On ergodic two-armed bandits (Q417067)

This paper focuses on the Narenda two-armed bandit algorithm under assumption that the payoff sequences are unknown and deterministic. The authors consider the Narenda algorithm under the conditions required by \textit{D. Lamberton, G. Pagès} and \textit{P. Tarrès} [Ann. Appl. Probab. 14, No. 3, 1424--1454 (2004; Zbl 1048.62079)] without monotonicity and under weak ergodic assumptions. The obtained results point out that, even with strongly dependent outcomes, the payoff probability accumulates statistical information on the ergodic behaviour of the two arms to induce a corresponding appropriate decision.