations on factored transition systems, with the construction of abstractions as only one possible application. For each transformation, we study desirable properties such as conservativeness (overapproximating the original transition system), inducedness (absence of spurious states and transitions), and refinability (reconstruction of paths in the original transition system from the transformed one). We provide the first complete characterizations of the conditions under which these desirable properties can be achieved. We also provide the first full formal account of factored mappings, the mechanism used within the merge-and-shrink framework to establish the relationship between states in the original and transformed factored transition system.

Unlike earlier attempts to develop a theory for merge-and-shrink, our approach is fully compositional: the properties of a sequence of transformations can be entirely understood by the properties of the individual transformations involved. This aspect is key to the use of merge-and-shrink as a general toolbox for transforming factored transition systems. New transformations can easily be added to our theory, with compositionality taking care of the seamless integration with the existing components. Similarly, new properties of transformations can be integrated into the theory by showing their compositionality and studying under which conditions they are satisfied by the building blocks of merge-and-shrink.

Cost partitioning admissibly combines the information from multiple heuristics for optimal state-space search. One of the strongest cost partitioning algorithms is saturated cost partitioning. It considers the heuristics in sequence and assigns to each heuristic the minimal fraction of the remaining costs that are needed for preserving all heuristic estimates. Saturated cost partitioning has recently been generalized in two directions: first, by allowing to use different costs for the transitions induced by the same operator, and second, by preserving the heuristic estimates for only a subset of states. In this work, we unify these two generalizations and show that the resulting subset-saturated transition cost partitioning algorithm usually yields stronger heuristics than the two generalizations by themselves.

Cost partitioning is a general method for admissibly summing up heuristic estimates for optimal state-space search. Most cost partitioning algorithms can optimize the resulting cost-partitioned heuristic for a specific state. Since computing a new cost-partitioned heuristic for each evaluated state is usually too expensive in practice, the strongest planners based on cost partitioning over abstraction heuristics precompute a set of cost-partitioned heuristics before the search and maximize over their estimates during the search. This makes state evaluations very fast, but since there is no better termination criterion than a time limit, it requires a long precomputation phase, even for the simplest planning tasks. A prototypical example for this is the Scorpion planner which computes saturated cost partitionings over abstraction heuristics offline before the search. Using Scorpion as a case study, we show that by incrementally extending the set of cost-partitioned heuristics online during the search, we drastically speed up the planning process and even often solve more tasks.

The benchmarks from previous International Planning Competitions (IPCs) are the de-facto standard for evaluating planning algorithms. The IPC set is both a collection of planning domains and a selection of instances from these domains. Most of the domains come with a parameterized generator that generates new instances for a given set of parameter values. Due to the steady progress of planning research some of the instances that were generated for past IPCs are inadequate for evaluating current planners. To alleviate this problem, we introduce Autoscale, an automatic tool that selects instances for a given domain. Autoscale takes into account constraints from the domain designer as well as the performance of current planners to generate an instance set of appropriate difficulty, while avoiding too much bias with respect to the considered planners. We show that the resulting benchmark set is superior to the IPC set and has the potential of improving empirical evaluation of planning research.

Landmarks of a planning task denote properties that must be satisfied by all plans. Existing landmark heuristics exploit that each landmark must be achieved at least once. However, if the orderings between the landmarks induce cyclic dependencies, one of the landmarks in each cycle must be achieved an additional time. We propose two novel heuristics for cost-optimal planning that consider cyclic dependencies between landmarks in addition to the cost for achieving all landmarks once. We show that our heuristics dominate the minimum hitting set solution over any set of landmarks as well as h + if all delete-relaxation landmarks are considered. An experimental evaluation on benchmarks from the International Planning Competition shows that exploiting cyclic dependencies can lead to improved heuristics.

Optimal cost partitioning can produce high quality heuristic estimates even from small abstractions. It can be computed with a linear program (LP) but the size of this LP often makes this impractical. Recent work used Lagrangian decomposition to speed up the computation. Here we use a different decomposition technique called Dantzig-Wolfe decomposition to tackle the problem. This gives new insights into optimal cost partitioning and has several advantages over Lagrangian decomposition: our method detects when a cost partition is optimal; it can deal with general cost functions; and it does not consider abstractions in the linear program that do not contribute to the heuristic value. We also show the advantage of the method empirically and investigate several improvements that are useful for all cost partitioning methods.

Planning is an attractive approach to solving problems because of its generality and its ease of use. A domain expert often has knowledge that could improve the performance of such a solution but this knowledge may be vague or not directly actionable. For example, an expert could be aware that a certain property is important but not have a good way of using this knowledge to speed up the search for a solution. In this paper we evaluate concept languages as a possible input language for injecting expert domain knowledge into a planning system. We also explore mixed integer programming as a way to use this knowledge to improve search efficiency and to help the user find and refine useful domain knowledge.

It is usually prohibitively expensive to recompute landmarks in every state of a heuristic search algorithm that bases its heuristic estimate on landmarks. Successful planning systems like LAMA compute landmarks just for the initial state instead and only keep track for each state encountered during search which landmarks of the initial state are accepted or required again. The LM-A* algorithm describes a similar idea for cost-optimal heuristic search. However, it lacks support for reasonable orderings, which have recently been shown to be informative when cyclic dependencies between landmarks are exploited in the heuristic computation. We propose the novel LM-BFS framework for best-first heuristic search with landmark-based heuristics which allows to describe a variety of algorithms in terms of five components: landmark generation, landmark graph progression, landmark graph merging, landmark graph extension, and exploration strategy. We use LM-BFS to show that considering reasonable orderings in the landmark graph progression component leads to flawed landmark graphs in LAMA , and propose a novel landmark graph extension that captures the same information flawlessly. We observe empirically that considering reasonable orderings in the extension rather than the progression component has a positive effect both for LAMA and for cost-optimal planning.

Saturated cost partitioning and post-hoc optimization are two powerful cost partitioning algorithms for optimal classical planning. The main idea of saturated cost partitioning is to give each considered heuristic only the fraction of remaining operator costs that it needs to prove its estimates. We show how to apply this idea to post-hoc optimization and obtain a heuristic that dominates the original both in theory and on the IPC benchmarks.

Landmarks are a core component of LAMA, a state-of-the-art satisficing planning system based on heuristic search. It uses landmarks to estimate the goal distance by summing up the costs of their cheapest achievers. This procedure ignores synergies between different landmarks: The cost of an action is counted multiple times if it is the cheapest achiever of several landmarks. Common admissible landmark heuristics tackle this problem by underapproximating the cost of a minimum hitting set of the landmark achievers. We suggest to overapproximate it by computing suboptimal hitting sets instead if admissibility is not a requirement. As our heuristics consider synergies between landmarks, we further propose to relax certain restrictions LAMA imposes on the number of landmarks and synergies between them. Our experimental evaluation shows a reasonable increase in the number of landmarks that leads to better guidance when used with our new heuristics.

Classical planning problems are defined using some specification language, such as PDDL. The domain expert defines action schemas, objects, the initial state, and the goal. One key aspect of PDDL is that the set of objects cannot be modified during plan execution. While this is fine in many domains, sometimes it makes modeling more complicated. This may impact the performance of planners, and it requires the domain expert to bound the number of required objects beforehand, which can be a challenge. We introduce an extension to the classical planning formalism, where action effects can create and remove objects. This problem is semi-decidable, but it becomes decidable if we can bound the number of objects in any given state, even though the state space is still infinite. On the practical side, we extend the Powerlifted planning system to support this PDDL extension. Our results show that this extension improves the performance of Powerlifted while supporting more natural PDDL models.

