Reinforcement Learning with TensorFlow

Architecture and properties of AlphaGo 

All methods have relied on a tree search combined with some domain knowledge and human intervention. AlphaGo uses both a tree search and two kinds of CNNs (policy and value networks) to guide the tree search. These CNNs are kind of similar to the evaluation function used in DeepBlue, with one difference, that is, CNNs learn the evaluation function while the evaluation function used in DeepBlue was hand-crafted. 

The tree search used earlier was a brute force approach, while CNN is a learning-based approach, which provides a sort of intuition-based game playing. Thus, the first task would be reducing the search space (which is in the order of 10¹⁷⁰ for Go). This can be done by two approaches:

• Reducing   ...