Genetic Learning of Digital Three-Layer Perceptrons for Implementation of Binary Cellular Automata.

This paper presents digital three-layer perceptrons (ab. DLPs ) having binary connection parameters and consider their application to analysis and synthesis of binary cellular automata (ab. BCAs ). Dynamics of BCS is governed by a rule table that is a Boolean function and can cause rich spatiotemporal patterns. Regarding a rule table of a BCA as a teacher signals we apply a GA-based learning algorithm to synthesize a DLP. Performing basic numerical experiments, we suggest the following: 1) The GA-based learning algorithm runs successfully and can realize rule table of BCA. 2) The number of hidden neurons may be an important factor to evaluate complexity of spatiotemporal pattern of BCA. 3) The DLP function can be preserved even if we reduce the teacher signals based on core teacher signals.