3D Space Cadet Pinball - Deep Reinforcement Learning is an open-source example of create an AI in .Net using (Unity ML-Agents) and a game ('3D Space Cadet Pinball'), to serve as a learning environment for training intelligent agents to play pinball without supervision.

The Pinball agent is rewarded by the score going up, and all actions and frames that lead to a higher score are remembered. Initially the actions will be random (using a curiosity model) but over time the agent will choose actions that lead to the highest reward (score).

• A PinballAgent that can learn from an external game (3D Pinball) by watching the screen as a visual observation.
• (TODO:) Use behaviour cloning to watch a recording of the highest score for 3d Space Cadet Pinball
• (TODO:) Fix scaling limitation.

• Download and install ('3D Space Cadet Pinball')
• Download, install and configure (Unity)
• Setup a (Python 3.6 Virtual Environment)

• Setup a (training session)

• The brain (Pinball Agent) implemented in .Net
• The eyes (External Window Manager) used to screen capture external window, and send keys to it. You could say one eye watches the window frames and the other looks at the score.
• The AI config (trainer_config.yaml) the unity ML agents training config.

• The 3D Space Cadet Window must be in focus during training, with no windows overlapping it.
• Currently the PinballAgent can only read the score when Windows is set to 100% scale when training. This is due to the method chosen for detecting the score by comparing pixels.

To change a display scaling size using the recommended settings, use these steps:

1. Open Settings.
2. Click on System.
3. Click on Display.
4. Under the "Scale and layout" section, use the drop-down menu and select the 100% scale setting.

• Unity ML-Agents

Initial concept and code developed by Elliot Wood

This is an open-source project and we encourage and welcome contributions. Please add detail when submitting new feature branches.

If you run into any problems using this project, submit an issue and make sure to include as much detail as possible.

Apache License 2.0