Balancing Coordination and Synchronization Cost in Cooperative Situated Multi-Agent Systems with Imperfect Communication

We propose a new Markov team decision model to the decentralized control of cooperative multi-agent systems with imperfect communication. Informational classes capture system’s communication semantics and uncertainties about transmitted information and stochastic transmission models, including delayed and lost messages, summarize characteristics of communication devices and protocols. This model provides a quantitative solution to the problem of balancing coordination and synchronization cost in cooperative domains, but its exact solution is computationally infeasible. We propose a generic heuristic approach, based on a off-line centralized team plan. Decentralized decision-making relies on Bayesian dynamic system estimators and decision-theoretic policy generators. These generators use system estimators to express agent’s uncertainty about system state and also to quantify expected effects of communication on local and external knowledge. Probabilities of external team behavior, a byproduct of policy generators, are used into system estimators to infer state transition. Experimental results concerning two previously proposed multi-agent tasks are presented, including limited communication range and reliability.