In this talk, a class of stochastic decentralized control systems called Mean Field Teams (MFT) are introduced. Such systems consist of an arbitrary number of coupled subsystems; the dynamics of the subsystem are coupled through the mean-field (empirical distribution) of the states; the per-step cost is arbitrarily coupled. The controller of each subsystem observes the local state and the mean-field of the system. Such systems arise in a variety of applications including smart grids and communication networks.
By exploiting the exchangeability of subsystems, we identify an information state and obtain a dynamic programming decomposition that identifies team-optimal strategies. The solution complexity increases polynomially with the number of subsystems that enables us to solve problems with moderate (100s to 1000s) number of subsystems. Both Markov chain and LQG variations are presented.
Joint work with Aditya Mahajan.
Bienvenue à tous!