Planning domains are an important notion, e.g. when it comes to restricting the input for generalized planning or learning approaches. However, domains as specified in PDDL cannot fully capture the intuitive understanding of a planning domain. We close this semantic gap and propose using PDDL axioms to characterize the (typically infinite) set of legal tasks of a domain. A minor extension makes it possible to express all properties that can be determined in polynomial time. We demonstrate the suitability of the approach on established domains from the International Planning Competition.

Classical planning tasks are commonly described in a first-order language. However, most classical planners translate tasks by grounding them and then rewriting them into a propositional language. In recent years, the grounding step has become a larger bottleneck. In this work, we study how to improve it. We build on top of the most common grounder for planning tasks which uses Datalog to find all reachable atoms and actions. Inspired by recent progress in lifted planning, database theory, and algorithmics, we develop a new method to ground these Datalog programs. Our algorithm can ground more instances than the baseline, and most tasks it cannot ground are out of reach from any ground planner.

Hybrid systems represent an important and powerful formalism for modeling real-world applications such as embedded systems. A verification tool like SpaceEx is based on the exploration of a symbolic search space (the region space). As a verification tool, it is typically optimized towards proving the absence of errors. In some settings, e.g., when the verification tool is employed in a feedback-directed design cycle, one would like to have the option to call a version that is optimized towards finding an error trajectory in the region space. A recent approach in this direction is based on guided search. Guided search relies on a cost function that indicates which states are promising to be explored, and preferably explores more promising states first. In this paper, we propose an abstraction-based cost function based on coarse-grained space abstractions for guiding the reachability analysis. For this purpose, a suitable abstraction technique that exploits the flexible granularity of modern reachability analysis algorithms is introduced. The new cost function is an effective extension of pattern database approaches that have been successfully applied in other areas. The approach has been implemented in the SpaceEx model checker. The evaluation shows its practical potential.

Directed model checking is a well-established approach for detecting error states in concurrent systems. A popular variant to find shortest error traces is to apply the A* search algorithm with distance heuristics that never overestimate the real error distance. An important class of such distance heuristics is the class of pattern database heuristics. Pattern database heuristics are built on abstractions of the system under consideration. In this paper, we propose downward pattern refinement, a systematic approach for the construction of pattern database heuristics for concurrent systems of timed automata. First, we propose a general framework for pattern databases in the context of timed automata and show that desirable theoretical properties hold for the resulting pattern database. Afterward, we formally define a concept to measure the accuracy of abstractions. Based on this concept, we propose an algorithm for computing succinct abstractions that are still accurate to produce informed pattern databases. We evaluate our approach on large and complex industrial problems. The experiments show the practical potential of the resulting pattern database heuristic.

Many heuristics for cost-optimal planning are based on linear programming. We cover several interesting heuristics of this type by a common framework that fixes the objective function of the linear program. Within the framework, constraints from different heuristics can be combined in one heuristic estimate which dominates the maximum of the component heuristics. Different heuristics of the framework can be compared on the basis of their constraints. We present theoretical results on the relation between existing heuristics and experimental results that demonstrate the potential of the proposed framework.

We propose the application of variable neighborhood search to address the problem of local optima during the hill-climbing procedure of iPDB.

We prove that general merge-and-shrink representations are strictly more powerful than linear ones by showing that there exist problem families that can be represented compactly with general merge-and-shrink representations but not with linear ones. We also give a precise bound that quantifies the necessary blowup incurred by conversions from general merge-and-shrink representations to linear representations or BDDs/ADDs. Our theoretical results suggest an untapped potential for non-linear merging strategies and for the use of non-linear merge-and-shrink-like representations within symbolic search.

Potential heuristics, recently introduced by Pommerening et al., characterize admissible and consistent heuristics for classical planning as a set of declarative constraints. Every feasible solution for these constraints defines an admissible heuristic, and we can obtain heuristics that optimize certain criteria such as informativeness by specifying suitable objective functions.

The original paper only considered one such objective function: maximizing the heuristic value of the initial state. In this paper, we explore objectives that attempt to maximize heuristic estimates for all states (reachable and unreachable), maximize heuristic estimates for a sample of reachable states, maximize the number of detected dead ends, or minimize search effort. We also search for multiple heuristics with complementary strengths that can be combined to obtain even better heuristics.

Operator cost partitioning is a well-known technique to make admissible heuristics additive by distributing the operator costs among individual heuristics. Planning tasks are usually defined with non-negative operator costs and therefore it appears natural to demand the same for the distributed costs. We argue that this requirement is not necessary and demonstrate the benefit of using general cost partitioning. We show that LP heuristics for operator-counting constraints are cost-partitioned heuristics and that the state equation heuristic computes a cost partitioning over atomic projections. We also introduce a new family of potential heuristics and show their relationship to general cost partitioning.

Many recent planning heuristics are based on LP optimization. However, planning experts mostly use LP solvers as a black box and it is often not obvious to them which LP techniques would be most suitable for their specific applications.

To foster the communication between the planning and the optimization community, this paper gives an easily accessible overview over these recent LP-based heuristics, namely the optimal cost partitioning heuristic for abstractions, the post-hoc optimization heuristic, a landmark heuristic, the state-equation heuristic, and a delete relaxation heuristic. All these heuristics fit the framework of so-called operator-counting constraints, which we also present.

Pattern Databases (PDBs) are a common form of abstraction-based heuristic which are often compressed so that a large PDB can fit in memory. Partial Pattern Databases (PPDBs) achieve this by storing only layers of the PDB which are close to the goal. This paper studies the problem of how to best compress and use the 457 GB 12-edge Rubik's cube PDB, suggesting a number of ways that Bloom filters can be used to effectively compress PPDBs. We then develop a theoretical model of the common min compression approach and our Bloom filters, showing that the original method of compressed PPDBs can never be better than min compression. We conclude with experimental results showing that Bloom filter compression of PPDBs provides superior performance to min compression in Rubik's cube.

Label reduction is a technique for simplifying families of labeled transition systems by dropping distinctions between certain transition labels. While label reduction is critical to the efficient computation of merge-and-shrink heuristics, current theory only permits reducing labels in a limited number of cases. We generalize this theory so that labels can be reduced in every intermediate abstraction of a merge-and-shrink tree. This is particularly important for efficiently computing merge-and-shrink abstractions based on non-linear merge strategies. As a case study, we implement a non-linear merge strategy based on the original work on merge-and-shrink heuristics in model checking by Dräger et al.

Many heuristics for cost-optimal planning are based on linear programming. We cover several interesting heuristics of this type by a common framework that fixes the objective function of the linear program. Within the framework, constraints from different heuristics can be combined in one heuristic estimate which dominates the maximum of the component heuristics. Different heuristics of the framework can be compared on the basis of their constraints. With this new method of analysis, we show dominance of the recent LP-based state-equation heuristic over optimal cost partitioning on single-variable abstractions. We also show that the previously suggested extension of the state-equation heuristic to exploit safe variables cannot lead to an improved heuristic estimate. We experimentally evaluate the potential of the proposed framework on an extensive suite of benchmark tasks.

We have recently shown how counterexample-guided abstraction refinement can be used to derive informative Cartesian abstraction heuristics for optimal classical planning. In this work we introduce two methods for producing diverse sets of heuristics within this framework, one based on goal facts, the other based on landmarks. In order to sum the heuristic estimates admissibly we present a novel way of finding cost partitionings for explicitly represented abstraction heuristics. We show that the resulting heuristics outperform other state-of-the-art abstraction heuristics on many benchmark domains.

The iPDB procedure by Haslum et al. is the state-of-the-art method for computing additive abstraction heuristics for domain-independent planning. It performs a hill-climbing search in the space of pattern collections, combining information from multiple patterns in the so-called canonical heuristic.

We show how stronger heuristic estimates can be obtained through linear programming. An experimental evaluation demonstrates the strength of the new technique on the IPC benchmark suite.

