Probabilistic plan evaluation in a classical hierarchical planner

The authors describe an approach to probabilistic plan evaluation that has been implemented as part of a hierarchical adversarial planning system. Probability evaluations occur at multiple levels of abstraction. Probability evaluations at lower levels of abstraction are used to update probability evaluations at higher levels. The impact of an adversary´s possible actions are considered in the probability calculations. Probability evaluations are a function of the planning agent´s assessment of its own and the adversary´s planning capability. This approach separates the probability that a plan will succeed from the probability that a plan will lead to success. The latter is based on estimates of a planning agent´s ability to adopt a plan to unanticipated events