Solve game of the amazons with neural networks and distributed computing
Yuanxing Chang, Hongkun Qiu · 2018
A training model was constructed with convolution neural network. Its policy network was trained by supervised learning. Then the parameters of the policy network were constantly adjusted by self-play and the game strength was enhanced. So it made it surpassed the existing game programs that simply used - pruning algorithm or Monte Carlo algorithm. By using distributed computing to update the data asynchronously, multiple game groups were allocated to different machines at the same time. In this way, a number of common configuration computers could be made full use of.