An evaluator specification is either a newly created evaluator instance or an evaluator that has been defined previously. This page describes how one can specify a new evaluator instance. For re-using evaluators, see Evaluator Predefinitions.

If the evaluator is a heuristic, definitions of properties in the descriptions below:

This plugin type can be predefined using --evaluator. The old predefinition key --heuristic is still supported but deprecated.

Additive heuristic

add(transform=no_transform(), cache_estimates=true)

Language features supported:

Properties:

Potential heuristic optimized for all states

The algorithm is based on

all_states_potential(max_potential=1e8, lpsolver=CPLEX, transform=no_transform(), cache_estimates=true)

Note: to use an LP solver, you must build the planner with LP support. See LPBuildInstructions.

Language features supported:

Properties:

Blind heuristic

Returns cost of cheapest action for non-goal states, 0 for goal states

blind(transform=no_transform(), cache_estimates=true)

Language features supported:

Properties:

Context-enhanced additive heuristic

cea(transform=no_transform(), cache_estimates=true)

Language features supported:

Properties:

Additive CEGAR heuristic

See the paper introducing Counterexample-guided Abstraction Refinement (CEGAR) for classical planning:

and the paper showing how to make the abstractions additive:

cegar(subtasks=[landmarks(),goals()], max_states=infinity, max_transitions=1000000, max_time=infinity, pick=MAX_REFINED, use_general_costs=true, debug=false, transform=no_transform(), cache_estimates=true, random_seed=-1)

Language features supported:

Properties:

Causal graph heuristic

cg(transform=no_transform(), cache_estimates=true)

Language features supported:

Properties:

Diverse potential heuristics

The algorithm is based on

diverse_potentials(num_samples=1000, max_num_heuristics=infinity, max_potential=1e8, lpsolver=CPLEX, transform=no_transform(), cache_estimates=true, random_seed=-1)

Note: to use an LP solver, you must build the planner with LP support. See LPBuildInstructions.

Language features supported:

Properties:

FF heuristic

ff(transform=no_transform(), cache_estimates=true)

Language features supported:

Properties:

Goal count heuristic

goalcount(transform=no_transform(), cache_estimates=true)

Language features supported:

Properties:

h^m heuristic

hm(m=2, transform=no_transform(), cache_estimates=true)

Language features supported:

Properties:

Max heuristic

hmax(transform=no_transform(), cache_estimates=true)

Language features supported:

Properties:

Potential heuristic optimized for initial state

The algorithm is based on

initial_state_potential(max_potential=1e8, lpsolver=CPLEX, transform=no_transform(), cache_estimates=true)

Note: to use an LP solver, you must build the planner with LP support. See LPBuildInstructions.

Language features supported:

Properties:

Landmark-count heuristic

For the inadmissible variant see the papers

and

For the admissible variant see the papers

and

lmcount(lm_factory, admissible=false, optimal=false, pref=false, alm=true, lpsolver=CPLEX, transform=no_transform(), cache_estimates=true)

Optimal search: When using landmarks for optimal search (admissible=true), you probably also want to add this heuristic as a lazy_evaluator in the A* algorithm to improve heuristic estimates.

Note: To use optimal=true, you must build the planner with LP support. See LPBuildInstructions.

Differences to the literature: This heuristic differs from the description in the literature (see references above) in the set of preferred operators computed. The original implementation described in the literature computes two kinds of preferred operators:

  1. If there is an applicable operator that reaches a landmark, all such operators are preferred.
  2. If no such operators exist, perform an FF-style relaxed exploration towards the nearest landmarks (according to the landmark orderings) and use the preferred operators of this exploration.

Our implementation of the heuristic only considers preferred operators of the first type and does not include the second type. The rationale for this change is that it reduces code complexity and helps more cleanly separate landmark-based and FF-based computations in LAMA-like planner configurations. In our experiments, only considering preferred operators of the first type reduces performance when using the heuristic and its preferred operators in isolation but improves performance when using this heuristic in conjunction with the FF heuristic, as in LAMA-like planner configurations.

Note: to use an LP solver, you must build the planner with LP support. See LPBuildInstructions.

Language features supported:

Properties:

Landmark-cut heuristic

lmcut(transform=no_transform(), cache_estimates=true)

Language features supported:

Properties:

Max evaluator

Calculates the maximum of the sub-evaluators.

max(evals)

Merge-and-shrink heuristic

This heuristic implements the algorithm described in the following paper:

For a more exhaustive description of merge-and-shrink, see the journal paper

Please note that the journal paper describes the "old" theory of label reduction, which has been superseded by the above conference paper and is no longer implemented in Fast Downward.

