CoDreamer: Communication-Based Decentralised World Models
TLDR
CoDreamer extends Dreamer to multi-agent RL using GNN-based communication for decentralized world models, improving sample efficiency and cooperation.
Reasoning
The paper introduces a novel extension of Dreamer to multi-agent settings with a two-level communication system, addressing partial observability and cooperation. Strengths include clear methodology and demonstrated superiority over baselines, but weaknesses include lack of real-world validation and limited detail on environments.
Read-first score
Read-first score 60.8, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 58.
Field roles
Rank sensitivity
Stability: volatile; rank range: 569.
Keyword Scores
Deep Analysis
Innovations
- Extension of the Dreamer algorithm to multi-agent environments
- Two-level communication system using Graph Neural Networks for world models and policies
- Separate communication within learned world models and within learned policies to enhance modelling and task-solving
Methodology
CoDreamer extends the Dreamer algorithm to multi-agent settings by incorporating Graph Neural Networks for a two-level communication system. Communication is utilized both within the learned world models and within the policies of each agent to address partial observability and inter-agent cooperation. The approach is evaluated against baseline methods across various multi-agent environments.
Key Results
CoDreamer offers greater expressive power than a naive application of Dreamer and demonstrates superiority over baseline methods across various multi-agent environments.