DataTerminology.comNEW
Also known as: DreamerV2, DreamerV3
A model-based reinforcement learning algorithm that learns a latent world model to plan and act efficiently.
DESCRIPTION
Dreamer trains a latent dynamics model from high-dimensional observations, then optimizes an actor and critic using imagined rollouts. It achieves strong sample efficiency on continuous control tasks and has evolved into DreamerV2 and DreamerV3 with improved stability and performance. It illustrates the power of planning in learned latent spaces over raw pixel spaces.
ALTERNATIVE DEFINITIONS
arXiv
We present Dreamer, an algorithm that learns behaviors purely from images via training world models and optimizing actions by using imagination.
Emphasizes image-based learning and imagination with world models.
Danijar Hafner
Dreamer learns a world model from the agent’s experience and trains an actor and critic purely in imagination to achieve sample-efficient control.
Highlights actor-critic training in latent imagination.
arXiv
DreamerV2 scales model-based RL by learning value functions and policies from imagined trajectories, achieving competitive performance across tasks.
Addresses scalability and performance in Dreamer’s successor.
RELATED TERMS
Model-Based Reinforcement Learning World Model Actor-Critic Latent Dynamics Imagination Rollouts Representation Learning MuZero Planning Policy Optimization Sample Efficiency
COMPETING TERMS
Model-Free Reinforcement Learning Deep Q-Network
KNOWLEDGE AREAS
TARGET AUDIENCE
AI Engineer Data Scientist MLOps Engineer
EXTERNAL SOURCES
- Wikipedia - Reinforcement learning
- arXiv - Original Dreamer paper
- arXiv - DreamerV2 paper
- arXiv - DreamerV3 improvements
TOP QUESTIONS
What makes Dreamer different from model-free RL?
Dreamer learns a world model and plans in imagination, improving sample efficiency, while model-free methods rely on direct environment interactions.
Category: Alternatives
Where does Dreamer perform well?
Continuous control tasks and pixel-based environments where planning in latent space aids learning efficiency.
Category: Example
What are key components of Dreamer?
Latent dynamics model, actor-critic optimization, imagined rollouts, and representation learning.
Category: Requirements
What’s a hot topic around Dreamer?
Combining world models with large-scale representation learning and planning for generalist agents.
Category: Topic
How does Dreamer handle high-dimensional inputs?
It encodes observations into a latent space and models dynamics, avoiding expensive pixel-level prediction.
Category: Misunderstanding
Is Dreamer suitable for real-world robotics?
Promising due to sample efficiency, but requires careful modeling of noise, partial observability, and safety constraints.
Category: Opinion
What are common pitfalls implementing Dreamer?
Unstable training from poor model capacity, inadequate rollout length, and misaligned actor-critic learning rates.
Category: Mistakes
How is Dreamer evaluated?
Standard RL benchmarks (e.g., DM Control) and sample efficiency metrics compared to model-free baselines.
Category: Testing
What’s the relation of Dreamer to MuZero?
Both use learned models; MuZero plans via MCTS on latent dynamics, while Dreamer trains actor-critic via imagined trajectories.
Category: Example