Learning Collaborative Behavior by Observation

This paper presents a multi-agent framework capable of learning teamwork by observation. The system combines proven single entity learning by observation techniques with a multi-agent system shown to exhibit effective teamwork. An effective simulated production team is observed. An off-line training algorithm uses the observed data to develop behavior maps for a Collaborative Context-based Reasoning framework. The Collaborative Context-based Reasoning framework provides generic base classes capable of recreating the behavior of the original agents using the behavior maps developed by the training algorithm. The resulting prototype effectively replaces the original team of agents and is capable of reproducing its behavior and generalizing the behavior to encompass similar situations.