Online generation and use of macro-actions in forward-chaining planning

This thesis presents a technique for online learning and management of macroactions in forward chaining planning. Macro-actions are learnt on plateaux, areas of the search landscape where the heuristic cannot offer good search guidance, and are reused when future plateaux are encountered. Libraries of macro-actions are generated, storing macro-actions for use on future problems. Several strategies for the management and pruning of such libraries are considered allowing the planner to maintain a smaller collection of macro-actions to minimise potential negative effects on performance. The work is extended to investigate the potential for the simulation of macroactions without increasing the branching factor during search. Actions are reordered during enforced hill-climbing (EHC) search, and best-first search, based on the number of times they have followed the action currently at the tail of the plan in past solutions. In EHC this is equivalent to suggesting two-step macroactions without increasing the branching factor. The techniques are evaluated across a wide range of domains with differing properties under the relaxed planning graph heuristic. The results show that a library of plateau-escaping macro-actions can improve planner performance using only online learning techniques to select the best macro-actions to keep. Search time pruning is shown to be very effective; whilst pruning the library of macro-actions based on the number of times each action is used can offer further improvements. Action reordering has been shown to offer performance improvements, albeit not as great as those produced by using macro-actions, but without the occasional degradation in performance on some problems that is often associated with macro-actions.