Reproduce “Hierarchical Imitation Learning of Team Behavior from Heterogeneous Demonstrations” 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
Sangwon Seo, Vaibhav Unhelkar
Deep Team Imitation Learner (DTIL) is a hierarchical multi-agent imitation learning algorithm designed to learn multimodal team behaviors from heterogeneous demonstrations in partially observable environments. The method iteratively infers latent subtasks and updates policies via distribution matching, achieving expert-level performance in complex multi-agent tasks and accurately inferring underlying expert strategies.
Key insights
- A one-to-one correspondence between an agent's hierarchical policy and its observation-action-subtask-previous subtask occupancy measure is established for partially observable multi-agent settings.
- The overall distribution matching problem can be factored into two marginal measures (observation-action-subtask and observation-subtask-previous subtask) for simplified optimization.
- Leveraging online samples and deep function approximators enables DTIL to scale to complex, high-dimensional tasks, overcoming limitations of prior hierarchical multi-agent imitation learning methods.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.