| name | reinformed-dreamer-asymmetric-world-model |
| description | Reinforced Dreamer methodology for asymmetric reinforcement learning using latent guidance to improve world model representations and behaviors in model-based RL. |
| paper_id | 2607.26040 |
Reinforced Dreamer: Asymmetric World Model Training with Latent Guidance
Overview
This methodology addresses limitations in privileged information representations learned by asymmetric model-based reinforcement learning algorithms like Informed Dreamer. It proposes a novel asymmetric representation learning objective using latent guidance, resulting in the Reinforced Dreamer algorithm.
Key Contributions
- Identifies limitations in privileged information representations in existing asymmetric model-based RL approaches
- Proposes a novel asymmetric representation learning objective using latent guidance
- Demonstrates consistent improvement over standard Dreamer across multiple benchmarks
- Effective under both partial observability (with additional state info) and full observability (with refined state info)
Implementation Guidelines
- World Model Architecture: Use a standard Dreamer-style world model with observation encoder, transition model, and reward predictor
- Asymmetric Learning Setup: Provide additional supervision beyond rewards during training:
- Under partial observability: use additional state information
- Under full observability: use more refined state information
- Latent Guidance Objective: Implement the proposed latent guidance objective to improve privileged information representations
- Training Protocol: Train the asymmetric world model using the latent guidance objective alongside standard RL objectives
Use Cases
- Model-based reinforcement learning with partial observability
- Scenarios where additional state information is available during training but not deployment
- Applications requiring improved representation learning in world models
- Benchmark environments where standard Dreamer shows inconsistent performance
Activation Keywords
reinforced dreamer, asymmetric reinforcement learning, latent guidance, world model, model-based RL, privileged information
References
- arXiv:2607.26040 [cs.LG]
- Original Dreamer paper
- Informed Dreamer methodology