Hybrid systems represent an important and powerful formalism for modeling real-world applications such as embedded systems. A verification tool like SpaceEx is based on the exploration of a symbolic search space (the region space). As a verification tool, it is typically optimized towards proving the absence of errors. In some settings, e.g., when the verification tool is employed in a feedback-directed design cycle, one would like to have the option to call a version that is optimized towards finding an error path in the region space. A recent approach in this direction is based on guided search. Guided search relies on a cost function that indicates which states are promising to be explored, and preferably explores more promising states first. In this paper, an abstraction-based cost function based on pattern databases for guiding the reachability analysis is proposed. For this purpose, a suitable abstraction technique that exploits the flexible granularity of modern reachability analysis algorithms is introduced. The new cost function is an effective extension of pattern database approaches that have been successfully applied in other areas. The approach has been implemented in the SpaceEx model checker. The evaluation shows its practical potential.

Counterexample-guided abstraction refinement (CEGAR) is a methodological framework for incrementally computing abstractions of transition systems.

We propose a CEGAR algorithm for computing abstraction heuristics for optimal classical planning. Starting from a coarse abstraction of the planning task, we iteratively compute an optimal abstract solution, check if and why it fails for the concrete planning task and refine the abstraction so that the same failure cannot occur in future iterations.

A key ingredient of our approach is a novel class of abstractions for classical planning tasks that admits efficient and very fine-grained refinement. Our implementation performs tens of thousands of refinement steps in a few minutes and produces heuristics that are often significantly more accurate than pattern database heuristics of the same size.

Despite their general success in the heuristic search community, pattern database (PDB) heuristics have, until very recently, not been used by the most successful classical planning systems.

We describe a new efficient implementation of pattern database heuristics within the Fast Downward planner. A planning system using this implementation is competitive with the state of the art in optimal planning, significantly improving over results from the previous best PDB heuristic implementation in planning.

In domain-independent planning, dependencies of operators and variables often prevent the effective application of planning techniques that rely on "loosely coupled" problems (like factored planning or partial order reduction). In this paper, we propose a generic approach for factorizing a classical planning problem into an equivalent problem with fewer operator and variable dependencies. Our approach is based on variable factorization, which can be reduced to the well-studied problem of graph factorization. While the state spaces of the original and the factorized problems are isomorphic, the factorization offers the potential to exponentially reduce the complexity of planning techniques like factored planning and partial order reduction.

Recent work has introduced fork-decoupled search, addressing classical planning problems where a single center component provides preconditions for several leaf components. Given a fixed center path π^C, the leaf moves compliant with π^C can then be scheduled independently for each leaf. Fork-decoupled search thus searches over center paths only, maintaining the compliant paths for each leaf separately. This can yield dramatic benefits. It is empirically complementary to partial order reduction via strong stubborn sets, in that each method yields its strongest reductions in different benchmarks. Here we show that the two methods can be combined, in the form of strong stubborn sets for fork-decoupled search. This can yield exponential advantages relative to both methods. Empirically, the combination reliably inherits the best of its components, and often outperforms both.

Symmetry reduction has significantly contributed to the success of classical planning as heuristic search. However, it is an open question if symmetry reduction techniques can be lifted to fully observable nondeterministic (FOND) planning. We generalize the concepts of structural symmetries and symmetry reduction to FOND planning and specifically to the LAO* algorithm. Our base implementation of LAO* in the Fast Downward planner is competitive with the LAO*-based FOND planner myND. Our experiments further show that symmetry reduction can yield strong performance gains compared to our base implementation of LAO*.

The sleep sets technique is a path-dependent pruning method for state space search. In the past, the combination of sleep sets with graph search algorithms that perform duplicate elimination has often shown to be error-prone. In this paper, we provide the theoretical basis for the integration of sleep sets with common search algorithms in AI that perform duplicate elimination. Specifically, we investigate approaches to safely integrate sleep sets with optimal (best-first) search algorithms. Based on this theory, we provide an initial step towards integrating sleep sets within A* and additional state pruning techniques like strong stubborn sets. Our experiments show slight, yet consistent improvements on the number of generated search nodes across a large number of standard domains from the international planning competitions.

Symmetries provide the basis for well-established approaches to tackle the state explosion problem in state space search and in AI planning. However, although by now there are various symmetry-based techniques available, these techniques have not yet been empirically evaluated and compared to each other in a common setting. In particular, it is unclear which of them should be preferably applied, and whether there are techniques with stronger performance than others. In this paper, we shed light on this issue by providing an empirical case study. We combine and evaluate several symmetry-based techniques for cost-optimal planning as heuristic search. For our evaluation, we use state-of-the-art abstraction heuristics on a large set of benchmarks from the international planning competitions.

Pruning techniques based on partial order reduction and symmetry elimination have recently found increasing attention for optimal planning. Although these techniques appear to be rather different, they base their pruning decisions on similar ideas from a high level perspective. In this paper, we propose safe integrations of partial order reduction and symmetry elimination for cost-optimal classical planning. We show that previously proposed symmetry-based search algorithms can safely be applied with strong stubborn sets. In addition, we derive the notion of symmetrical strong stubborn sets as a more tightly integrated concept. Our experiments show the potential of our approaches.

Planning in hybrid domains poses a special challenge due to the involved mixed discrete-continuous dynamics. A recent solving approach for such domains is based on applying model checking techniques on a translation of PDDL+ planning problems to hybrid automata. However, the proposed translation is limited because must behavior is only over-approximated, and hence, processes and events are not reflected exactly. In this paper, we present the theoretical foundation of an exact PDDL+ translation. We propose a schema to convert a hybrid automaton with must transitions into an equivalent hybrid automaton featuring only may transitions.

We describe transition normal form (TNF) for classical planning tasks, where there is a unique goal state and variables occur in an operator precondition iff they appear in the effect. Tasks can be efficiently converted to TNF, all common planning heuristics are invariant under the transformation, and tasks in normal form are easier to study theoretically.

The stubborn set method is a state-space reduction technique, originally introduced in model checking and then transfered to classical planning. It was shown that stubborn sets significantly improve the performance of optimal deterministic planners by considering only a subset of applicable operators in a state. Fully observable nondeterministic planning (FOND) extends the formalism of classical planning by nondeterministic operators. We show that stubborn sets are also beneficial for FOND problems. We introduce nondeterministic stubborn sets, stubborn sets which preserve strong cyclic plans. We follow two approaches: Fast Incremental Planning with stubborn sets from classical planning and LAO* search with nondeterministic stubborn sets. Our experiments show that both approaches increase coverage and decrease node generations when compared to their respective baselines.

Merge-and-shrink heuristics crucially rely on effective reduction techniques, such as bisimulation-based shrinking, to avoid the combinatorial explosion of abstractions. We propose the concept of factored symmetries for merge-and-shrink abstractions based on the established concept of symmetry reduction for state-space search. We investigate under which conditions factored symmetry reduction yields perfect heuristics and discuss the relationship to bisimulation. We also devise practical merging strategies based on this concept and experimentally validate their utility.

Algorithms are usually shown to be correct on paper, but bugs in their implementations can still lead to incorrect results. In the case of classical planning, it is fortunately straightforward to check whether a computed plan is correct. For optimal planning, however, plans are additionally required to have minimal cost, which is significantly more difficult to verify. While some domain-specific approaches exists, we lack a general tool to verify optimality for arbitrary problems. We bridge this gap and introduce two approaches based on the principle of certifying algorithms, which provide a computer-verifiable certificate of correctness alongside their answer. We show that both approaches are sound and complete, analyze whether they can be generated and verified efficiently, and show how to apply them to concrete planning algorithms. The experimental evaluation shows that verifying optimality comes with a cost, but is still practically feasible. Furthermore, it confirms that the tested planner configurations provide optimal plans on the given instances, as all certificates were verified successfully.

In classical planning, the aim is to find a sequence of deterministic actions leading from the initial to a goal state. In this work, we consider the scenario where a party who knows the solution to a planning task, called the prover, wants to convince a second party, the verifier, that it has the solution without revealing any information about the solution itself. This is relevant in domains where privacy is important, for example when plans contain sensitive information or when the solution should not be revealed upfront. We achieve this by introducing a zero-knowledge protocol for plan existence. By restricting ourselves to tasks with polynomially-bounded plan length, we are able to construct a protocol that can be run efficiently by both prover and verifier. The resulting protocol does not rely on any reduction, has a constant number of rounds, and runs in time polynomial in the size of the task.

The goal of generalized planning is to find a solution that works for all tasks of a specific planning domain. Ideally, this solution is also efficient (i.e., polynomial) in all tasks. One possible approach is to learn such a solution from training examples and then prove that this generalizes for any given task. However, such proofs are usually pen-and-paper proofs written by a human. In our paper, we aim at automating these proofs so we can use a theorem prover to show that a solution generalizes for any task. Furthermore, we want to prove that this generalization works while still preserving efficiency. Our focus is on generalized potential heuristics encoding tiered measures of progress, which can be proven to lead to a find in a polynomial number of steps in all tasks of a domain. We show our ongoing work in this direction using the interactive theorem prover Isabelle/HOL. We illustrate the key aspects of our implementation using the Miconic domain and then discuss possible obstacles and challenges to fully automating this pipeline.

The goal of generalized planning is to find a solution that works for all tasks of a specific planning domain. Ideally, this solution is also efficient (i.e., polynomial) in all tasks. One possible approach is to learn such a solution from training examples and then prove that this generalizes for any given task. However, such proofs are usually pen-and-paper proofs written by a human. In our paper, we aim at automating these proofs so we can use a theorem prover to show that a solution generalizes for any task. Furthermore, we want to prove that this generalization works while still preserving efficiency. Our focus is on generalized potential heuristics encoding tiered measures of progress, which can be proven to lead to a find in a polynomial number of steps in all tasks of a domain. We show our ongoing work in this direction using the interactive theorem prover Isabelle/HOL. We illustrate the key aspects of our implementation using the Miconic domain and then discuss possible obstacles and challenges to fully automating this pipeline.

How can we train neural network (NN) heuristic functions for classical planning, using only states as the NN input? Prior work addressed this question by (a) per-instance imitation learning and/or (b) per-domain learning. The former limits the approach to instances small enough for training data generation, the latter to domains where the necessary knowledge generalizes across instances. Here we explore three methods for (a) that make training data generation scalable through bootstrapping and approximate value iteration. In particular, we introduce a new bootstrapping variant that estimates search effort instead of goal distance, which as we show converges to the perfect heuristic under idealized circumstances. We empirically compare these methods to (a) and (b), aligning three different NN heuristic function learning architectures for cross-comparison in an experiment of unprecedented breadth in this context. Key lessons are that our methods and imitation learning are highly complementary; that per-instance learning often yields stronger heuristics than per-domain learning; and the LAMA planner is still dominant but our methods outperform it in one benchmark domain.

Testing is a promising way to gain trust in neural action policies π. Previous work on policy testing in sequential decision making targeted environment behavior leading to failure conditions. But if the failure is unavoidable given that behavior, then π is not actually to blame. For a situation to qualify as a “bug” in π, there must be an alternative policy π0 that does better. We introduce a generic policy testing framework based on that intuition. This raises the bug confirmation problem, deciding whether or not a state is a bug. We analyze the use of optimistic and pessimistic bounds for the design of test oracles approximating that problem. We contribute an implementation of our framework in classical planning, experimenting with several test oracles and with random-walk methods generating test states biased to poor policy performance and/or state novelty. We evaluate these techniques on policies π learned with ASNets. We find that they are able to effectively identify bugs in these π, and that our random-walk biases improve over uninformed baselines.

Lifted planners are useful to solve tasks that are too hard to ground. Still, computing informative lifted heuristics is difficult: directly adapting ground heuristics to the lifted setting is often too expensive, and extracting heuristics from the lifted representation can be uninformative. A natural alternative for lifted planners is to use width-based search. These algorithms are among the strongest for ground planning, even the variants that do not access the action model. In this work, we adapt best-first width search to the lifted setting and show that this yields state-of-the-art performance for hard-to-ground planning tasks.

Heuristics for lifted planning are not yet as informed as the best heuristics for ground planning. Recent work introduced the idea of using Datalog programs to compute the additive heuristic over lifted tasks. Based on this work, we show how to compute the more informed FF heuristic in a lifted manner. We extend the Datalog program with executable annotations that can also be used to define other delete-relaxation heuristics. In our experiments, we show that a planner using the lifted FF implementation produces state-of-the-art results for lifted planners. It also reduces the gap to state-of-the-art ground planners in domains where grounding is feasible.

Since no classical planner consistently outperforms all others, it is important to select a planner that works well for a given classical planning task. The two strongest approaches for planner selection use image and graph convolutional neural networks. They have the drawback that the learned models are complicated and uninterpretable. To obtain explainable models, we identify a small set of simple task features and show that elementary and interpretable machine learning techniques can use these features to solve roughly as many tasks as the complex approaches based on neural networks.

Recent research in classical planning has shown the importance of search techniques that operate directly on the lifted representation of the problem, particularly in domains where the ground representation is prohibitively large. In this paper, we show how to compute the additive and maximum heuristics from the lifted representation of a problem. We do this by adapting well-known reachability analysis techniques based on a Datalog formulation of the delete relaxation of the problem. Our adaptation allows us to obtain not only the desired heuristic value, but also other useful heuristic information such as helpful actions. Our empirical evaluation shows that our lifted version of the additive heuristic is competitive with its ground counterpart on most of the standard international competition benchmarks, and significantly outperforms other state-of-the-art lifted heuristic methods in the literature.

How can we train neural network (NN) heuristic functions for classical planning, using only states as the NN input? Prior work addressed this question by (a) supervised learning and/or (b) per-domain learning generalizing over problem instances. The former limits the approach to instances small enough for training data generation, the later to domains and instance distributions where the necessary knowledge generalizes across instances. Clearly, reinforcement learning (RL) on large instances can potentially avoid both difficulties. We explore this here in terms of three methods drawing on previous ideas relating to bootstrapping and approximate value iteration, including a new bootstrapping variant that estimates search effort instead of goal distance. We empirically compare these methods to (a) and (b), aligning three different NN heuristic function learning architectures for cross-comparison in an experiment of unprecedented breadth in this context. Key lessons from this experiment are that our methods and supervised learning are highly complementary; that per-instance learning often yields stronger heuristics than per-domain learning; and that LAMA is still dominant but is outperformed by our methods in one benchmark domain.

Planning is an attractive approach to solving problems because of its generality and its ease of use. A domain expert often has knowledge that could improve the performance of such a solution but this knowledge may be vague or not directly actionable. For example, an expert could be aware that a certain property is important but not have a good way of using this knowledge to speed up the search for a solution. In this paper we evaluate concept languages as a possible input language for injecting expert domain knowledge into a planning system. We also explore mixed integer programming as a way to use this knowledge to improve search efficiency and to help the user find and refine useful domain knowledge.

While classical planning systems can usually detect if a task is unsolvable, only recent research introduced a way to verify such a claim. These methods have already been applied to a variety of explicit and symbolic search algorithms, but so far no planning technique based on SAT has been covered with them. We fill this gap by showing how property directed reachability can produce proofs while only minimally altering the framework by allowing to utilize certificates for unsolvable SAT queries within the proof. We additionally show that a variant of the algorithm that does not use SAT calls can produce proofs that fit into the existing framework without requiring any changes.

The standard PDDL language for classical planning uses several first-order features, such as schematic actions. Yet, most classical planners ground this first-order representation into a propositional one as a preprocessing step. While this simplifies the design of other parts of the planner, in several benchmarks the grounding process causes an exponential blowup that puts otherwise solvable tasks out of reach of the planners. In this work, we take a step towards planning with lifted representations. We tackle the successor generation task, a key operation in forward-search planning, directly on the lifted representation using well-known techniques from database theory. We show how computing the variable substitutions that make an action schema applicable in a given state is essentially a query evaluation problem. Interestingly, a large number of the action schemas in the standard benchmarks are acyclic conjunctive queries, for which query evaluation is tractable. Our empirical results show that our approach is competitive with the standard (grounded) successor generation techniques in a few domains and outperforms them on benchmarks where grounding is challenging or infeasible.

Informed heuristics are essential for the success of heuristic search algorithms. But, it is difficult to develop a new heuristic which is informed on various tasks. Instead, we propose a framework that trains a neural network as heuristic for the tasks it is supposed to solve. We present two reinforcement learning approaches to learn heuristics for fixed state spaces and fixed goals. Our first approach uses approximate value iteration, our second approach uses searches to generate training data. We show that in some domains our approaches outperform previous work, and we point out potentials for future improvements.

Since no classical planner consistently outperforms all others, it is important to select a planner that works well for a given classical planning task. The two strongest approaches for planner selection use image and graph convolutional neural networks. They have the drawback that the learned mod- els are not interpretable. To obtain explainable models, we identify a small set of simple task features and show that elementary and interpretable machine learning techniques can use these features to solve as many tasks as the approaches based on neural networks.

There exists no planning algorithm that outperforms all others. Therefore, it is important to know which algorithm works well on a task. A recently published approach uses either image or graph convolutional neural networks to solve this problem and achieves top performance. Especially the transforma- tion from the task to an image ignores a lot of information. Thus, we would like to know what the network is learning and if this is reasonable. As this is currently not possible, we take one step back. We identify a small set of simple graph features and show that elementary and interpretable machine learning techniques can use those features to outperform the neural network based approach. Furthermore, we evaluate the importance of those features and verify that the performance of our approach is robust to changes in the training and test data.

Neural networks (NN) have been shown to be powerful state-value predictors in several complex games. Can similar successes be achieved in classical planning? Towards a systematic exploration of that question, we contribute a study of hyperparameter space in the most canonical setup: input = state, feed-forward NN, supervised learning, generalization only over initial state. We investigate a broad range of hyperparameters pertaining to NN design and training. We evaluate these techniques through their use as heuristic functions in Fast Downward. The results on IPC benchmarks show that highly competitive heuristics can be learned, yielding substantially smaller search spaces than standard techniques on some do- mains. But the heuristic functions are costly to evaluate, and the range of domains where useful heuristics are learned is limited. Our study provides the basis for further research improving on current weaknesses.

Generalized planning aims at computing solutions that work for all instances of the same domain. In this paper, we show that several interesting planning domains possess compact generalized heuristics that can guide a greedy search in guaranteed polynomial time to the goal, and which work for any instance of the domain. These heuristics are weighted sums of state features that capture the number of objects satisfying a certain first-order logic property in any given state. These features have a meaningful interpretation and generalize naturally to the whole domain. Additionally, we present an approach based on mixed integer linear programming to compute such heuristics automatically from the observation of small training instances. We develop two variations of the approach that progressively refine the heuristic as new states are encountered. We illustrate the approach empirically on a number of standard domains, where we show that the generated heuristics will correctly generalize to all possible instances.

We describe a new algorithm for generating pattern collections for pattern database heuristics in optimal classical planning. The algorithm uses the counterexample-guided abstraction refinement (CEGAR) principle to guide the pattern selection process. Our experimental evaluation shows that a single run of the CEGAR algorithm can compute informative pattern collections in a fairly short time. Using multiple CEGAR algorithm runs, we can compute much larger pattern collections, still in shorter time than existing approaches, which leads to a planner that outperforms the state-of-the-art pattern selection methods by a significant margin.

The LM-Cut heuristic is currently the most successful heuristic in optimal STRIPS planning but it cannot be applied in the presence of conditional effects. Keyder, Hoffmann and Haslum recently showed that the obvious extensions to such effects ruin the nice theoretical properties of LM-Cut. We propose a new method based on context splitting that preserves these properties.

In heuristic search and especially in optimal classical planning the computation of accurate heuristic values can take up the majority of runtime. In many cases, the heuristic computations for a search node and its successors are very similar, leading to significant duplication of effort. For example most landmarks of a node that are computed by the LM-cut algorithm are also landmarks for the node's successors.

We propose to reuse these landmarks and incrementally compute new ones to speed up the LM-cut calculation. The speed advantage obtained by incremental computation is offset by higher memory usage. We investigate different search algorithms that reduce memory usage without sacrificing the faster computation, leading to a substantial increase in coverage for benchmark domains from the International Planning Competitions.

We adapt a partial order reduction technique based on stubborn sets, originally proposed for detecting dead ends in Petri Nets, to the setting of optimal planning. We demonstrate that stubborn sets can provide significant state space reductions on standard planning benchmarks, outperforming the expansion core method.

Optimal plans of delete-free planning tasks are interesting both in domains that have no delete effects and as the relaxation heuristic h+ in general planning. Many heuristics for optimal and satisficing planning approximate the h+ heuristic, which is well-informed and admissible but intractable to compute. In this work, branch-and-bound and IDA* search are used in a search space tailored to delete-free planning together with an incrementally computed version of the LM-cut heuristic. The resulting algorithm for optimal delete-free planning exceeds the performance of A* with the LM-cut heuristic in the state-of-the-art planner Fast Downward.

The pancake problem is a famous search problem where the objective is to sort a sequence of objects (pancakes) through a minimal number of prefix reversals (flips). The best approaches for the problem are based on heuristic search with abstraction (pattern database) heuristics. We present a new class of abstractions for the pancake problem called relative-order abstractions. Relative-order abstractions have three advantages over the object-location abstractions considered in previous work. First, they are size-independent, i.e., do not need to be tailored to a particular instance size of the pancake problem. Second, they are more compact in that they can represent a larger number of pancakes within abstractions of bounded size. Finally, they can exploit symmetries in the problem specification to allow multiple heuristic lookups, significantly improving search performance over a single lookup. Our experiments show that compared to object-location abstractions, our new techniques lead to an improvement of one order of magnitude in runtime and up to three orders of magnitude in the number of generated states.

We empirically examine several ways of exploiting the information of multiple heuristics in a satisficing best-first search algorithm, comparing their performance in terms of coverage, plan quality, speed, and search guidance. Our results indicate that using multiple heuristics for satisficing search is indeed useful. Among the combination methods we consider, the best results are obtained by the alternation method of the "Fast Diagonally Downward" planner.

The desire for persistent autonomy in ambitious robotic space missions has seen renewed interest in the use of on-board planning. This poses particular challenges in terms of the limited computational power of radiation-hardened hardware, and also the need to integrate with functions external to the planner itself. In this paper, we discuss the development of a forward-chaining temporal planner for use in this setting. We present results illustrating its lightweight resource footprint compared to contemporary planners, and its application to two proposed robotic space scenarios.

The European Robotic Goal-Oriented Autonomous Controller (ERGO) (http://www.h2020-ergo.eu) is one of the PERASPERA SRC first call projects. The focus of this project is to develop a framework for long-range autonomy that allows commanding a spacecraft via high-level goals. The developed ERGO framework achieves this aim by providing a new paradigm based on components and tools. Two main components, that include Artificial Intelligence technology, are at the core of ERGO: an on-board mission planner, able to dynamically generate plans on-board from high-level goals, and a scientific agent, that detects targets of interest from images. Additionally, other autonomous capabilities that can be used and tailored for different robotics platforms, such as rover navigation and robotic arm motion planning, are provided by ERGO. In this paper we discuss the results and performances achieved in the field tests of two applications developed with the ERGO framework: an orbital and a planetary mission.

