Learning neural search policies for classical planning

Heuristic forward search is currently the dominant paradigmin classical planning. Forward search algorithms typicallyrely on a single, relatively simple variation of best-first searchand remain fixed throughout the process of solving a plan-ning problem. Existing work combining multiple search tech-niques usually aims at supporting best-first search with anadditional exploratory mechanism, triggered using a hand-crafted criterion. A notable exception is very recent workwhich combines various search techniques using a trainablepolicy. That approach, however, is confined to a discrete ac-tion space comprising several fixed subroutines.In this paper, we introduce a parametrized search algorithmtemplate which combines various search techniques withina single routine. The template’s parameter space defines aninfinite space of search algorithms, including, among others,BFS, local and random search. We then propose a neural ar-chitecture for designating the values of the search parametersgiven the state of the search. This enables expressing neuralsearch policies that change the values of the parameters asthe search progresses. The policies can be learned automat-ically, with the objective of maximizing the planner’s per-formance on a given distribution of planning problems. Weconsider a training setting based on a stochastic optimizationalgorithm known as thecross-entropy method(CEM). Exper-imental evaluation of our approach shows that it is capable offinding effective distribution-specific search policies, outper-forming the relevant baselines