Macro Learning in Planning as Parameter Configuration

摘要

In AI planning, macro learning is the task of finding subsequences of operators that can be added to the planning domain to improve planner performance. Typically, a single set is added to the domain for all problem instances. A number of techniques have been developed to generate such a macro set based on offline analysis of problem instances. We build on recent work on instance-specific and fixed-set macros, and recast the macro generation problem as parameter configuration: the macros in a domain are viewed as parameters of the planning problem. We then apply an existing parameter configuration system to reconfigure a domain either once or per problem instance. Our empirical results demonstrate that our approach outperforms existing macro acquisition and filtering tools. For instance-specific macros, our approach almost always achieves equal or better performance than a complete evaluation approach, while often being an order of magnitude faster offline.