A unified approach to mean-field team: homogeneity, heterogeneity and quasi-exchangeability

TL;DR

This paper introduces a unified method for large-scale stochastic team optimization that transforms heterogeneous agent systems into a more manageable homogeneous form, enabling analysis and establishing asymptotic optimality.

Contribution

It develops a novel unified approach converting continuum heterogeneity into homogeneity, addressing a complex forward-backward stochastic system with double projections and mean-field interactions.

Findings

Proposes a new unified approach for heterogeneous agent systems.

Establishes well-posedness of a novel forward-backward stochastic system.

Demonstrates asymptotic optimality of the proposed method.

Abstract

This paper aims to systematically solve stochastic team optimization of large-scale system, in a rather general framework. Concretely, the underlying large-scale system involves considerable weakly-coupled cooperative agents for which the individual admissible controls: (\textbf{i}) enter the diffusion terms, (\textbf{ii}) are constrained in some closed-convex subsets, and (\textbf{iii}) subject to a general \emph{partial decentralized information} structure. A more important but serious feature: (\textbf{iv}) all agents are heterogenous with \emph{continuum} instead \emph{finite} diversity. Combination of (\textbf{i})-(\textbf{iv}) yields a quite general modeling of stochastic team-optimization, but on the other hand, also fails current existing techniques of team analysis. In particular, classical team consistency with continuum heterogeneity collapses because of (\textbf{i}). As the resolution, a novel \emph{unified approach} is proposed under which the intractable \emph{continuum} \emph{heterogeneity} can be converted to a more tractable \emph{homogeneity}. As a trade-off, the underlying randomness is augmented, and all agents become (quasi) weakly-exchangeable. Such approach essentially involves a subtle balance between homogeneity v.s. heterogeneity, and left (prior-sampling)- v.s. right (posterior-sampling) information filtration. Subsequently, the consistency condition (CC) system takes a new type of forward-backward stochastic system with \emph{double-projections} (due to (\textbf{ii}), (\textbf{iii})), along with \emph{spatial mean} on continuum heterogenous index (due to (\textbf{iv})). Such system is new in team literature and its well-posedness is also challenging. We address this issue under mild conditions. Related asymptotic optimality is also established.