RL Training with PPO on NoWings using MLAgents

Required setup to view/run trained models:

1. Download Unity 2019.4.25f1
2. python 3.8.x (3.10.x does not work with pytorch!)
3. pip install torch==1.7.0 -f https://download.pytorch.org/whl/torch_stable.html
4. pip install mlagents

See agent playing NoWings with trained model using PPO:

1. add cloned folder as a project in Unityhub, then go to assets/scenes/noWings
2. go to Hierarchy>jank, make sure there is an .onnx file attached to jankgent in the inference model category
3. hit play

Play the game in Unity:

1. If havn't, add cloned folder as a project in Unityhub, then go to assets/scenes/noWings
2. go to Hierarchy>gameobjects, check that the jankgent component is on Heuristic Only
3. hit play
4. dive with the down arrow key

To train new models on the game: (example with testing.yaml in the ppo config folder) mlagents-learn config/ppo/testing.yaml

See mlagents-learn --help for other possible flags.

We ran a total of three experiments with varied reward systems and yielded the following results:

More about the experiments are detailed in their respective folders.