In this chapter, we studied reinforcement learning algorithms for one of the most complex and difficult games in the world, Go. In particular, we explored Monte Carlo tree search, a popular algorithm that learns the best moves over time. In AlphaGo, we observed how MCTS can be combined with deep neural networks to make learning more efficient and powerful. Then we investigated how AlphaGo Zero revolutionized Go agents by learning solely and entirely from self-play experience while outperforming all existing Go software and players. We then implemented this algorithm from scratch.
We also implemented AlphaGo Zero, which is the lighter version of AlphaGo since it does not depend on human game data. However, as noted, AlphaGo Zero requires enormous amounts of computational resources. Moreover, as you may have noticed, AlphaGo Zero depends on a myriad of hyperparameters, all of which require fine-tuning. In short, training AlphaGo Zero fully is a prohibitive task. We don't expect the...
