Jian Huang, Adrian R. Pearce
In distributed systems, learning does not necessarily involve the participation of agents directly in the inductive process itself. Instead, many systems frequently employ multiple instances of induction separately. In this paper, we develop and evaluate a new approach for learning in distributed systems that tightly integrates processes of induction between agents, based on inductive logic programming techniques. The paper's main contribution is the integration of an epistemic approach to reasoning about knowledge with inverse entailment during induction. The new approach facilitates a systematic approach to the sharing of knowledge and invention of predicates only when required. We illustrate the approach using the well-known path planning problem and compare results empirically to (multiple instances of) single agent-based induction over varying distributions of data. Given a chosen path planning algorithm, our algorithm enables agents to combine their local knowledge in an effective way to avoid central control while significantly reducing communication costs.
Subjects: 7.1 Multi-Agent Systems; 12. Machine Learning and Discovery
Submitted: Oct 15, 2006