Improving Curiosity in Reinforcement Learning
Research at UNSW on a simpler curiosity model for game-playing agents.
My thesis research at UNSW, from February to July 2020.
What it does
- Builds on the Intrinsic Curiosity Module, which rewards an agent for reaching states it can’t predict yet.
- Makes the curiosity model simpler while keeping its performance.
- The idea came from the plum pudding model of the atom.
- Compares it with the original on Ms. Pac-Man, Breakout and Space Invaders.
Built with
- Python
- PyTorch
- OpenAI Gym
- Rainbow DQN, PPO and ICM