We consider a learning problem where the decision maker interacts with a standard Markov decision process, with the exception that the reward functions vary arbitrarily over time. We show that, against every possible realization of the reward process, the agent can perform as well—in hindsight—as every stationary policy. This generalizes the classical no-regret result for repeated games. Specifically, we present an efficient online algorithm—in the spirit of reinforcement learning—that ensures that the agent's average performance loss vanishes over time, provided that the environment is oblivious to the agent's actions. Moreover, it is possible to modify the basic algorithm to cope with instances where reward observations are limited to the agent's trajectory. We present further modifications that reduce the computational cost by using function approximation and that track the optimal policy through infrequent changes.

Abstract. We consider the setting of stochastic bandit problems with a continuum of arms indexed by [0, 1] d . We first point out that the strategies considered so far in the literature only provided theoretical guarantees of the form: given some tuning parameters, the regret is small with respect to a class of environments that depends on these parameters. This is however not the right perspective, as it is the strategy that should adapt to the specific bandit environment at hand, and not the other way round. Put differently, an adaptation issue is raised. We solve it for the special case of environments whose mean-payoff functions are globally Lipschitz. More precisely, we show that the minimax optimal orders of magnitude L d/(d+2) T (d+1)/(d+2) of the regret bound over T time instances against an environment whose mean-payoff function f is Lipschitz with constant L can be achieved without knowing L or T in advance. This is in contrast to all previously known strategies, which require to some extent the knowledge of L to achieve this performance guarantee.

We consider a sequential decision problem where the rewards are generated by a piecewise-stationary distribution. However, the different reward distributions are unknown and may change at unknown instants. Our approach uses a limited number of side observations on past rewards, but does not require prior knowledge of the frequency of changes. In spite of the adversarial nature of the reward process, we provide an algorithm whose regret, with respect to the baseline with perfect knowledge of the distributions and the changes, is O(k log(T )), where k is the number of changes up to time T . This is in contrast to the case where side observations are not available, and where the regret is at least Ω( √ T ).