Factory for pattern collections

combo

combo(max_states=1000000, verbosity=normal)

Disjoint CEGAR

This pattern collection generator uses the CEGAR algorithm to compute a pattern for the planning task. See below for a description of the algorithm and some implementation notes. The original algorithm (called single CEGAR) is described in the paper

disjoint_cegar(max_pdb_size=1000000, max_collection_size=10000000, max_time=infinity, use_wildcard_plans=true, verbosity=normal, random_seed=-1)

Short description of the CEGAR algorithm

The CEGAR algorithm computes a pattern collection for a given planning task and a given (sub)set of its goals in a randomized order as follows. Starting from the pattern collection consisting of a singleton pattern for each goal variable, it repeatedly attempts to execute an optimal plan of each pattern in the concrete task, collects reasons why this is not possible (so-called flaws) and refines the pattern in question by adding a variable to it. Further parameters allow blacklisting a (sub)set of the non-goal variables which are then never added to the collection, limiting PDB and collection size, setting a time limit and switching between computing regular or wildcard plans, where the latter are sequences of parallel operators inducing the same abstract transition.

Implementation notes about the CEGAR algorithm

The following describes differences of the implementation to the original implementation used and described in the paper.

Conceptually, there is one larger difference which concerns the computation of (regular or wildcard) plans for PDBs. The original implementation used an enforced hill-climbing (EHC) search with the PDB as the perfect heuristic, which ensured finding strongly optimal plans, i.e., optimal plans with a minimum number of zero-cost operators, in domains with zero-cost operators. The original implementation also slightly modified EHC to search for a best-improving successor, chosen uniformly at random among all best-improving successors.

In contrast, the current implementation computes a plan alongside the computation of the PDB itself. A modification to Dijkstra's algorithm for computing the PDB values stores, for each state, the operator leading to that state (in a regression search). This generating operator is updated only if the algorithm found a cheaper path to the state. After Dijkstra finishes, the plan computation starts at the initial state and iteratively follows the generating operator, computes all operators of the same cost inducing the same transition, until reaching a goal. This constitutes a wildcard plan. It is turned into a regular one by randomly picking a single operator for each transition.

Note that this kind of plan extraction does not consider all successors of a state uniformly at random but rather uses the previously deterministically chosen generating operator to settle on one successor state, which is biased by the number of operators leading to the same successor from the given state. Further note that in the presence of zero-cost operators, this procedure does not guarantee that the computed plan is strongly optimal because it does not minimize the number of used zero-cost operators leading to the state when choosing a generating operator. Experiments have shown (issue1007) that this speeds up the computation significantly while not having a strongly negative effect on heuristic quality due to potentially computing worse plans.

Two further changes fix bugs of the original implementation to match the description in the paper. The first bug fix is to raise a flaw for all goal variables of the task if the plan for a PDB can be executed on the concrete task but does not lead to a goal state. Previously, such flaws would not have been raised because all goal variables are part of the collection from the start on and therefore not considered. This means that the original implementation accidentally disallowed merging patterns due to goal violation flaws. The second bug fix is to actually randomize the order of parallel operators in wildcard plan steps.

Genetic Algorithm Patterns

The following paper describes the automated creation of pattern databases with a genetic algorithm. Pattern collections are initially created with a bin-packing algorithm. The genetic algorithm is used to optimize the pattern collections with an objective function that estimates the mean heuristic value of the the pattern collections. Pattern collections with higher mean heuristic estimates are more likely selected for the next generation.

genetic(pdb_max_size=50000, num_collections=5, num_episodes=30, mutation_probability=0.01, disjoint=false, random_seed=-1, verbosity=normal)

Note: This pattern generation method uses the zero/one pattern database heuristic.

Implementation Notes

The standard genetic algorithm procedure as described in the paper is implemented in Fast Downward. The implementation is close to the paper.

Supported language features:

Hill climbing

This algorithm uses hill climbing to generate patterns optimized for the Canonical PDB heuristic. It it described in the following paper:

For implementation notes, see:

hillclimbing(pdb_max_size=2000000, collection_max_size=20000000, num_samples=1000, min_improvement=10, max_time=infinity, random_seed=-1, verbosity=normal)

Note: The pattern collection created by the algorithm will always contain all patterns consisting of a single goal variable, even if this violates the pdb_max_size or collection_max_size limits.

Note: This pattern generation method generates patterns optimized for use with the canonical pattern database heuristic.

