A multi-agent framework for simulating proactive teamwork

Teamwork becomes very important in many dynamic, multi-agent environments. Psychological studies about teamwork have shown that an effective team often can anticipate information needs of teammates based on a shared mental model. Unfortunately, existing multi-agent systems for teamwork are limited in supporting proactive information exchange. To address this issue, we have developed and implemented a multi-agent architecture called CAST that simulates and supports proactive information exchange in a dynamic environment. In this dissertation, we first introduce CAST, its architecture, and its kernel. Then we present a knowledge representation language called MALLET, which is used to represent the knowledge regarding the structure and process of a team. We present our approach toward proactive information exchange, its formal model that serves as the theoretical foundation for CAST, its algorithm and the communication support in CAST for proactive information exchange. Team process knowledge represented in MALLET is compiled into a Petri-Net based model and the states of the process are kept track of through tokens in the Petri-Net. Based on this model, an algorithm called DIARG has been developed to detect possible information flows among CAST agents and provide information proactively to those who may need the information. Agent teams can be dynamically formed through a dynamic role selection algorithm based on certain preference constraints of agent roles, which can help make teamwork more efficient and balance the workload among agents. Two sets of experiments have been designed and run on a Multi-Agent Wumpus World testbed application. The experiment results showed that with the support of proactive information exchange, CAST enhanced the effectiveness of teamwork among agents without sacrificing a high cost of their communication. The results also showed that with the support of dynamic role selection algorithm, CAST improved the efficiency of teamwork especially in a complex problem domain.