NonZero: Interaction-Guided Exploration for Multi-Agent Monte Carlo Tree Search

Sizhe Tang, Zuyuan Zhang, Mahdi Imani, Tian Lan · 2026

Read NonZero: Interaction-Guided Exploration for Multi-Agent Monte Carlo Tree Search by Sizhe Tang, Zuyuan Zhang, Mahdi Imani, Tian Lan (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-00751.txt.

Description

Sizhe Tang, Zuyuan Zhang, Mahdi Imani, and Tian Lan present NonZero, a method that keeps Monte Carlo tree search workable when many agents must act together. Joint actions grow exponentially, so normal search cannot explore enough options on a realistic budget. NonZero instead searches over a compact nonlinear representation guided by surrogate models and an interaction-based proposal rule.

Single-agent changes are ranked by predicted gain, while pairs are scored by a mixed-difference measure that can reveal teamwork benefits no single agent shows alone. The proposal step is framed as a bandit problem with a sublinear local-regret guarantee. On MatGame, SMAC, and SMACv2, NonZero improves sample efficiency and final performance against strong baselines under matched budgets.

Catalog record

CREATORSizhe Tang, Zuyuan Zhang, Mahdi Imani, Tian Lan
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE75 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.00751

Browse the full library · About the Tower of Babel