The Neural Networks Realization of Quasi-Gradient System to Search for Equilibrium Solutions in Game Theory and Their Dynamical Behavior

Abstract

Mutually coupled plural Neural Networks (N.N.) modules are proposed from the view point of noncooperative game theory. First, new dynamical models, which is called “Quasi-Gradient System”, to search the Nash Equilibrium (NE) points under [0, 1] -interval or nonnegative constraints are proposed. The stability of the proposed searching models is analyzed by the linearization approach. In addition, relations between the Lotka-Volterra's ecological model or the population genetics model and the proposed searching models are indicated. Second, new mutually coupled plural N.N. modules are introduced to realize the proposed searching model for problems with quadratic objective functions. the asymmetric Hopfield type N.N. can be regarded as a special class of the proposed N.N. modules. Last, by simulations for simple problems, the biffurcations in dynamical behavior such as converging to different NE points, cyclic state transition with no NE points and other exceptional cases are shown.