World Model for RL

Key Insight

Once you have a world model, you can train an agent without touching the real environment: let the model imagine thousands of future rollouts and have a policy learn from those dreamed trajectories instead of slow, expensive real experience — the core idea of DreamerV3. This project builds a lightweight version: learn a world model of a simple environment, roll it forward under candidate actions, and train a policy purely inside that "dream." This guide owns only the generative side — predicting the next state given an action; the policy-update loop and the reward objective belong to RL Phase 6 (Model-Based RL). The payoff is sample efficiency: imagined experience is essentially free once the model is trained, so the agent can practice far more than the real world would ever allow.