Tuesday Januari 7 2003 at 14h00 in Celestijnenlaan 200A, room 03.14 (Cafetaria)
Relational Reinforcement Learning: The early years.by Kurt Driessens (Dept. CW, KUL)
This talk will be about the idea of relational reinforcement learning. In reinforcement learning, an agent faces the problem of learning a beneficial behavior through his own interaction with his environment. The agent gets feedback from his actions through some kind of reward and must learn to optimize his behavior accordingly. In Q-learning, an existing reinforcement learning technique, this is done by building a Quality- or Q-function which expresses the benefit of executing a certain action in a certain world-state. The major problem that Q-learning encounters when dealing with large environments is that the number of different Q-values that have to be computed becomes too large. Relational reinforcement learning combines Q-learning with the representational power of relational learning techniques. It uses different relational regression techniques to represent a generalized Q-function. The first order regression engines that are being developed include an incremental first order regression tree algorithm, a relational instance based regression algorithm and neural logic programs, a first order extension of artificial neural networks. As a consequence of using a relational generalization of the Q-function, relational reinforcement learning is able to handle more difficult tasks, with larger environments and more possible actions, than standard reinforcement learning techniques, and reuse experience from simple learning tasks in more difficult but related tasks.