Abstract: | This dissertation describes a behaviour-based mobile robot, Asterix, which uses reinforcement learning to learn a box-pushing task. Two reinforcement learning algorithms, Z-learning and Q(lambda)-learning were used to investigate the feasibility of using reinforcement learning to automatically program a robot. Asterix was built and a software interface with a workstation written. The primary sensors, the infra-red detectors, were examined to determine their properties. The robot successfully learned a slightly sub-optimal behaviour, revealing some of the complexities involved when using reinforcement learning to program a behaviour-based robot. The behaviour of the robot highlighted the need for both internal and external metrics of success in a task in order to evaluate the robot's controller.
|