Beating the Multiplicative Weights Update Algorithm

TL;DR

This paper analyzes the robustness of a specific multiplicative weights update algorithm against adversaries, showing it converges in expected linear time despite adversarial corruption, outperforming previous Byzantine consensus bounds.

Contribution

It introduces a detailed adversarial model for MWU algorithms, provides empirical analysis of convergence under attack, and compares its efficiency to existing Byzantine consensus algorithms.

Findings

MWU converges in expected O(n) rounds under adversarial attack

The algorithm outperforms the previous O(n^3) Byzantine consensus bounds

Empirical results validate the robustness of the proposed MWU approach

Abstract

Multiplicative weights update algorithms have been used extensively in designing iterative algorithms for many computational tasks. The core idea is to maintain a distribution over a set of experts and update this distribution in an online fashion based on the parameters of the underlying optimization problem. In this report, we study the behavior of a special MWU algorithm used for generating a global coin flip in the presence of an adversary that tampers the experts' advice. Specifically, we focus our attention on two adversarial strategies: (1) non-adaptive, in which the adversary chooses a fixed set of experts a priori and corrupts their advice in each round; and (2) adaptive, in which this set is chosen as the rounds of the algorithm progress. We formulate these adversarial strategies as being greedy in terms of trying to maximize the share of the corrupted experts in the final weighted advice the MWU computes and provide the underlying optimization problem that needs to be solved to achieve this goal. We provide empirical results to show that in the presence of either of the above adversaries, the MWU algorithm takes $\mathcal{O}(n)$ rounds in expectation to produce the desired output. This result compares well with the current state of the art of $\mathcal{O}(n^3)$ for the general Byzantine consensus problem. Finally, we briefly discuss the extension of these adversarial strategies for a general MWU algorithm and provide an outline for the framework in that setting.