Decentralized decision making with strategic users
Many real-world dynamic decision-making problems consist of multiple decision-makers with asymmetric information. Some examples include Markets, social learning, traffic management, autonomous vehicles, cyber-physical systems, internet of things and many more. In these systems, there are multiple decision-makers (DMs) who make some common and private observations of the ‘state’ of the systems with the goal to minimize their own cost (dynamic games) or total cost incurred by everybody (dynamic teams). In this talk, I will present a general sequential decomposition framework to study such problems. This framework extends currently known results in decentralized stochastic control for team problems. For strategic users, it presents a novel methodology to compute (Markovian) Perfect Bayesian equilibria (PBE), which was an open problem in the theory of dynamic games. I present a running public-goods example to study its PBE. In general, our results extend the ideas of dynamic programming to general multi-agent dynamic optimization problems to study ‘signaling’ behavior i.e. how players’ actions reveal their private information in the system which affects other users’ utilities.