Stochastic processes with optimization: A model of learning in higher systems (Q799308)

A special class of discrete-parameter stochastic processes is considered, where at each step of time an optimization procedure is performed, which leads the process to its next state. The long-run behaviour of this process is investigated. This stepwise evolution of the system in dependence of an optimization procedure may be viewed as a learning process. It is shown that under certain conditions a repeated presentation of a stimulus pattern (serving as basis for the optimization procedure) leads to a convergent process in the sense that the best action is chosen. The author names processes of this kind ''stochastic processes with optimization and relaxation''. Concerning their convergence properties two classes of them are considered: monoergodic and biergodic processes. Sufficient conditions for these properties are given.