Imagined Training

Imagined training refers to training policies entirely within generated simulated worlds produced by a world model, rather than in the real environment or recorded demonstrations from it.

In imagined training, a world model serves as a generative simulator, producing plausible states and transitions that did not occur in any real trajectory. Policies trained in these imagined worlds may generalize to unseen environments and unseen interacting partners without requiring access to the original game.

This approach addresses the coverage problem in reinforcement learning: passively collected demonstrations cover only a small fraction of meaningful interactions. World models enable active generation of novel scenarios, including adversarial situations, rare events, and entirely new level configurations.

Imagined training in multi-agent settings enables competitive agents, cooperative agents, and mixed populations that learn entirely within generated environments.