Prior structure for online learning

摘要

Online learning and adaptation capabilities are important for robot systems which operate in the real world and have to react to changes in the environment and the task requirements. Such learning schemes, however often suffer from problems of complexity, rendering them intractable for online learning in complex domains. To address these problems, the approach presented here introduces prior structure which drastically reduces the size of the state and action spaces. Deriving control from a set of predictable control laws, the system is modeled more abstractly as a discrete event dynamic system which represents all possible system behavior. Use of this framework also permits the automatic incorporation of knowledge in the form of constraints, providing a basis for the explicit representation of safety conditions and "maturational" effects which can be used to further enhance learning performance. To illustrate this structure and to show its applicability, the acquisition of a rotation gait on a four-legged walking platform is demonstrated.