/2048-DeepQLearning

Implementation of an agent able to play 2048. Report, evaluation and implementation details in the GitHub repository. During the project I gained experience with Reinforcement learning technologies such as the gym framework and the Keras-rl library.

Primary LanguagePython

2048 A Deep Q-Learning-Approach

Project for the Machine Learning course - La Sapienza Università di Roma

Implementing an agent for the 2048 game using a reinforcement learning approach.

For more details look at: ML_2048_report.pdf

Let's see in action: https://youtu.be/swc_otVJJrA

Requirements

  • tensorflow==1.14.0
  • keras-rl
  • keras==2.2.0
  • gym
  • matplotlib
  • pandas
  • numpy

Graphic Interface: visualize=True in test/fit

Evaluation:

Training results with 5 millions steps

alt text

Dataframe training results with 5.000 steps

alt text

Dataframe training results with 5.000.000 steps

alt text