Evaluation of reduced neural network models for predicting go game moves

Abstract

With increasing processing power and the introduction of GPUs, convolutional neural networks are getting more and more complex. While these networks are able to solve more complex tasks, they are less suited for use on a mobile platform where there are stricter memory and power constraints. We will look at neural network reduction methods, which aim to reduce the memory and power requirements of convolutional neural networks, whilst maintaining their quality. These methods are applied and evaluated with a Go move predicting network. Additionally an Android App is developed that is able to recognize a Go board and stone positions in order to use the reduced network to predict the next best moves.