Comparative criteria for partially observable contingent planning

In contingent planning under partial observability with sensing actions, agents actively use sensing to discover meaningful facts about the world. The solution can be represented as a plan tree or graph, branching on various possible observations. Typically in contingent planning one seeks a satisfying plan leading to a goal state at each leaf. In many applications, however, one may prefer some satisfying plans to others, such as plans that lead to the goal with a lower average cost. However, methods such as average cost make an implicit assumption concerning the probabilities of outcomes, which may not apply when the stochastic dynamics of the environment are unknown. We focus on the problem of providing valid comparative criteria for contingent plan trees and graphs, allowing us to compare two plans and decide which one is preferable. We suggest a set of such comparison criteria—plan simplicity, dominance, and best and worst plan costs.We also argue that in some cases certain branches of the plan correspond to an unlikely combination of mishaps, and can be ignored, and provide methods for pruning such unlikely branches before comparing the plan graphs. We explain these criteria, and discuss their validity, correlations, and application to real world problems. We also suggest efficient algorithms for computing the comparative criteria where needed. We provide experimental results, showing that existing contingent planners provide diverse plans, that can be compared using these criteria.