New research shows high entropy leads to symmetry equivariant policies in Dec-POMDPs

A new paper explores how high entropy regularization can lead to symmetry-equivariant policies in Decentralized Partially Observable Markov Decision Processes (Dec-POMDPs). The research demonstrates that sufficiently high entropy ensures policy gradient flow converges to a compatible joint policy across different initializations. Empirical tests on environments like Hanabi and Overcooked show that increasing the entropy coefficient significantly impacts cross-play returns, with potential for improvement by greedifying policies post-training. AI

IMPACT Suggests higher entropy coefficients for Dec-POMDP hyperparameter tuning, potentially improving multi-agent policy compatibility.

RANK_REASON This is a research paper published on arXiv detailing theoretical and empirical findings. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• arXiv
• Dec-POMDP
• Hanabi
• Overcooked
• Yokai

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →