APPLICATION OF CELLULAR AUTOMATA IN THE DEVELOPMENT OF NEURAL NETWORKS

Abstract

Evolutionary computing is a branch of artificial intelligence that uses methods that describe natural processes. Such methods include finding solutions to complex nonlinear problems such as pattern recognition, finding the best solution among many, modelling the mutation process, etc. Evolutionary algorithms are also used to automate the process of finding the best combination of parameters for a machine learning model. The use of cellular automata to model the evolutionary process using a neural network is possible only at the beginning of the process. In this case, machine learning models can be represented as cellular automata.  The state of each cell in the automaton represents a parameter or a part of the model architecture. Then the evolution of the automaton will be governed by a set of rules that determine the change of the automaton over time, depending on the current state of the automaton. 

This approach expands the class of models whose development process is described by traditional evolutionary methods. Processes based on cellular automata are more complex and flexible. The use of cellular automata can provide a more intuitive way of representing an evolutionary machine learning model. Since cellular automata have local interaction between elements, this allows improving the properties of a neural network and getting rid of such a drawback as overfitting. 

To evaluate the quality of the built models, we compared each type of network on the same data sample with the same parameters. The accuracy, speed, convergence, and efficiency of each approach were evaluated.  

Keywords. neural network, evolutionary algorithms, cellular automata, accuracy, modeling of dynamic systems.