ERGO (European Robotic Goal-Oriented Autonomous Controller) is one of the six space robotic projects in the frame of the first call of the PERASPERA SRC. ERGO is aimed to future space missions, in which space robots will require a higher level of autonomy (e.g. Exomars or Mars2020). As a framework, ERGO provides a set of components that can be reused and tailored for robots space missions (Orbital, Deep Space or Planetary Explorations) in which the on-board system has to work autonomously, performing complex operations in hazardous environments without human intervention. The concept of autonomy can be applied to a whole set of operations to be performed on-board with no human supervision, such as Martian exploration rovers, deep space probes, or in- orbit assembly robots. In the last decades, the advantages of increasing the level of autonomy in spacecraft have been demonstrated in planetary rovers. At the same time, orbital space missions have already successfully applied autonomy concepts on board, in particular for autonomous event detection and on-board activities planning.

ERGO provides a framework for on-board autonomy systems based on a specific paradigm aimed to facilitate an easy integration and/or expansion covering future mission needs; by using this paradigm, both reactive and deliberative capabilities can be orchestrated on-board. In ERGO, deliberative capabilities are provided via AI techniques: automated planning and machine-learning based vision systems. ERGO also provides a set of tools for developing safety-critical space mission applications and FDIR systems. Moreover, specific components for motion planning, path planning, hazard avoidance and trajectory control are also part of the framework. Finally, ERGO is integrated with the TASTE middleware. All ERGO components are now being tested in an orbital and a planetary scenario.

This paper discusses the ERGO components, its main characteristics, and how they have been applied to an orbital and a planetary scenario. It provides an overview of the evolution of the ERGO system; its main components and the future extensions planned for it.

The European Robotic Goal-Oriented Autonomous Controller ERGO is one of the six space robotic projects in the frame of the PERASPERA SRC. Its main objective is to provide an autonomous framework for future space robots that will be able to perform its activities without the need of constant human supervision and control. Future space missions, in particular those aimed at Deep Space or planetary exploration, such as Exomars, or Mars2020 demand a greater level of autonomy. The concept of autonomy applies here to a whole set of operations to be performed on-board without human supervision; for instance, a Martian rover has to avoid getting stuck in the sand while traversing, autonomously recharge its batteries periodically, and communicate with Earth occasionally each sol.

Additionally, it will need to be able to detect serendipitous events (e.g. a rock that has a specific property). A deep space probe has to take the right measurements to approach an asteroid, and due to the latency of the communication with Ground, these measurements need to be taken autonomously on board. Orbital space missions have already successfully applied autonomy concepts on board, in particular for autonomous event detection and on- board activities planning.

In ERGO we provide a framework for autonomy aimed to cover a wide set of a capabilities, ranging from reactive capabilities (i.e. capabilities that demand a quick response) to deliberative capabilities (that consider different courses of actions, and evaluate among the different possibilities the best alternative).

This paper will discuss the process of the design of robotic systems using the paradigm provided by this framework applied to two different scenarios: a Sample Fetching Rover (SFR), and also an On-Orbit Servicing mission, where a damaged spacecraft can have one or several of its modules replaced autonomously by a servicer spacecraft. We will describe the methodology, the main problems found, the design decisions taken to overcome these problems, as well as an overview of the final design of both systems.

The European Robotic Goal-Oriented Autonomous Controller ERGO (http://www.h2020-ergo.eu/) is one of the six space robotic projects in the frame of the PERASPERA SRC (http://www.h2020-peraspera.eu/). Its goal is to provide an Autonomy Framework capable of operating at different levels of autonomy, from tele- operations to full on-board autonomy. Even though it has been originally conceived for space robotics, its domain independent design facilitates its application to any terrestrial robotic system. This paper presents the approach followed, current status and future steps.

Cost partitioning is a general and principled approach for constructing additive admissible heuristics for state-space search. Cost partitioning approaches for optimal classical planning include optimal cost partitioning, uniform cost partitioning, zero-one cost partitioning, saturated cost partitioning, post-hoc optimization and the canonical heuristic for pattern databases. We compare these algorithms theoretically, showing that saturated cost partitioning dominates greedy zero-one cost partitioning. As a side effect of our analysis, we obtain a new cost partitioning algorithm dominating uniform cost partitioning. We also evaluate these algorithms experimentally on pattern databases, Cartesian abstractions and landmark heuristics, showing that saturated cost partitioning is usually the method of choice on the IPC benchmark suite.

Symmetry-based state space pruning techniques have proved to greatly improve heuristic search based classical planners. Similarly, abstraction heuristics in general and pattern databases in particular are key ingredients of such planners. However, only little work has dealt with how the abstraction heuristics behave under symmetries. In this work, we investigate the symmetry properties of the popular canonical pattern databases heuristic. Exploiting structural symmetries, we strengthen the canonical pattern databases by adding symmetric pattern databases, making the resulting heuristic invariant under structural symmetry, thus making it especially attractive for symmetry-based pruning search methods. Further, we prove that this heuristic is at least as informative as using symmetric lookups over the original heuristic. An experimental evaluation confirms these theoretical results.

Pattern databases are among the strongest known heuristics for many classical search benchmarks such as sliding-tile puzzles, the 4-peg Towers of Hanoi puzzles, Rubik's Cube, and TopSpin. Min-compression is a generally applicable technique for augmenting pattern database heuristics that has led to marked experimental improvements in some settings, while being ineffective in others. We provide a theoretical explanation for these experimental phenomena by studying the interaction between the ranking function used to order abstract states in a pattern database, the compression scheme used to abstract states, and the dependencies between state variables in the problem representation.

Symmetry-based state space pruning techniques have proved to greatly improve heuristic search based classical planners. Similarly, abstraction heuristics in general and pattern databases in particular are key ingredients of such planners. However, only little work has dealt with how the abstraction heuristics behave under symmetries. In this work, we investigate the symmetry properties of the popular canonical pattern databases heuristic. Exploiting structural symmetries, we strengthen the canonical pattern databases by adding symmetric pattern databases, making the resulting heuristic invariant under structural symmetry, thus making it especially attractive for symmetry-based pruning search methods. Further, we prove that this heuristic is at least as informative as using symmetric lookups over the original heuristic. An experimental evaluation confirms these theoretical results.

Cost partitioning is a general method for adding multiple heuristic values admissibly. In the setting of optimal classical planning, saturated cost partitioning has recently been shown to be the cost partitioning algorithm of choice for pattern database heuristics found by hill climbing, systematic pattern database heuristics and Cartesian abstraction heuristics. To evaluate the synergy of the three heuristic types, we compute the saturated cost partitioning over the combined sets of heuristics and observe that the resulting heuristic is outperformed by the heuristic that simply maximizes over the three saturated cost partitioning heuristics computed separately for each heuristic type. Our new algorithm for choosing the orders in which saturated cost partitioning considers the heuristics allows us to compute heuristics outperforming not only the maximizing heuristic but even state-of-the-art planners.

We propose techniques for efficiently determining optimal solutions to large logistics planning domain problems. We map a problem instance to a directed graph and show that no more than one vehicle per weakly connected component of the graph is needed for an optimal solution. We propose techniques for efficiently finding the vehicles which must be employed for an optimal solution. Also we develop a strong admissible heuristic based on the analysis of a directed graph, the cycles of which represent situations in the problem state in which a vehicle must visit a location more than once. To the best of our knowledge, ours is the first method that determines optimal solutions for large logistics instances (including the largest instances in the IPC 1998 and IPC 2000 problem sets).

