A framework for interleaving planning-while-learning and execution

Interacting with the environment in presence of incomplete information requires the ability to acquire new knowledge from the interaction with the environment and to employ it when deliberating about which actions to execute. The ability to identify a particular environmental behaviour by inspecting perceptual feedback greatly contributes to completing the knowledge available to the agent. This paper introduces a formal framework for interleaving planning-while-learning and execution in partially specified environments. Planning-while-learning combines conventional planning with the search of the environmental behaviour model that best fits the experienced behaviour of the environment. Heuristics for early termination of planning and assumptions are used in order to reduce the cost of planning. Sufficiency conditions are given that guarantee the soundness and the completeness of the agent's control system w.r.t. the environmental model and the goal.