On Adaptation of Loss Functions in Decentralized Adaptive Control

Decentralized adaptive agents/controller are able to operate in autonomous regimes, where they estimate model of their own neighborhood and design their control strategies with respect to their local aims. When each agent designs its strategy using only its model, the resulting control will be suboptimal since local models do not allow prediction of consequences of actions of the neighbors. Information about expected future trajectory as perceived by the neighbor can be obtained by means of communication. The task is to design a method of control strategy design that makes use of this information to improve the overall performance. In this paper, we propose to use the predictor to modify the loss function of the receiving agent. Informal justification of the approach is presented and it properties are illustrated in simulation experiments with decentralized LQG control.