The following paper describes how to improve the DFP merge strategy with tie-breaking, and presents two new merge strategies (dyn-MIASM and SCC-DFP):

merge_and_shrink(transform=no_transform(), cache_estimates=true, merge_strategy, shrink_strategy, label_reduction=<none>, prune_unreachable_states=true, prune_irrelevant_states=true, max_states=-1, max_states_before_merge=-1, threshold_before_merge=-1, verbosity=verbose)

Note: Conditional effects are supported directly. Note, however, that for tasks that are not factored (in the sense of the JACM 2014 merge-and-shrink paper), the atomic transition systems on which merge-and-shrink heuristics are based are nondeterministic, which can lead to poor heuristics even when only perfect shrinking is performed.

Note: When pruning unreachable states, admissibility and consistency is only guaranteed for reachable states and transitions between reachable states. While this does not impact regular A* search which will never encounter any unreachable state, it impacts techniques like symmetry-based pruning: a reachable state which is mapped to an unreachable symmetric state (which hence is pruned) would falsely be considered a dead-end and also be pruned, thus violating optimality of the search.

Note: A currently recommended good configuration uses bisimulation based shrinking, the merge strategy SCC-DFP, and the appropriate label reduction setting (max_states has been altered to be between 10000 and 200000 in the literature):

merge_and_shrink(shrink_strategy=shrink_bisimulation(greedy=false),merge_strategy=merge_sccs(order_of_sccs=topological,merge_selector=score_based_filtering(scoring_functions=[goal_relevance,dfp,total_order])),label_reduction=exact(before_shrinking=true,before_merging=false),max_states=50000,threshold_before_merge=1)

Note that for versions of Fast Downward prior to 2016-08-19, the syntax differs. See the recommendation in the file merge_and_shrink_heuristic.cc for an example configuration.

Language features supported:

Properties:

Operator counting heuristic

An operator counting heuristic computes a linear program (LP) in each state. The LP has one variable Count_o for each operator o that represents how often the operator is used in a plan. Operator counting constraints are linear constraints over these varaibles that are guaranteed to have a solution with Count_o = occurrences(o, pi) for every plan pi. Minimizing the total cost of operators subject to some operator counting constraints is an admissible heuristic. For details, see

operatorcounting(constraint_generators, lpsolver=CPLEX, transform=no_transform(), cache_estimates=true)

Note: to use an LP solver, you must build the planner with LP support. See LPBuildInstructions.

Language features supported:

Properties:

Sample-based potential heuristics

Maximum over multiple potential heuristics optimized for samples. The algorithm is based on

sample_based_potentials(num_heuristics=1, num_samples=1000, max_potential=1e8, lpsolver=CPLEX, transform=no_transform(), cache_estimates=true, random_seed=-1)

Note: to use an LP solver, you must build the planner with LP support. See LPBuildInstructions.

Language features supported:

Properties:

Basic Evaluators

Constant evaluator

Returns a constant value.

const(value=1)

g-value evaluator

Returns the g-value (path cost) of the search node.

g()

Preference evaluator

Returns 0 if preferred is true and 1 otherwise.

pref()

Sum evaluator

Calculates the sum of the sub-evaluators.

sum(evals)

Weighted evaluator

Multiplies the value of the evaluator with the given weight.

weight(eval, weight)

Pattern Database Heuristics

Canonical PDB

The canonical pattern database heuristic is calculated as follows. For a given pattern collection C, the value of the canonical heuristic function is the maximum over all maximal additive subsets A in C, where the value for one subset S in A is the sum of the heuristic values for all patterns in S for a given state.

cpdbs(patterns=systematic(1), max_time_dominance_pruning=infinity, transform=no_transform(), cache_estimates=true)

Language features supported:

Properties:

iPDB

This pattern generation method is an adaption of the algorithm described in the following paper:

For implementation notes, see:

ipdb(pdb_max_size=2000000, collection_max_size=20000000, num_samples=1000, min_improvement=10, max_time=infinity, random_seed=-1, max_time_dominance_pruning=infinity, transform=no_transform(), cache_estimates=true)

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 uses the canonical pattern collection 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.

Language features supported:

Properties:

Pattern database heuristic

TODO

pdb(pattern=greedy(), transform=no_transform(), cache_estimates=true)

Language features supported:

Properties:

Zero-One PDB

The zero/one pattern database heuristic is simply the sum of the heuristic values of all patterns in the pattern collection. In contrast to the canonical pattern database heuristic, there is no need to check for additive subsets, because the additivity of the patterns is guaranteed by action cost partitioning. This heuristic uses the most simple form of action cost partitioning, i.e. if an operator affects more than one pattern in the collection, its costs are entirely taken into account for one pattern (the first one which it affects) and set to zero for all other affected patterns.

zopdbs(patterns=systematic(1), transform=no_transform(), cache_estimates=true)

Language features supported:

Properties:

FastDownward: Doc/Evaluator (last edited 2018-12-13 22:14:51 by XmlRpcBot)