This implements Q-learning similarly to Mnih et al. 2013 "Playing Atari with Deep Reinforcement Learning" http://arxiv.org/pdf/1312.5602.pdf

• Python 2.7 with numpy/scipy
• Theano
• Lasagne

Implements a Connect-Four game board (6 rows, 7 columns by default)

Implements a base class and various policies for playing Connect-Four. A policy takes a game state (a Board object) and returns an optimal action (in some sense)

Attempts to train a Convolutional Neural Network (CNN) using Q-learning to predictQ(s,a) the value of taking each action a in state s The input is s and there is one output neuron for each a corresponding to Q(s,a).

Trains a CNN to predict the winner(s) of Connect-Four and other stuff in a supervised way: random boards (X) are generated and the answers (y) are computed. The CNN is then trained using gradient descent on examples (X,y).