Mastering the game of Go from scratch

In this report we pursue a transfer-learning inspired approach to learning to play the game of Go through pure self-play reinforcement learning. We train a policy network on a 5 ⇥ 5 Go board, and evaluate a mechanism for transferring this knowledge to a larger board size. Although our model did learn a few interesting strategies on the 5 ⇥ 5 board, it never achieved human level, and the transfer learning to a larger board size yielded neither faster convergence nor better play.