Using Plan Decomposition for Continuing Plan Optimisation and Macro Generation

Abstract

This thesis addresses three problems in the field of classical AI planning: decomposing a plan into meaningful subplans, continuing plan quality optimisation, and macro generation for efficient planning. The importance and difficulty of each of these problems is outlined below. (1) Decomposing a plan into meaningful subplans can facilitate a number of postplan generation tasks, including plan quality optimisation and macro generation – the two key concerns of this thesis. However, conventional plan decomposition techniques are often unable to decompose plans because they consider dependencies among steps, rather than subplans. (2) Finding high quality plans for large planning problems is hard. Planners that guarantee optimal, or bounded suboptimal, plan quality often cannot solve them In one experiment with the Genome Edit Distance domain optimal planners solved only 11.5% of problems. Anytime planners promise a way to successively produce better plans over time. However, current anytime planners tend to reach a limit where they stop finding any further improvement, and the plans produced are still very far from the best possible. In the same experiment, the LAMA anytime planner solved all problems but found plans whose average quality is 1.57 times worse than the best known. (3) Finding solutions quickly or even finding any solution for large problems within some resource constraint is also difficult. The best-performing planner in the 2014 international planning competition still failed to solve 29.3% of problems. Re-engineering a domain model by capturing and exploiting structural knowledge in the form of macros has been found very useful in speeding up planners. However, existing planner independent macro generation techniques often fail to capture some promising macro candidates because the constituent actions are not found in sequence in the totally ordered training plans. This thesis contributes to plan decomposition by developing a new plan deordering technique, named block deordering, that allows two subplans to be unordered even when their constituent steps cannot. Based on the block-deordered plan, this thesis further contributes to plan optimisation and macro generation, and their implementations in two systems, named BDPO2 and BloMa. Key to BDPO2 is a decomposition into subproblems of improving parts of the current best plan, rather than the plan as a whole. BDPO2 can be seen as an application of the large neighbourhood search strategy to planning. We use several windowing strategies to extract subplans from the block deordering of the current plan, and on-line learning for applying the most promising subplanners to the most promising subplans. We demonstrate empirically that even starting with the best plans found by other means, BDPO2 is still able to continue improving plan quality, and often produces