Heuristic Guidance for Forward-Chaining Planning with Numeric Uncertainty

Abstract

Uncertainty hinders many interesting applications of planning - it may come in the form of sensor noise, unpredictable environments, or known limitations in problem models. In this paper we explore heuristic guidance for forward-chaining planning with continuous random variables, while ensuring a probability of plan success. We extend the Metric Relaxed Planning Graph heuristic to capture a model of uncertainty, providing better guidance in terms of heuristic estimates and dead-end detection. By tracking the accumulated error on numeric values, our heuristic is able to check if preconditions in the planning graph are achievable with a sufficient degree of confidence; it is also able to consider acting to reduce the accumulated error. Results indicate that our approach offers improvements in performance compared to prior work where a less-informed relaxation was used.