Trust-Based Cooperative Games and Control Strategies for Autonomous Military Convoys.

摘要

The future presence of autonomous military robots in heterogeneous teams will introduce new trust-based vulnerabilities that previously did not exist in homogeneous human teams. Among these vulnerabilities is their exposure to cyber attacks, which can disrupt and/or take over these systems. Evidence exists that other nation-states are actively developing their cyber attack capabilities to break into U.S. military unmanned systems. Given this problem, our general research goal was to determine the feasibility of computational trust as a defensive capability against unacceptable behaviors in autonomous multi-agent systems. To meet this goal, we sought to develop new or improved computational trust models, algorithms, and frameworks for trust cultivation, aggregation, and propagation in distributed teams. Since autonomous convoy operations are expected to be one of the near-term, large-scale applications of autonomous military technologies, we chose it to be the application focus of our research. Our work produced two major results. The first was the cooperative trust game -- a new mathematical framework to predict coalition formation in response to trust-based interactions. Using this theory, we developed the convoy trust game and proved that the most optimal trust payoff in centralized convoys occurs when the lead vehicle acts as the trusted third- party for all follower vehicles. For decentralized convoys, we discovered that the trust payoff can be maximized if agents view immediate leaders and followers as surrogates for the whole system of agents in front and behind them, respectively. The second major result was the development of the RoboTrust model -- a new computational trust model that assigns a trust value equal to the smallest value in a set of maximum-likelihood estimates based on different historical observations. RoboTrust explicitly separates the context of an observation and the actual trust calculation.