Learning Zero-Sum Linear Quadratic Games with Improved Sample Complexity and Last-Iterate Convergence

TL;DR

This paper introduces a simplified nested zeroth-order algorithm for zero-sum linear quadratic games, achieving improved sample complexity and guaranteed last-iterate convergence in both deterministic and model-free settings.

Contribution

It presents the first global last-iterate linear convergence result for zero-sum LQ games and enhances sample efficiency with a novel nested ZO algorithm.

Findings

Achieves (1/)^2 sample complexity in the model-free setting.

Establishes the first last-iterate linear convergence for zero-sum LQ games.

Improves sample complexity by several orders of magnitude over previous methods.

Abstract

Zero-sum Linear Quadratic (LQ) games are fundamental in optimal control and can be used (i)~as a dynamic game formulation for risk-sensitive or robust control and (ii)~as a benchmark setting for multi-agent reinforcement learning with two competing agents in continuous state-control spaces. In contrast to the well-studied single-agent linear quadratic regulator problem, zero-sum LQ games entail solving a challenging nonconvex-nonconcave min-max problem with an objective function that lacks coercivity. Recently, Zhang et al. showed that an~$\epsilon$-Nash equilibrium (NE) of finite horizon zero-sum LQ games can be learned via nested model-free Natural Policy Gradient (NPG) algorithms with poly$(1/\epsilon)$ sample complexity. In this work, we propose a simpler nested Zeroth-Order (ZO) algorithm improving sample complexity by several orders of magnitude and guaranteeing convergence of the last iterate. Our main results are two-fold: (i) in the deterministic setting, we establish the first global last-iterate linear convergence result for the nested algorithm that seeks NE of zero-sum LQ games; (ii) in the model-free setting, we establish a~$\widetilde{\mathcal{O}}(\epsilon^{-2})$ sample complexity using a single-point ZO estimator. For our last-iterate convergence results, our analysis leverages the Implicit Regularization (IR) property and a new gradient domination condition for the primal function. Our key improvements in the sample complexity rely on a more sample-efficient nested algorithm design and a finer control of the ZO natural gradient estimation error utilizing the structure endowed by the finite-horizon setting.