robianmcd/chrome-dino-ai

AI bot that learns to play the offline dinosaur game in Chrome

Chrome Dino AI

AI bot that learns to play the offline dinosaur game in Chrome.

Install

pip install -r requirements.txt

On windows you can optionally install win32gui which will automatically focus the chrome window. Download the .whl file for your python version and architecture from https://www.lfd.uci.edu/~gohlke/pythonlibs/#pywin32 and install it with pip install pywin32‑224‑cp36‑cp36m‑win_amd64.whl

Run

• open chrome dev tools and "undock in separate window"
• Select the offline checkbox in the dev tools
• try to go to a website and put chrome in full screen
  • If you installed win32gui then go to https://github.com. It will look for this in the window title if there are multiple chrome windows open.
• run python dino_bot.py
• click on the chrome window so the bot can send it keyboard input
• wait a couple days
• profit

Note: if you have multiple monitors, uncomment the "show image" code in src/dino_img_util.py to see which screen the bot is looking at. You can control which monitor it uses by setting monitor_id in dino_bot.py

References

• Beat Atari with Deep Reinforcement Learning!
• Keras plays catch

TODOs / Experiments

• Try DQN improvements outlined here https://medium.freecodecamp.org/improvements-in-deep-q-learning-dueling-double-dqn-prioritized-experience-replay-and-fixed-58b130cc5682
• use an RNN with stateful=True
  • https://machinelearningmastery.com/cnn-long-short-term-memory-networks/
  • https://stackoverflow.com/questions/43882796/when-does-keras-reset-an-lstm-state
• Save/Load random_action_chance so you can kill the process and pickup where you left off
• Add ability to train initial model on everything in the memory
• Compress pixels to binary in the memory (need to use full size image instead of scaled down one so there is no anti-aliasing)
• Profile game loop. See how high the fps can go
• See how many duplicate images are in the memory
• Selectively prefer memories that are surprising to the model
• Unit test utility functions
• Move frame rate and timing into GameEnvironment