רון שטרן

Numeric planning addresses sequential decision-making in domains involving both discrete and continuous state variables. While effective, it requires a model of actions’ preconditions and effects, which might be unavailable or hard to model manually. Therefore, prior work developed algorithms that automatically learn numeric action models from execution traces. However, these approaches assumed that such traces are available, which may not hold in realistic settings where past interactions are unavailable. In this work, we introduce NOAM, the first online action model learning algorithm for numeric planning. NOAM can learn from both successful and failed observations. It iteratively refines two types of action models: a safe, risk-averse model and a more optimistic, exploratory action model. It performs goal-oriented exploration and prioritizes actions that are expected to be informative. We evaluated NOAM on several classical numeric benchmark domains and found that the models it learns enable solving most problems within those domains.