Interleaving execution and planning for nondeterministic, partially observable domains
Piergiorgio Bertoli, Alessandro Cimatti, Paolo Traverso · European Conference on Artificial Intelligence · 2004
Methods that interleave planning and execution are a practical solution to deal with complex planning problems in non-deterministic domains under partial observability. However, most of the existing approaches do not tackle in a principled way the important issue of termination of the planning-execution loop, or only do so considering specific assumptions over the domains. In this paper, we tackle the problem of interleaving planning and execution relying on a general framework, which is able to deal with nondeterministic, partially observable planning domains. We propose a new, general planning algorithm that guarantees the termination of the interleaving of planning and execution: either the goal is achieved, or the system detects that there is no longer a guarantee to progress toward it. Our experimental analysis shows that our algorithm can efficiently solve planning problems that cannot be tackled with a state of the art off-line planner for nondeterministic domains under partial observability, MBP. Moreover, we show that our algorithm can efficiently detect situations where progress toward the goal can be no longer guaranteed.