Can an AI earn a fire cape? I built a simulator of RuneScape’s Fight Caves and trained an agent to survive all 63 waves, including the final boss, Jad. It uses Proximal Policy Optimization (PPO), a deep reinforcement learning method, to learn through trial and error. Jad is not known for constructive feedback.
How it went
- At its best evaluation, the agent defeated Jad in 94.9% of attempts in the simulator, with food and potions available.
- The agent learned from scratch, without demonstrations or a scripted walkthrough.
- A playable 3D viewer lets me watch the agent or take the controls myself.
Curious how it learned? Read the story behind the project →