Keras Reinforcement Learning Projects

The AlphaGo algorithm

The algorithm behind AlphaGo consists of two parts:

• A tree-search procedure
• Convolutional networks that guide the tree-search procedure

In total, three convolutional networks of two different kinds are trained: two policy networks and one value network.

AlphaGo's algorithm uses a combination of machine learning and tree-search techniques, along with an extensive gaming and human learning phase. It uses the MCTS for the selection of moves, guided by two deep neural networks (value network and policy network). Before being sent to neural networks, the input is analyzed in a preprocessing phase to extract some features (for example, the adherence of the moves to a series of common patterns).

In the first phase of the ...