Online distributed algorithms for mixed equilibrium problems in dynamic environments

TL;DR

This paper introduces online distributed algorithms for solving mixed equilibrium problems with time-varying data in dynamic multi-agent environments, ensuring sublinear growth of regret and constraint violations.

Contribution

It proposes novel online distributed algorithms using mirror descent and primal-dual strategies for dynamic environments with time-varying data and constraints.

Findings

Algorithms achieve sublinear regret growth.

Performance holds under noisy gradient estimates.

Simulation results validate theoretical guarantees.

Abstract

In this paper, the mixed equilibrium problem with coupled inequality constraints in dynamic environments is solved by employing a multi-agent system, where each agent only has access to its own bifunction, its own constraint function, and can only communicate with its immediate neighbors via a time-varying digraph. At each time, the goal of agents is to cooperatively find a point in the constraint set such that the sum of local bifunctions with a free variable is non-negative. Different from existing works, here the bifunctions and the constraint functions are time-varying and only available to agents after decisions are made. To tackle this problem, first, an online distributed algorithm involving accurate gradient information is proposed based on mirror descent algorithms and primal-dual strategies. Of particular interest is that dynamic regrets, whose offline benchmarks are to find the solution at each time, are employed to measure the performance of the algorithm. Under mild assumptions on the graph and the bifunctions, we prove that if the deviation in the solution sequence grows within a certain rate, then both the dynamic regret and the violation of coupled inequality constraints increase sublinearly. Second, considering the case where each agent only has access to a noisy estimate on the accurate gradient, we propose an online distributed algorithm involving the stochastic gradients. The result shows that under the same conditions as in the first case, if the noise distribution satisfies the sub-Gaussian condition, then dynamic regrets, as well as constraint violations, increase sublinearly with high probability. Finally, several simulation examples are presented to corroborate the validity of our results.