Exploration by Optimisation in Partial Monitoring

TL;DR

This paper introduces a simple, efficient algorithm for adversarial partial monitoring games that achieves near-optimal regret bounds, matching the best known theoretical limits and extending to various game settings.

Contribution

The paper presents a new algorithm that attains optimal regret bounds in adversarial partial monitoring and performs well across multiple game types.

Findings

Achieves regret bound of 6(d+1)k^{3/2}√(n log k) in partial monitoring.

Matches the best known information-theoretic upper bounds.

Performs near-optimally in full information, bandit, and globally observable games.

Abstract

We provide a simple and efficient algorithm for adversarial $k$-action $d$-outcome non-degenerate locally observable partial monitoring game for which the $n$-round minimax regret is bounded by $6(d+1) k^{3/2} \sqrt{n \log(k)}$, matching the best known information-theoretic upper bound. The same algorithm also achieves near-optimal regret for full information, bandit and globally observable games.