This code is based on a fork of https://github.com/HaniAlmousli/minesweeperRLDemo . The minesweeper code originates from https://github.com/cash/minesweeper .

The main changes from @HaniAlmousli are:

• In the original implementation, a regressor is trained per action to predict the Q value given a state, and at every update, a brand new set of regressors are trained.
  Instead, we propose to use a regressor with a partial_fit method: http://scikit-learn.org/stable/modules/scaling_strategies.html#incremental-learning

• AgentDQN.py has been added to experiment with the DQN approach using Keras.

CNN model:

The advantage of using a CNN is that the model can be trained on a specific game size but can then be applied to games of any size.

Random Agent:

DQN Agent:

Random Agent:

Ridge Regressor Agent:

A Linear Model that takes the whole Minesweeper grid as input works better than a dense CNN that only looks at the game with a sliding window (twice better in this case). But the CNN approach has the advantage of being independent of the size of the game (for instance it was trained on a 16x16 grid and tested on a 10x10 grid).