kevin-fang / 2048-dnn-solver

An 2048 solver using deep neural nets [unfinished]

2048-solver

An 2048 solver using TensorFlow and its abstraction library TFLearn. The 2048 game is implemented (and I believe it works fine), but the neural network is currently not working.

How to run

Note that TensorFlow, TFLearn, and NumPy are all required dependencies.

To generate training data, run python generate_training_data.py.

To train the neural net on the generated training data, run python train_nn.py

Once the neural net is trained, you can check the accuracy by running python test_nn.py.

If you want to see how well the neural net does compared to a random player, try python random_test.py.

With the neural net testing, for some reason it will hit a max accuracy of about .55 with a loss around 1. When testing is done, the score is about 160 over 1000 runs. For reference, a random player will get on average 180.

How it works/how it's supposed to work:

generate_training_data.py produces a numpy array called saved_training.npy containing random games, that had a threshold of at least score_requirement. Each entry in the numpy array looks like:

    [[0, 2, 4, 8, 2, 8, 4, 8, 16, 32, 2, 4, 8, 16, 32, 64], [0, 1, 0, 0]]

Where the first element is the flattened (raveled) array of the 2048 game, and the second element is a one-hotted array of the move made at that position, with the positions meaning [up, down, left, right].

train_nn.py opens saved_training.npy and trains the neural net using the following formula with the model provided in nn_model.py, and saves the model in models/trained_nn.model.

test_nn.py opens the model located at models/trained_nn.model and runs 1000 (or whatever num_tests is) games and calculates the average accuracy and the amount that each button was pressed