Title:Deep Teams with Risk-Sensitive Linear Quadratic Models: A Gauge Transformation

Abstract:Motivated from invariance principle and recent developments in artificial intelligence, we introduce a risk-sensitive linear quadratic control system whose solution resembles a deep neural network. In particular, we consider a population of decision makers that are partitioned into multiple sub-populations wherein the decision makers in each sub-population are coupled in both dynamics and cost function through a few linear regressions of the states and actions of the decision makers. Two non-classical information structures are considered: deep-state sharing and partial deep-state sharing, where deep state refers to the linear regression of the states of the decision makers in each sub-population. For deep-state sharing structure, a closed-form representation of the globally optimal strategy is obtained in terms of a deep Riccati equation, whose dimension is independent of the number of decision makers in each sub-population, hence scalable. In addition, two sub-optimal sequential strategies under partial deep-state sharing information structure are proposed by introducing two Kalman filters, one based on the finite-population model and the other one based on the infinite-population model. It is shown that the prices of information associated with the above sub-optimal solutions converge to zero as the number of decision makers goes to infinity.
Two numerical examples of a supply-chain management system are presented to demonstrate the efficacy of the obtained results.