Hybrid Reinforcement Learning Agent
In Progress Reinforcement LearningHybrid MethodsDecision Making
Developing a hybrid RL architecture that combines model-free and model-based approaches to improve sample efficiency and generalisation. The agent maintains a learned world model for planning while retaining a policy network for reactive behaviour in high-variance environments.
- Combines Dreamer-style latent planning with SAC policy learning
- Tested on continuous control benchmarks (MuJoCo, DMControl)
- Reward shaping via learned auxiliary objectives