Improve the performance of neural network training with accurate information: take Connect6 for example

Abstract

DeepMind introduced a general reinforcement learning algorithm called AlphaZero to learn through self-play without any human knowledge. It got a superhuman success not only in Go but also in Chess and Shogi. Nevertheless, AlphaZero needs huge computational resources to train a high quality neural network. Most institutions have no such huge computational resources or cannot invest an enormous number of resources to support a research project. Therefore, this paper proposes to embed accurate information into the training phase for improving the performance of the neural network under limited resources. In competition with Zeta-180, the win rate of FD-60 far surpasses all other modifications. The results of experiments indicate that embedding accurate information into the training-phase can effectively improve the performance of the neural network under limited resources.