Search arXiv⌕ Search
Home/Archive/arXiv

arXiv · 2601.07142

Dynamics of Multi-Agent Actor-Critic Learning in Stochastic Games: from Multistability and Chaos to Stable Cooperation

Yuxin Geng·Wolfram Barfuss·Feng Fu·Xingru Chen
View original source ↗Read source document ↗Export metadata ↓

Abstract

Achieving robust coordination and cooperation is a central challenge in multi-agent reinforcement learning (MARL). Uncovering the mechanisms underlying such emergent behaviors calls for a dynamical understanding of learn processes. In this work, we investigate the dynamics of actor-critic agents in stochastic games, focusing on the impact of entropy regularization. By leveraging time-scale separation, we derive the system's evolution equations, which are then formally analyzed using dynamical systems theory. We find that in the constant-sum game of Matching Pennies, the system exhibits chaotic behavior. Entropy regularization mitigates this chaos and drives the dynamics toward convergence to fair cooperation. In contrast, in the general-sum game of the Prisoner's Dilemma, the system displays multistability. Interestingly, the three stable equilibria of the system correspond to the well-known ALLC (Always Cooperate), ALLD (Always Defect), and GRIM (Grim Trigger) strategies from evolutionary game theory (EGT). Entropy regularization strengthens system resilience by enlarging the basin of attraction of the cooperative equilibrium. Our findings reveal a close link between the mechanism of direct reciprocity in EGT and how cooperation emerges in MARL, offering insights for designing more robust and collaborative multi-agent systems.

Explore related subjects

physics.soc-ph

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Yuxin Geng, Wolfram Barfuss, Feng Fu, Xingru Chen. 2026-01-12. Dynamics of Multi-Agent Actor-Critic Learning in Stochastic Games: from Multistability and Chaos to Stable Cooperation. https://arxiv.org/abs/2601.07142

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

Self-Reference in Large Language Models: The Introspection Threshold for Recursive Self-Improvement

The pursuit of self-evolving AI raises a critical question: when is autonomous self-improvement sustainable rather than degenerative? Drawing an analogy to von Neumann's complexity threshold for self-reproducing automata, we argue that sustainable recursive self-improvement in Large Language Models (LLMs) requires a functional analogue: introspection -- the system's capacity to simulate its own operations and target modifications. Grounded in Kleene's Second Recursion Theorem, we demonstrate the theoretical existence of such introspective programs. However, an empirical review reveals that while current LLMs exhibit quasi-introspection (e.g., partial metacognition), they fall short of true introspection due to structural bottlenecks: a lack of complete self-access, the feedforward nature of the Transformer, and computational class constraints that prevent fixed-point iteration. We conclude by outlining architectural paths to cross this complexity threshold and discussing the associated safety implications.

physics.soc-ph↗
arXiv2026

Multilayer Analysis of the Global Trade Network

Global trade is more than a single network of aggregate flows. Beneath the observable exchange of products among economies lies a complex multilayer structure, formed by thousands of product-specific trade relationships that differ in their similarity, interdependence, and temporal evolution. Using the CEPII's BACI database, which records bilateral product-level trade flows between economies, we represent the global trade network from 1995 to 2024 as a temporal multilayer network, with economies as nodes and directed weighted trade flows as edges. To investigate product-level organisation and cross-layer similarity, temporal structural change, and the structural role of individual economies, we introduce a random-walk-based similarity measure that provides a unified framework for comparing weighted and directed trade layers. Our results show that the global trade network remains relatively stable over short periods but undergoes gradual structural change over longer timescales. We also find that similarity-based product communities only partially align with the official product taxonomy, indicating that products assigned to the same official category do not necessarily exhibit similar trade-network structures. Finally, we show that an economy's structural influence is not always determined by its trade volume. These results highlight the value of multilayer network analysis for revealing patterns in global trade that remain hidden at the aggregate level.

physics.soc-ph↗
arXiv2026

Detectability limits of scaling laws

Power law scaling relations between size and output are central to quantitative theories of cities, organisms, and other complex systems. Competing theories predict scaling exponents that differ by small fractions, but there is no existing theory for verifying whether a given dataset can even distinguish exponents at the required resolution to address such discrepancies. Here we derive a resolution limit for scaling exponents, giving the smallest exponent difference that any method of analysis can detect. We find that the Hurst exponents governing the evolution of systems' sizes and deviations from the scaling law determine how long a record of growing systems must be before it can separate competing scaling theories. Empirical results suggest that many available data panels are insufficient for reliable scaling model selection.

physics.soc-ph↗