patrickvonplaten / TRexGameRL

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

This repo contains a private project applying Reinforcement learning for the TRex game provided by chromium


  • chrome
  • chromedriver (sudo apt-get install chromium-chromedriver && sudo ln -s /usr/lib/chromium-browser/chromedriver /usr/bin/chromedriver)
  • Tkinter (apt-get install python-tk, for pyenv users check this)
  • pip install -r requirements.txt

The motivation is to master the TRex game purely based on visual input in form of screenshots of the game. This way, the model has access to the same and only the same information a human play would have for the Game.

The DQN model architecture is taken from Playing Atari with Deep Reinforcement Learning.

Iteratively, the model was improved by implementing the duel networks architecture, prioritized experience replay and double q-learning.

An average score of ~800 (see python/plots/score_plot_trial19_mem_50k_lr_0.0001_decay_5000.png) is achieved by using a large memory of 50.000 and Adam optimization.

Using a smaller memory of only 12.000 with Adam only a score of ~400 could be achieved (see python/plots/score_plot_trial3_adam_12k_2nd.png).

ezoic increase your site revenue



Language:JavaScript 51.2%Language:Python 28.3%Language:HTML 17.3%Language:Shell 1.7%Language:CSS 1.4%