AI-Pathfinding-SMAstar version 0.05 =================================== NAME AI::Pathfinding::SMAstar - Memory-bounded A* Search SYNOPSIS use AI::Pathfinding::SMAstar; EXAMPLE ################################################################## # # This example uses a hypothetical object called FrontierObj, and # shows the functions that FrontierObj must feature in order to # perform a path search in a solution-space populated by # FrontierObj objects. # ################################################################## my $smastar = AI::Pathfinding::SMAstar->new( # evaluates f(n) = g(n) + h(n), returns a number _state_eval_func => \&FrontierObj::evaluate, # when called on a node, returns 1 if it is a goal _state_goal_p_func => \&FrontierObj::goal_test, # must return the number of successors of a node _state_num_successors_func => \&FrontierObj::get_num_successors, # must return *one* successor at a time _state_successors_iterator => \&FrontierObj::get_successors_iterator, # can be any suitable string representation _state_get_data_func => \&FrontierObj::string_representation, # gets called once per iteration, useful for showing algorithm progress _show_prog_func => \&FrontierObj::progress_callback, ); # you can start the search from multiple start-states # Add the initial states to the smastar object before starting the search. foreach my $frontierObj (@start_states){ $smastar->add_start_state($frontierObj); } # Start the search. If successful, frontierGoalObj will contain the # goal node. The optimal path to the goal node will be encoded in the # ancestry of the goal node. $frontierGoalObj->antecedent() contains # the goal object's parent, and so forth back to the start state. my $frontierGoalObj = $smastar->start_search( \&log_function, # returns a string used for logging progress \&str_function, # returns a string used to *uniquely* identify a node $max_states_in_queue, # indicate the maximum states allowed in memory $MAX_COST, # indicate the maximum cost allowed in search ); Explanation In the example above, a hypothetical object, FrontierObj, is used to represent a node in your search space. To use SMA* search to find a shortest path from a starting node to a goal in your search space, you must define what a node is, in your search space (or point, or state). A common example used for informed search methods, and one that is used in Russell's original paper, is a N-puzzle, such as an 8-puzzle or 15-puzzle. If trying to solve such a puzzle, a node in the search space could be defined as a particular configuration of that puzzle. In the /t directory of this module's distribution, SMA* is applied to the problem of finding the shortest palindrome that contains a minimum number of letters specified, over a given lexicon of words. Once you have a definition and representation of a node in your search space, SMA* search requires the following functions to work: ** State evaluation function (_state_eval_func above) This function must return the cost of this node in the search space. In all forms of A* search, this means the cost paid to arrive at this node along a path, plus the estimated cost of going from this node to a goal state. This function must be positive and monotonic, meaning that successor nodes mustn't be less expensive than their antecedent nodes. Monotonicity is ensured in this implementation of SMA*, so even if your function is not monotonic, SMA* will assign the antecedent node's cost to a successor if that successor costs less than the antecedent. * State goal predicate function (_state_goal_p_func above) This function must return 1 if the node is a goal node, or 0 otherwise. * State number of successors function (_state_num_successors_func above) This function must return the number of successors of this node, i.e. all nodes that are reachable from this node via a single operation. * State successors iterator (_state_iterator above) This function must return a *handle to a function* that returns next successor of this node, i.e. it must return an iterator that produces the successors of this node *one* at a time. This is necessary to maintain the memory-bounded constraint of SMA* search. * State get-data function (_state_get_data_func above) This function returns a string representation of this node. * State show-progress function (_show_prog_func above) This is a callback function for displaying the progress of the search. It can be an empty callback if you do not need this output. * log string function (log_function above) This is an arbitrary string used for logging. It also gets passed to the show-progress function above. * str_function (str_function above) This function returns a *unique* string representation of this node. Uniqueness is required for SMA* to work properly. * max states allowed in memory (max_states_in_queue above) An integer indicating the maximum number of expanded nodes to hold in memory at any given time. * maximum cost (MAX_COST above) An integer indicating the maximum cost, beyond which nodes will not be expanded. DESCRIPTION Overview Memory-bounded A* search (or SMA* search) addresses some of the limitations of conventional A* search, by bounding the amount of space required to perform a shortest-path