Strategies for autonomous adaptation and learning in dynamical networks

Abstract

The complexity of dynamical networks', which compute with diverse attractors, renders them inaccessible, at present anyway, to entirely analytical treatment. Therefore, exploration and development of a learning algorithm for such nets, would need to rely mostly on numerical simulations. Here, we discuss strategies for the development of autonomous adaptation and learning algorithms for dynamical networks that are driven by entropy related information theoretic measures. A net of parametrically coupled logistic processing elements, an instance of a dynamical network, is used to illustrate the rationale, detail, and features of the strategies developed.