Characterizing Oblivious Common-Reward Multi-Player Bandits: Universal Stationarity, Coordination Traps, and Certified Global Rates

Key points

• First, we prove that the independent gradient-bandit updates applied by the players jointly form an unbiased stochastic gradient estimator of a hidden product-policy objective.1
• Third, we show that global optimality can be recovered under structural assumptions on the reward table.1
• To make global optimality achievable, we introduce finite-table certificates that separately characterize product-gradient dominance and self-bounding stochastic noise.1
• We then establish certificate-conditioned bounds on the product-policy gap and pseudo-regret.1

Sources (1)

• [1]Characterizing Oblivious Common-Reward Multi-Player Bandits: Universal Stationarity, Coordination Traps, and Certified Global Rates
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 21, 12:49 PM

  This paper presents a theoretical analysis of the oblivious common-reward multi-player multi-armed bandit (OC-MPMAB), a minimal multi-player bandit model in which multiple players select arms simultaneously and observe only a shared reward.

  First, we prove that the independent gradient-bandit updates applied by the players jointly form an unbiased stochastic gradient estimator of a hidden product-policy objective.

Extractive summary: sentences quoted from the sources.