Mean Field (NCE) stochastic control: Populations of major and egoist-altruist agents

For noncooperative games the Nash Certainty Equivalence (NCE), or Mean Field (MF) methodology [1], [2] provides decentralized strategies which asymptotically yield Nash equilibria. An extension of this theory to populations of altruistic agents (defined with so-called social cost functions) and to mixed populations was carried out in [3], and a theory treating populations of egoistic agents and one or more so-called major agents was developed in [4]. In this paper we study the equilibria and the overall stability of dynamic LQG games, where (i) there is a single major agent and a large population of mixed minor agents, and (ii) the cost for each minor agent is a convex combination of its own cost and the social cost of the minor agents. We analyse the resulting equilibria, provide experimental results, and present a mean field stochastic control algorithm, which when applied by all agents in the system, gives rise to system behaviour where (i) all agents systems are L² stable, (ii) the set of controls yields an ε-Nash equilibrium for all ε, and (iii) if each minor agent in the system only considers the social cost, then the difference between (i) the cost observed by each minor agent and (ii) the social cost that would be observed if a centralized controller minimizes the social cost tends to zero as the population size grows to infinity.