Reproduce “Learning Symbolic Task Decompositions for Multi-Agent Teams” 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

Ameesh Shah, Niklas Lauffer, Thomas Chen, Nikhil Pitta, Sanjit A. Seshia

The LOTaD framework enables multi-agent teams to automatically discover optimal ways to divide complex tasks using reward machines and multi-armed bandit selection, leading to improved sample efficiency and coordination compared to monolithic and heuristic-based approaches. It consistently achieved higher rewards and learned faster than baselines across various multi-agent environments.

Key insights

  • Learning optimal task decompositions simultaneously with agent policies addresses the multi-agent credit assignment problem and improves learning efficiency without requiring manual design.
  • Conditioning individual agent policies on the overall task state, in addition to their subtasks, enables effective coordination and problem-solving in environments with interdependent agent dynamics.
  • Using reward machines provides a formal and structured way to represent complex tasks, facilitating their decomposition into manageable subtasks for multi-agent teams.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.