USC CS561 project. Implement a GO agent for 5x5 GO, no machine learning libraries allowed.
-
Feature extraction: Extract 1x1, 2x2 and 3x3 windows of location dependent and independent features on board.
-
Value estimation: Calculated by extracted features and their weights.
-
Learning algorithm: TD(0).
-
Search method: Monte-Carlo Tree Search.
Incorporate weight sharing among features to accelarate learning. Running efficiency improved by C++ in feature extraction and tree search.
Winrate against class championship player in previous years: 0.6