One common pattern database compression technique is to merge adjacent database entries and store the minimum of merged entries to maintain heuristic admissibility. In this paper we propose a compression technique that preserves every entry, but reduces the number of bits used to store each entry, therefore limiting the values that can be represented. Even when this technique throws away low values in the heuristic, it can still have better performance than the traditional approach. We develop a theoretical basis for selecting which values to keep and show improved performance in both unidirectional and bidirectional search.

We analyze how complex a heuristic function must be to directly guide a state-space search algorithm towards the goal. As a case study, we examine functions that evaluate states with a weighted sum of state features. We measure the complexity of a domain by the complexity of the required features. We analyze conditions under which the search algorithm runs in polynomial time and show complexity results for several classical planning domains.

We propose an efficient method for determining optimal solutions to such skill-based solitaire card games as Freecell. We use A* search with an admissible heuristic function based on analyzing a directed graph whose cycles represent deadlock situations in the game state. To the best of our knowledge, ours is the first algorithm that efficiently determines optimal solutions for Freecell games. We believe that the underlying ideas should be applicable not only to games but also to other classical planning problems which manifest deadlocks.

The merge-and-shrink framework provides a general basis for the computation of abstraction heuristics for factored transition systems. Recent experimental and theoretical research demonstrated the utility of non-linear merge strategies, which have not been studied in depth. We experimentally analyze the quality of state-of-the-art merge strategies by comparing them to random strategies and with respect to tie-breaking, showing that there is considerable room for improvement. We finally describe a new merge strategy that experimentally outperforms the current state of the art.

We propose and implement an efficient method for determining optimal solutions to such skill-based solitaire card games as Freecell and King Albert solitaire. We use A* search together with an admissible heuristic function that is based on analyzing a directed graph whose cycles represent deadlock situations in the game state. To the best of our knowledge, ours is the first algorithm that efficiently determines optimal solutions for Freecell games. We believe that the underlying ideas should be applicable not only to games but also to other classical planning problems which manifest deadlocks.

We analyze how complex a heuristic function must be to directly guide a state-space search algorithm towards the goal. As a case study, we examine functions that evaluate states with a weighted sum of state features. We measure the complexity of a domain by the complexity of the required features. We analyze conditions under which the search algorithm runs in polynomial time and show complexity results for several classical planning domains.

Heuristic search is a state-of-the-art approach to classical planning. Several heuristic families were developed over the years to automatically estimate goal distance information from problem descriptions. Orthogonally to the development of better heuristics, recent years have seen an increasing interest in symmetry-based state space pruning techniques that aim at reducing the search effort. However, little work has dealt with how the heuristics behave under symmetries. We investigate the symmetry properties of existing heuristics and reveal that many of them are invariant under symmetries.

Pruning techniques have recently been shown to speed up search algorithms by reducing the branching factor of large search spaces. One such technique is sleep sets, which were originally introduced as a pruning technique for model checking, and which have recently been investigated on a theoretical level for planning. In this paper, we propose a generalization of sleep sets and prove its correctness. While the original sleep sets were based on the commutativity of operators, generalized sleep sets are based on a more general notion of operator sequence redundancy. As a result, our approach dominates the original sleep sets variant in terms of pruning power. On a practical level, our experimental evaluation shows the potential of sleep sets and their generalizations on a large and common set of planning benchmarks.

Bounded intention planning provides a pruning technique for optimal planning that has been proposed several years ago. In addition, partial order reduction techniques based on stubborn sets have recently been investigated for this purpose. In this paper, we revisit bounded intention planning in the view of stubborn sets.

A crucial problem in GUI testing is the identification of accurate event sequences that encode corresponding user interactions with the GUI. Ultimately, event sequences should be both feasible (i.e., executable on the GUI) and relevant (i.e., cover as much of the code as possible). So far, most work on GUI testing focused on approaches to generate feasible event sequences. In addition, based on event dependency analyses, a recently proposed static analysis approach systematically aims at selecting both relevant and feasible event sequences. However, statically analysing event dependencies can cause the generation of a huge number of event sequences, leading to unmanageable GUI test suites that are not executable within reasonable time.

In this paper we propose a refined static analysis approach based on program slicing. On the theoretical side, our approach identifies and eliminates redundant event sequences in GUI test suites. Redundant event sequences have the property that they are guaranteed to not affect the test effectiveness. On the practical side, we have implemented a slicing-based test suite reduction algorithm that approximatively identifies redundant event sequences. Our experiments on six open source GUI applications show that our reduction algorithm significantly reduces the size of GUI test suites. As a result, the overall execution time could significantly be reduced without losing test effectiveness.

Strong stubborn sets have recently been analyzed and successfully applied as a pruning technique for planning as heuristic search. Strong stubborn sets are defined declaratively as constraints over operator sets. We show how these constraints can be relaxed to offer more freedom in choosing stubborn sets while maintaining the correctness and optimality of the approach. In general, many operator sets satisfy the definition of stubborn sets. We study different strategies for selecting among these possibilities and show that existing approaches can be considerably improved by rather simple strategies, eliminating most of the overhead of the previous state of the art.

A general and important problem of search-based planning techniques is the state explosion problem, which is usually tackled with approaches to reduce the branching factor of the planning task. Such approaches often implicitly exploit the observation that the number of available operators is higher than the number of operators that are actually needed to find a plan. In this paper, we propose a simple, but general under-approximation refinement framework for satisficing planning that explicitly exploits this observation. Our approach iteratively searches for plans with operator subsets, which are refined if necessary by adding operators that appear to be needed. Our evaluation shows that even a straight-forward instantiation of this framework yields a competitive planner that often finds plans with small operator sets.

In the last years, pruning techniques based on partial order reduction have found increasing attention in the planning community. One recent result is that the expansion core method is a special case of the strong stubborn sets method proposed in model checking. However, it is still an open question if there exist efficiently computable strong stubborn sets with strictly higher pruning power than expansion core. In this paper, we prove that the pruning power of strong stubborn sets is strictly higher than the pruning power of expansion core even for a straight-forward instantiation of strong stubborn sets. This instantiation is as efficiently computable as expansion core. Hence, our theoretical results suggest that strong stubborn sets should be preferred to expansion core. Our empirical evaluation on all optimal benchmarks from the international planning competitions up to 2011 supports the theoretical results.

As classical planning is known to be computationally hard, no single planner is expected to work well across many planning domains. One solution to this problem is to use online portfolio planners that select a planner for a given task. These portfolios perform a classification task, a well-known and well-researched task in the field of machine learning. The classification is usually performed using a representation of planning tasks with a collection of hand-crafted statistical features. Recent techniques in machine learning that are based on automatic extraction of features have not been employed yet due to the lack of suitable representations of planning tasks. In this work, we alleviate this barrier. We suggest representing planning tasks by images, allowing to exploit arguably one of the most commonly used and best developed techniques in deep learning. We explore some of the questions that inevitably rise when applying such a technique, and present various ways of building practically useful online portfolio-based planners. An evidence of the usefulness of our proposed technique is a planner that won the cost-optimal track of the International Planning Competition 2018.

Generalized planning is concerned with the computation of plans that solve not one but multiple instances of a planning domain. Recently, it has been shown that generalized plans can be expressed as mappings of feature values into actions, and that they can often be computed with fully observable non-deterministic (FOND) planners. The actions in such plans, however, are not the actions in the instances themselves, which are not necessarily common to other instances, but abstract actions that are defined on a set of common features. The formulation assumes that the features and the abstract actions are given. In this work, we address this limitation by showing how to learn them automatically. The resulting account of generalized planning combines learning and planning in a novel way: a learner, based on a Max SAT formulation, yields the features and abstract actions from sampled state transitions, and a FOND planner uses this information, suitably transformed, to produce the general plans. Correctness guarantees are given and experimental results on several domains are reported.

