Learning action selection in autonomous agents

This paper focuses on learning in autonomous agents under dynamic environments. Autonomous agent control has been dominated by two major Artificial Intelligence (AI) approaches, Planning-based and Behavior-based. We concentrate on Behavior-based control. Multiple behavior modules from which a Behavior-based agent is constituted may have conflicting outputs (actions) for various input situations. It is becoming increasingly difficult to 'hard-wire' or 'pre-fix' the actions. "Learning" to select an appropriate action emerges as a strong alternative to hard-wired schemes. A Behavior-based agent is constructed for the research study. The application task chosen for the agent is to learn to navigate in a real indoor environment avoiding static and moving obstacles. The strategy is as follows: learning to avoid obstacles is to be achieved by learning to select an appropriate action in any input situation. The learning algorithm, based on reinforcement learning principles, chooses with high probability, an appropriate action based on the performance statistics (activations and reinforcements) of the conflicting behaviors. Learning experiments conducted to observe the performance of the agent were encouraging.