Parallel Strategies for Best-First Generalized Planning

The paper focuses on parallelizing Best-First Generalized Planning (BFGP), a heuristic-based approach for solving multiple planning instances from the same domain. BFGP uses novel heuristics and a new solution space, independent of the number of input instances, to compute general algorithmic solutions. The authors propose two shared-memory parallel strategies that scale linearly with the number of cores, aiming to bridge the performance gap between state-of-the-art planning solvers and generalized planning. The first strategy sequentially expands nodes until there are at least N nodes per thread, then starts a parallel search where each thread is independent. The second strategy distributes promising nodes during the parallel search phase, balancing the tradeoff between searching promising states and minimizing communication overhead. The results show that BFGP is well-suited for parallelization, with potential for handling more complex problems from IPC planning domains. The paper also discusses various aspects of generalized planning, including methods for efficient navigation of the planning space, representation and synthesis of C++ programs, and a new planning search-space that leverages pointers over objects. The 2023 International Planning Competition highlights advancements and challenges in the field of generalized planning. In summary, the paper presents parallelization strategies for the Best-First Generalized Planning algorithm, demonstrating its suitability for handling complex planning problems and closing the performance gap between state-of-the-art planning solvers and generalized planning. The work contributes to the field of AI research focused on automated synthesis of algorithmic-like solutions for multiple planning instances, with potential for advancements in efficient navigation, representation, and synthesis of planning solutions.