LESSON wm.1.2 · Theory · 150 min

Dreamer: train an agent inside latent imagination

Dreamer: train an agent inside latent imagination. Learn RSSM, reconstruction, reward prediction, imagined trajectories, and actor-critic and complete: Compare learning from real t

Learning objectives

  1. Explain what problem “Dreamer: train an agent inside latent imagination” solves without hiding behind terminology.
  2. Trace the variables and causal links across RSSM, reconstruction, reward prediction.
  3. Complete “Compare learning from real trajectories with imagined rollouts” and judge the result with evidence rather than intuition.

Core concepts

RSSM

RSSM connects Dreamer’s deterministic memory and stochastic latent state. Representation learning incorporates observations, dynamics unfolds without new frames, and reward-value heads train the actor-critic; compare imagined and real returns as prediction horizon grows.

reconstruction

reconstruction connects Dreamer’s deterministic memory and stochastic latent state. Representation learning incorporates observations, dynamics unfolds without new frames, and reward-value heads train the actor-critic; compare imagined and real returns as prediction horizon grows.

reward prediction

reward prediction connects Dreamer’s deterministic memory and stochastic latent state. Representation learning incorporates observations, dynamics unfolds without new frames, and reward-value heads train the actor-critic; compare imagined and real returns as prediction horizon grows.

imagined trajectories

imagined trajectories connects Dreamer’s deterministic memory and stochastic latent state. Representation learning incorporates observations, dynamics unfolds without new frames, and reward-value heads train the actor-critic; compare imagined and real returns as prediction horizon grows.

and actor-critic

and actor-critic connects Dreamer’s deterministic memory and stochastic latent state. Representation learning incorporates observations, dynamics unfolds without new frames, and reward-value heads train the actor-critic; compare imagined and real returns as prediction horizon grows.

Build and verify

Compare learning from real trajectories with imagined rollouts

  • Predict: write the expected output, trend, or failure before running code.
  • Build: implement only the minimum components needed to answer the question.
  • Verify: compare with a baseline or trusted implementation; save seeds, parameters, and raw outputs.
  • Transfer: change one shape, dataset, scale, or workload condition and explain whether the conclusion still holds.

Open the complete interactive lesson