Investigation of Elementary Cellular Automata for Reservoir Computing

Abstract

Reservoir computing is an approach to machine learning. Typical reservoir computing approaches use large, untrained artificial neural networks to transform an input signal. To produce the desired output, a readout layer is trained using linear regression on the neural network.

Recently, several attempts have been made using other kinds of dynamic systems instead of artificial neural networks. Cellular automata are an example of a dynamic sys- tem that has been proposed as a replacement.-

This thesis attempts to discover whether cellular automata are a viable candidate for use in reservoir computing. Four different tasks solved by other reservoir computing sys- tems are attempted with elementary cellular automata, a limited subset of all possible cellular automata. The effect of changing different properties of the cellular automata are investigated, and the results are compared with the results when performing the same experiments with typical reservoir computing systems.

Reservoir computing seems like a potentially very interesting utilization of cellular automata. However, it is evident that more research into this field is necessary to reach performance comparable to existing reservoir computing systems.