Stochastic Differential Games and Intricacy of Information Structures

This paper discusses, in both continuous time and discrete time, the issue of certainty equivalence in two-player zero-sum stochastic differential/dynamic games when the players have access to state information through a common noisy measurement channel. For the discrete-time case, the channel is also allowed to fail sporadically according to an independent Bernoulli process, leading to intermittent loss of measurements, where the players are allowed to observe past realizations of this process. A complete analysis of a parametrized two-stage stochastic dynamic game is conducted in terms of existence, uniqueness and characterization of saddlepoint equilibria (SPE), which is shown to admit SPE of both certainty-equivalent (CE) and non-CE types, in different regions of the parameter space; for the latter, the SPE involves mixed strategies by the maximizer. The insight provided by the analysis of this game is used to obtain through an indirect approach SPE for three classes of differential/dynamic games: (i) linear-quadratic-Gaussian (LQG) zerosum differential games with common noisy measurements, (ii) discrete-time LQG zero-sum dynamic games with common noisy measurements, and (iii) discrete-time LQG zero-sum dynamic games with intermittently missing perfect state measurements. In all cases CE is a generalized notion, requiring two separate filters for the players, even though they have a common communication channel. Discussions on extensions to other classes of stochastic games, including nonzero-sum stochastic games, and on the challenges that lie ahead conclude the paper.