Tech Report CS-89-17
Implementing a Learning System for Subsumption Architectures
John Shewchuk and Paul Viola
chuk We have built a robot that wanders about the environment and learns from novel experiences. The robot learns to perform those actions that bring about a predetermined desirable state. We argue that some of the difficulties associated with learning are reduced by embedding those modules capable of learning within a subsumption architecture. The claim that subsumption architectures are well suited to the creation of real-time robust control systems is not new. However, until now little has been said regarding the incorporation of learning. We claim, and hope to demonstrate, that an augmented subsumption architecture is especially well suited to the task of learning behaviors. We have found that a simple learning algorithm guided by a utility function can enable a robot to discover and exhibit new behaviors. In addition, by calculating a certainty factor associated with every learned response, the robot can moderate the effects of learning on its behavior.
(complete text in pdf)