Convolutional Neural Networks for Cellular Automata Classification

Wolfram famously developed a four-way classification of CA behaviour, with Class IV containing CAs that generate complex, localised structures. However, finding Class IV rules is far from straightforward, and can require extensive, time-consuming searches. This work presents a Convolutional Neural Network (CNN) that was trained on visual examples of CA behaviour, and learned to classify CA images with a high degree of accuracy. I propose that a refinement of this system could serve as a useful aid to CA research, automatically identifying possible candidates for Class IV behaviour and universality, and significantly reducing the time required to find interesting CA rules.