Learning Action Models for Reactive Autonomous Agents

To be maximally e ective, autonomous agents such as robots must be able both to react appropriately in dynamic environments and to plan new courses of action in novel situations. Reliable planning requires accurate models of the e ects of actions| models which are often more appropriately learned through experience than designed. This thesis describes TRAIL (Teleo-Reactive Agent with Inductive Learning), an integrated agent architecture which learns models of actions based on experiences in the environment. These action models are then used to create plans that combine both goal-directed and reactive behaviors. Previous work on action-model learning has focused on domains that contain only deterministic, atomic action models that explicitly describe all changes that can occur in the environment. The thesis extends this previous work to cover domains that contain durative actions, continuous variables, nondeterministic action e ects, and actions taken by other agents. Results have been demonstrated in several robot simulation environments and the Silicon Graphics, Inc. ight simulator. The main emphasis in this thesis is on the action-model learning process within TRAIL. The agent begins the learning process by recording experiences in its environment either by observing a trainer or by executing a plan. Second, the agent identi es instances of action success or failure during these experiences using a new analysis demonstrating nine possible causes of action failure. Finally, a variant of the Inductive Logic Programming algorithm DINUS is used to induce action models based on the action instances. As the action models are learned, they can be used for constructing plans whose execution contributes to additional learning experiences. Diminishing reliance on the teacher signals successful convergence of the learning process.