Implementation Notes

The following will very briefly describe the algorithm and explain the differences between the original implementation from 2007 and the new one in Fast Downward.

The aim of the algorithm is to output a pattern collection for which the Canonical PDB yields the best heuristic estimates.

The algorithm is basically a local search (hill climbing) which searches the "pattern neighbourhood" (starting initially with a pattern for each goal variable) for improving the pattern collection. This is done as described in the section "pattern construction as search" in the paper, except for the corrected search neighbourhood discussed below. For evaluating the neighbourhood, the "counting approximation" as introduced in the paper was implemented. An important difference however consists in the fact that this implementation computes all pattern databases for each candidate pattern rather than using A* search to compute the heuristic values only for the sample states for each pattern.

Also the logic for sampling the search space differs a bit from the original implementation. The original implementation uses a random walk of a length which is binomially distributed with the mean at the estimated solution depth (estimation is done with the current pattern collection heuristic). In the Fast Downward implementation, also a random walk is used, where the length is the estimation of the number of solution steps, which is calculated by dividing the current heuristic estimate for the initial state by the average operator costs of the planning task (calculated only once and not updated during sampling!) to take non-unit cost problems into account. This yields a random walk of an expected lenght of np = 2 * estimated number of solution steps. If the random walk gets stuck, it is being restarted from the initial state, exactly as described in the original paper.

The section "avoiding redundant evaluations" describes how the search neighbourhood of patterns can be restricted to variables that are relevant to the variables already included in the pattern by analyzing causal graphs. There is a mistake in the paper that leads to some relevant neighbouring patterns being ignored. See the errata for details. This mistake has been addressed in this implementation. The second approach described in the paper (statistical confidence interval) is not applicable to this implementation, as it doesn't use A* search but constructs the entire pattern databases for all candidate patterns anyway. The search is ended if there is no more improvement (or the improvement is smaller than the minimal improvement which can be set as an option), however there is no limit of iterations of the local search. This is similar to the techniques used in the original implementation as described in the paper.

manual_patterns

manual_patterns(patterns, verbosity=normal)

Multiple CEGAR

This pattern collection generator implements the multiple CEGAR algorithm described in the paper

It is an instantiation of the 'multiple algorithm framework'. To compute a pattern in each iteration, it uses the CEGAR algorithm restricted to a single goal variable. See below for descriptions of the algorithms.

multiple_cegar(max_pdb_size=1M, max_collection_size=10M, pattern_generation_max_time=infinity, total_max_time=100.0, stagnation_limit=20.0, blacklist_trigger_percentage=0.75, enable_blacklist_on_stagnation=true, verbosity=normal, random_seed=-1, use_wildcard_plans=true)

Short description of the CEGAR algorithm

The CEGAR algorithm computes a pattern collection for a given planning task and a given (sub)set of its goals in a randomized order as follows. Starting from the pattern collection consisting of a singleton pattern for each goal variable, it repeatedly attempts to execute an optimal plan of each pattern in the concrete task, collects reasons why this is not possible (so-called flaws) and refines the pattern in question by adding a variable to it. Further parameters allow blacklisting a (sub)set of the non-goal variables which are then never added to the collection, limiting PDB and collection size, setting a time limit and switching between computing regular or wildcard plans, where the latter are sequences of parallel operators inducing the same abstract transition.

Implementation notes about the CEGAR algorithm

The following describes differences of the implementation to the original implementation used and described in the paper.

Conceptually, there is one larger difference which concerns the computation of (regular or wildcard) plans for PDBs. The original implementation used an enforced hill-climbing (EHC) search with the PDB as the perfect heuristic, which ensured finding strongly optimal plans, i.e., optimal plans with a minimum number of zero-cost operators, in domains with zero-cost operators. The original implementation also slightly modified EHC to search for a best-improving successor, chosen uniformly at random among all best-improving successors.

In contrast, the current implementation computes a plan alongside the computation of the PDB itself. A modification to Dijkstra's algorithm for computing the PDB values stores, for each state, the operator leading to that state (in a regression search). This generating operator is updated only if the algorithm found a cheaper path to the state. After Dijkstra finishes, the plan computation starts at the initial state and iteratively follows the generating operator, computes all operators of the same cost inducing the same transition, until reaching a goal. This constitutes a wildcard plan. It is turned into a regular one by randomly picking a single operator for each transition.

