Learning Quantitative Knowledge for Multiagent Coordination

Abstract

A central challenge of multiagent coordination is reasoning about how the actions of one agent affect the actions of another. Knowledge of these interrelationships can help coordinate agents --- preventing conflicts and exploiting beneficial relationships among actions. We explore three interlocking methods that learn quantitative knowledge of such non-local effects in TAEMS, a well-developed framework for multiagent coordination. The surprising simplicity and effectiveness of these methods demonstrates how agents can learn domain-specific knowledge quickly, extending the utility of coordination frameworks that explicitly represent coordination knowledge. Content Areas: Intelligent Agents: Tasks or Problems: Multi-agent communication, coordination, or collaboration; Intelligent Agents: Tasks or Problems: Learning and adaptation; Machine Learning and Discovery: Techniques or Algorithms: Specialized learning algorithms Introduction A major challenge of designing effective multiagent s...