Game-playing Reinforcement Learning Model Project: Python Code

Python notebook for training and evaluating models: Python Notebook

Notebook for viewing TensorBoard (training logs): Python Notebook

Custom Python libraries used:

• https://github.com/MatPoliquin/baselines-fix.git (a fork of OpenAI Baselines that fixes a bug which significantly improves training time)
• https://github.com/MatPoliquin/stable-retro.git (a fork of OpenAI Retro that has more game integrations)

Code library modifications (link to modified files)

• baselines/run.py:
  • line 57: added ‘MarioBros-Nes’ to _game_envs[retro] dict
• baselines/common/retro_wrappers.py:
  • line 239: added ‘MarioDiscretizer’
  • line 251: added call to MarioDiscretizer when building environment
  • line 253: modified WarpFrame parameters to specify width, height, and non-greyscale images
• baselines/common/cmd_util.py
  • line 80: modified initilization of the Gym Retro environment to NOT use the default action space
• retro/data/stable/MarioBros-Nes/data.json
  • Added “enemies” variable & memory location (for enemies remaining on level)
• retro/data/stable/MarioBros-Nes/updated_scenario_cropping.json
  • Added cropping of the game image
• retro/data/stable/MarioBros-Nes/updated_scenario_reward_func.json
  • Updated scenario for new reward function based on enemies remaining instead of game score
• retro/data/stable/MarioBros-Nes/rom.nes
  • Included ROM for game

---

Back