search. This module is an implementation of SMA*, which was first introduced by Stuart Russell in 1992. SMA* is a more efficient variation of the original MA* search introduced by Chakrabarti et al. in 1989. See references below. Motivation and Comparison to A* Search A* search A* Search is an optimal and complete algorithm for computing a sequence of operations leading from a system's start-state (node) to a specified goal. In this context, optimal means that A* search will return the shortest possible sequence of operations (path) leading to the goal, and complete means that A* will always find a path to the goal if such a path exists. In general, A* search works using a calculated cost function on each node along a path, in addition to an admissible heuristic estimating the distance from that node to the goal. The cost is calculated as: f(n) = g(n) + h(n) Where: * n is a state (node) along a path * g(n) is the total cost of the path leading up to n * h(n) is the heuristic function, or estimated cost of the path from n to the goal node. The to be admissible, the heuristic must never over-estimate the distance from the node to the goal. If the heuristic is set to zero, A* search reduces to Branch and Bound search. For a given heuristic function, it can be shown that A* search is optimally efficient, meaning that, in its calculation of the shortest path, it expands fewer nodes in the search space than any other algorithm. The space complexity of A* search is bounded by an exponential of the branching factor of the search-space and the length of the longest path examined during the search. This is can be a problem particularly if the branching factor is large, as the algorithm may run out of memory. SMA* Search SMA* search addresses the possibility of running out of memory during search by pruning the portion of the search-space that is being examined. It relies on the pathmax, or monotonicity constraint on f(n) to remove the shallowest of the highest-cost nodes from the search queue when there is no memory left to expand new nodes. It records the best costs of the pruned nodes within their antecedent nodes to ensure that crucial information about the search space is not lost. To facilitate this mechanism, the search queue is best maintained as a search-tree of search-trees ordered by cost and depth, respectively. The pruning of the search queue allows SMA* search to utilize all available memory for search without any danger of overflow. It can, however, make SMA* search significantly slower than a theoretical unbounded-memory search, due to the extra bookkeeping it must do, and because nodes may need to be re-expanded (the overall number of node expansions may increase). It can be shown that of the memory-bounded variations of A* search, such MA*, IDA*, Iterative Expansion, etc., SMA* search expands the least number of nodes on average. However, for certain classes of problems, guaranteeing optimality can be costly. This is particularly true in solution spaces where: * the branching factor of the search space is large * there are multiple equivalent optimal solutions (or shortest paths) For solution spaces with these characteristics, stochastic methods or approximation algorithms such as Simulated Annealing can provide a massive reduction in time and space requirements, while introducing a tunable probability of producing a sub-optimal solution. METHODS new() Creates a new SMA* search object. start_search() Initiates a memory-bounded search. You must pass a log_function for recording current status, a function that returns a *unique* string representing a node in the search-space, a maximum number of expanded states to store in the queue, and a maximum cost value, beyond which the search will cease. state_eval_func() Sets/gets the function that returns the cost of this node in the search space. state_goal_p_func() Sets/gets the function that returns 1 if the node is a goal node, or 0 otherwise. state_num_successors_func() Sets/gets the function that returns the number of successors of this node. state_successors_iterator() Sets/gets the function that returns the next successor of this node. state_get_data_func() Sets/gets the function that returns a string representation of this node. show_prog_func() sets/gets the callback function for displaying the progress of the search. It can be an empty callback if you do not need this output. EXPORT None by default. SEE ALSO Russell, Stuart. (1992) "Efficient Memory-bounded Search Methods" Proceedings of the 10th European conference on Artificial intelligence, pp. 1-5 Chakrabarti, P. P., Ghose, S., Acharya, A., and de Sarkar, S. C. (1989) "Heuristic search in restricted memory" Artificial Intelligence Journal, 41, pp. 197-221. AUTHOR Matthias Beebe, INSTALLATION To install this module type the following: perl Makefile.PL make make test make install DEPENDENCIES This module requires these other modules and libraries: Tree::AVL Test::More COPYRIGHT AND LICENCE Copyright (C) 2010 by Matthias Beebe This library is free software; you can redistribute it and/or modify it under the same terms as Perl itself, either Perl version 5.10.0 or, at your option, any later version of Perl 5 you may have available.