Reproduce “Policy Search, Retrieval, and Composition via Task Similarity in Collaborative Agentic Systems” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Saptarshi Nath, Christos Peridis, Eseoghene Benjamin, Xinran Liu, Soheil Kolouri, Peter Kinnell, Zexin Li, Cong Liu and 2 more
Researchers developed MOSAIC, an algorithm enabling decentralized agents to autonomously search, retrieve, and compose policies from peers based on task similarity and performance. This framework achieved up to 2.7x higher maximum total return and faster learning on various sparse-reward reinforcement learning benchmarks compared to isolated agents and outperformed several multi-task learning baselines.
Key insights
- Selective, similarity-driven policy transfer, guided by individual agent optimization, is more effective than isolated learning or global parameter sharing, enhancing sample efficiency and mitigating task interference.
- The integration of modular neural network masks, Wasserstein task embeddings for task similarity, and reward-guided policy composition facilitates decentralized and efficient knowledge sharing in agentic systems.
- Effective knowledge sharing can lead to the emergence of implicit curriculum learning, where agents solving simpler tasks inadvertently accelerate the learning of more complex tasks by others.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.