R3TR0Quan/SimplifiedRimWorldRL

Very simple implementation of RimWorld's mechanics in a gym env to train a reinforcement learning algorithm

Simplified RimWorld Env for RL

Very simple implementation of RimWorld's mechanics in a gym env to train a reinforcement learning algorithm

• A single deep Q learning algo learns how to kill enemies
• The friends are allowed instant movement on a 10 by 10 grid
• New enemy wave is spawned each time the last one is fully killed
• At the start waves are 1 enemy, at the half mark it goes to 2
• At the start the AI controls 1 actor, at the half mark it goes to 2
• Shots are instants
• There is a single box that provides cover

Rewards:

• 1 per kill
• -1 per actor that is killed
• -0.0000002 for moving to an occupied spot
• -0.0000002 for shooting an empty spot
• -0.05 for not doing anything for 10 consecutive steps
• +0.5 for killing all enemies in a round

Graph

The half point decrease in scores is because the amount of initial enemies and actors are upped.

Thanks:

• seungeunrho/minimalRL
• Learning to be a Bot: Reinforcement Learning in Shooter Games (Michelle McPartland and Marcus Gallagher, 2008)

---