Learning for repeated constrained games in counter-coalition space

The paper deals with the design and analysis of a learning gradient-type strategy for N-person averaged constrained game with incomplete information. Each player is modelled by a stochastic variable-structure learning automation (a simplest single state Markov chain). Using the "joint payoff function", the considered game problem is formulated in terms of, so-called, counter-coalition variables. A special δ-regularization is introduced. Such approach does not require the "diagonal concavity" conditions to guarantee the uniqueness of the Nash equilibrium. The asymptotic convergence of the suggested learning procedure is analyzed.