Overview

This project is an attempt to adapt Deep Q-Learning, as described in Playing Atari with Deep Reinforcement Learning by Mnih et al, for 3D Tictactoe

Installation Dependencies

• Python 2.7 or 3.5
• TensorFlow 0.10

TODOS:

• Enumerate best states for 2D tictactoe using minimax

• Implement q learning for 2D tictactoe

• Extend q learning for 3D tictactoe and see what breaks (couldn't finish enumerating states in state table - 80 million and counting)

• Implement deep q learning using a simple 2-layer neural net for 2D Tictactoe (then 3D Tictactoe)

• Implement policy gradient learning using a simple 2-layer neural net for 2D Tictactoe (then 3D Tictactoe)

• Establish reward rubrics and input format for tictactoe DQN pipeline

• Design model pipeline for DQN

• Design model pipeling for Policy Gradient Learning

• Experiment with model architecture to improve performance