Note that this kind of plan extraction does not consider all successors of a state uniformly at random but rather uses the previously deterministically chosen generating operator to settle on one successor state, which is biased by the number of operators leading to the same successor from the given state. Further note that in the presence of zero-cost operators, this procedure does not guarantee that the computed plan is strongly optimal because it does not minimize the number of used zero-cost operators leading to the state when choosing a generating operator. Experiments have shown (issue1007) that this speeds up the computation significantly while not having a strongly negative effect on heuristic quality due to potentially computing worse plans.

Two further changes fix bugs of the original implementation to match the description in the paper. The first bug fix is to raise a flaw for all goal variables of the task if the plan for a PDB can be executed on the concrete task but does not lead to a goal state. Previously, such flaws would not have been raised because all goal variables are part of the collection from the start on and therefore not considered. This means that the original implementation accidentally disallowed merging patterns due to goal violation flaws. The second bug fix is to actually randomize the order of parallel operators in wildcard plan steps.

Short description of the 'multiple algorithm framework'

This algorithm is a general framework for computing a pattern collection for a given planning task. It requires as input a method for computing a single pattern for the given task and a single goal of the task. The algorithm works as follows. It first stores the goals of the task in random order. Then, it repeatedly iterates over all goals and for each goal, it uses the given method for computing a single pattern. If the pattern is new (duplicate detection), it is kept for the final collection. The algorithm runs until reaching a given time limit. Another parameter allows exiting early if no new patterns are found for a certain time ('stagnation'). Further parameters allow enabling blacklisting for the given pattern computation method after a certain time to force some diversification or to enable said blacklisting when stagnating.

Implementation note about the 'multiple algorithm framework'

A difference compared to the original implementation used in the paper is that the original implementation of stagnation in the multiple CEGAR/RCG algorithms started counting the time towards stagnation only after having generated a duplicate pattern. Now, time towards stagnation starts counting from the start and is reset to the current time only when having found a new pattern or when enabling blacklisting.

Multiple Random Patterns

This pattern collection generator implements the 'multiple randomized causal graph' (mRCG) algorithm described in experiments of the paper

It is an instantiation of the 'multiple algorithm framework'. To compute a pattern in each iteration, it uses the random pattern algorithm, called 'single randomized causal graph' (sRCG) in the paper. See below for descriptions of the algorithms.

random_patterns(max_pdb_size=1M, max_collection_size=10M, pattern_generation_max_time=infinity, total_max_time=100.0, stagnation_limit=20.0, blacklist_trigger_percentage=0.75, enable_blacklist_on_stagnation=true, verbosity=normal, random_seed=-1, bidirectional=true)

Short description of the random pattern algorithm

The random pattern algorithm computes a pattern for a given planning task and a single goal of the task as follows. Starting with the given goal variable, the algorithm executes a random walk on the causal graph. In each iteration, it selects a random causal graph neighbor of the current variable. It terminates if no neighbor fits the pattern due to the size limit or if the time limit is reached.

Implementation notes about the random pattern algorithm

In the original implementation used in the paper, the algorithm selected a random neighbor and then checked if selecting it would violate the PDB size limit. If so, the algorithm would not select it and terminate. In the current implementation, the algorithm instead loops over all neighbors of the current variable in random order and selects the first one not violating the PDB size limit. If no such neighbor exists, the algorithm terminates.

Short description of the 'multiple algorithm framework'

This algorithm is a general framework for computing a pattern collection for a given planning task. It requires as input a method for computing a single pattern for the given task and a single goal of the task. The algorithm works as follows. It first stores the goals of the task in random order. Then, it repeatedly iterates over all goals and for each goal, it uses the given method for computing a single pattern. If the pattern is new (duplicate detection), it is kept for the final collection. The algorithm runs until reaching a given time limit. Another parameter allows exiting early if no new patterns are found for a certain time ('stagnation'). Further parameters allow enabling blacklisting for the given pattern computation method after a certain time to force some diversification or to enable said blacklisting when stagnating.

Implementation note about the 'multiple algorithm framework'

A difference compared to the original implementation used in the paper is that the original implementation of stagnation in the multiple CEGAR/RCG algorithms started counting the time towards stagnation only after having generated a duplicate pattern. Now, time towards stagnation starts counting from the start and is reset to the current time only when having found a new pattern or when enabling blacklisting.

Systematically generated patterns

Generates all (interesting) patterns with up to pattern_max_size variables. For details, see

systematic(pattern_max_size=1, only_interesting_patterns=true, verbosity=normal)

FastDownward: Doc/PatternCollectionGenerator (last edited 2024-01-11 22:26:38 by XmlRpcBot)