We study the impact of tie-breaking on the behavior of greedy best-first search with a fixed state space and fixed heuristic. We prove that it is NP-complete to determine the number of states that need to be expanded by greedy best-first search in the best case or in the worst case. However, the best- and worst-case behavior can be computed in polynomial time for undirected state spaces. We perform computational experiments on benchmark tasks from the International Planning Competitions that compare the best and worst cases of greedy best-first search to FIFO, LIFO and random tie-breaking. The experiments demonstrate the importance of tie-breaking in greedy best-first search.

If a planning system outputs a solution for a given problem, it is simple to verify that the solution is valid. However, if a planner claims that a task is unsolvable, we currently have no choice but to trust the planner blindly. We propose a sound and complete class of certificates of unsolvability, which can be verified efficiently by an independent program. To highlight their practical use, we show how these certificates can be generated for a wide range of state-of-the-art planning techniques with only polynomial overhead for the planner.

A classical result in optimal search shows that A* with an admissible and consistent heuristic expands every state whose f-value is below the optimal solution cost and no state whose f-value is above the optimal solution cost. For satisficing search algorithms, a similarly clear understanding is currently lacking. We examine the search behavior of greedy best-first search (GBFS) in order to make progress towards such an understanding.

Merge-and-shrink heuristics are a successful class of abstraction heuristics used for optimal classical planning. With the recent addition of generalized label reduction, merge-and-shrink can be understood as an algorithm framework that repeatedly applies transformations to a factored representation of a given planning task to compute an abstraction. In this paper, we describe an efficient implementation of the framework and its transformations, comparing it to its previous implementation in Fast Downward. We further discuss partial merge-and-shrink abstractions that do not consider all aspects of the concrete state space. To compute such partial abstractions, we stop the merge-and-shrink computation early by imposing simple limits on the resource consumption of the algorithm. Our evaluation shows that the efficient implementation indeed improves over the previous one, and that partial merge-and-shrink abstractions further push the efficiency of merge-and-shrink planners.

Counterexample-guided abstraction refinement (CEGAR) is a method for incrementally computing abstractions of transition systems. We propose a CEGAR algorithm for computing abstraction heuristics for optimal classical planning. Starting from a coarse abstraction of the planning task, we iteratively compute an optimal abstract solution, check if and why it fails for the concrete planning task and refine the abstraction so that the same failure cannot occur in future iterations. A key ingredient of our approach is a novel class of abstractions for classical planning tasks that admits efficient and very fine-grained refinement. Since a single abstraction usually cannot capture enough details of the planning task, we also introduce two methods for producing diverse sets of heuristics within this framework, one based on goal atoms, the other based on landmarks. In order to sum their heuristic estimates admissibly we introduce a new cost partitioning algorithm called saturated cost partitioning. We show that the resulting heuristics outperform other state-of-the-art abstraction heuristics in many benchmark domains.

While traditionally classical planning concentrated on finding plans for solvable tasks, detecting unsolvable instances has recently attracted increasing interest. To preclude wrong results, it is desirable that the planning system provides a certificate of unsolvability that can be independently verified. We propose a rule-based proof system for unsolvability where a proof establishes a knowledge base of verifiable basic statements and applies a set of derivation rules to infer the unsolvability of the task from these statements. We argue that this approach is more flexible than a recent proposal of inductive certificates of unsolvability and show how our proof system can be used for a wide range of planning techniques.

Relaxed reachability analysis is relevant to efficient grounding, invariant synthesis as well as the computation of relaxation-based heuristics. Planning domains are typically specified in a lifted representation, where the size of the tasks grows exponentially with the number of objects in the world. This growth also affects the analysis of relaxed reachability. We present a task reduction based on symmetries of the lifted representation that allows to perform the same analysis on smaller tasks.

While it is easy to verify that an action sequence is a solution for a classical planning task, there is no such verification capability if a task is reported unsolvable. We are therefore interested in certifi- cates that allow an independent verification of the absence of solutions. We identify promising con- cepts for certificates that can be generated by a wide range of planning approaches. We present a first proposal of unsolvability certificates and sketch ideas how the underlying concepts can be used as part of a more flexible unsolvability proof system.

The plans that planning systems generate for solvable planning tasks are routinely verified by independent validation tools. For unsolvable planning tasks, no such validation capabilities currently exist. We describe a family of certificates of unsolvability for classical planning tasks that can be efficiently verified and are sufficiently general for a wide range of planning approaches including heuristic search with delete relaxation, critical-path, pattern database and linear merge-and-shrink heuristics, symbolic search with binary decision diagrams, and the Trapper algorithm for detecting dead ends. We also augmented a classical planning system with the ability to emit certificates of unsolvability and implemented a planner-independent certificate validation tool. Experiments show that the overhead for producing such certificates is tolerable and that their validation is practically feasible.

Cost partitioning is a well-known technique to make admissible heuristics for classical planning additive. The optimal cost partitioning of explicit-state abstraction heuristics can be computed in polynomial time with a linear program, but the size of the model is often prohibitive. We study this model from a dual perspective and develop several simplification rules to reduce its size. We use these rules to answer open questions about extensions of the state equation heuristic and their relation to cost partitioning.

A classical result in optimal search shows that A* with an admissible and consistent heuristic expands every state whose f-value is below the optimal solution cost and no state whose f-value is above the optimal solution cost. For satisficing search algorithms, a similarly clear understanding is currently lacking. We examine the search behaviour of greedy best-first search (GBFS) in order to make progress towards such an understanding. We introduce the concept of high-water mark benches, which separate the search space into areas that are searched by a GBFS algorithm in sequence. High-water mark benches allow us to exactly determine the set of states that are not expanded under any GBFS tie-breaking strategy. For the remaining states, we show that some are expanded by all GBFS searches, while others are expanded only if certain conditions are met.

Potential heuristics for state-space search are defined as weighted sums over simple state features. Atomic features consider the value of a single state variable in a factored state representation, while binary features consider joint assignments to two state variables. Previous work showed that the set of all admissible and consistent potential heuristics using atomic features can be characterized by a compact set of linear constraints. We generalize this result to binary features and prove a hardness result for features of higher dimension. Furthermore, we prove a tractability result based on the treewidth of a new graphical structure we call the context-dependency graph. Finally, we study the relationship of potential heuristics to transition cost partitioning. Experimental results show that binary potential heuristics are significantly more informative than the previously considered atomic ones.

In classical planning, cost partitioning is a method for admissibly combining a set of heuristic estimators by distributing operator costs among the heuristics. An optimal cost partitioning is often prohibitively expensive to compute. Saturated cost partitioning is an alternative that is much faster to compute and has been shown to offer high-quality heuristic guidance on Cartesian abstractions. However, its greedy nature makes it highly susceptible to the order in which the heuristics are considered. We show that searching in the space of orders leads to significantly better heuristic estimates than with previously considered orders. Moreover, using multiple orders leads to a heuristic that is significantly better informed than any single-order heuristic. In experiments with Cartesian abstractions, the resulting heuristic approximates the optimal cost partitioning very closely.

Abstraction heuristics are a popular method to guide optimal search algorithms in classical planning. Cost partitionings allow to sum heuristic estimates admissibly by distributing action costs among the heuristics. We introduce state-dependent cost partitionings which take context information of actions into account, and show that an optimal state-dependent cost partitioning dominates its state-independent counterpart. We demonstrate the potential of our idea with a state-dependent variant of the recently proposed saturated cost partitioning, and show that it has the potential to improve not only over its state-independent counterpart, but even over the optimal state-independent cost partitioning. Our empirical results give evidence that ignoring the context of actions in the computation of a cost partitioning leads to a significant loss of information.

We suggest the use of linear temporal logic (LTL) for expressing declarative information about optimal solutions of search problems. We describe a general framework that associates LTLf formulas with search nodes in a heuristic search algorithm. Compared to previous approaches that integrate specific kinds of path information like landmarks into heuristic search, the approach is general, easy to prove correct and easy to integrate with other kinds of path information.