Markovian Bandits

The multi-armed bandit problem models the trade-off between exploration and exploitation. In this paper, we study the variant where exploration has a hard budget and precedes exploitation. Prior guarantees include a 1/3 approximation by Guha and Munagala using Lagrangian duality, and a 4/27 approximation by Ma against a primal linear programming relaxation for a more general model. We show that the analyses of both algorithms are tight, even when their relaxations are asymptotically exact. We provide a different efficient algorithm that achieves a 0.43 approximation. We then provide an efficient algorithm whose guarantee approaches 1-e^{-1} approx 0.63 for bounded exchangeable reward observations in a non-vanishing regime, which is asymptotically tight against natural linear programming relaxations. The technique also gives certificates for stopping sequential model evaluation.

We study the minimax estimation of the pseudo-spectral gap $γ_{ps}$ of a Markov chain from the observation of a single trajectory. Extending earlier works on finite-state space chains by Wolfer and Kontorovich (2024), we provide an estimator in the case of a compact, continuous state-space. Our estimator appears to be minimax-optimal on the class of chains with a transition operator with Lipschitz density, up to logarithmic terms. Our estimator, combined with Paulin's version of Bernstein inequality, allows to derive empirical concentration inequalities for Markov chains. This unlocks many applications, from confidence intervals, to active learning and generalization bounds. In particular, for rested bandits with Markov rewards, we provide a version of UCB with a regret bound similar to the one in the iid case, without assuming any knowledge on the pseudo-spectral gap of the chains.

Exploration--exploitation is a central trade-off in bandit learning. While classical algorithms such as upper confidence bound methods and Thompson Sampling effectively balance this trade-off in stationary environments, their exploration strategies mainly reduce uncertainty about the current optimal arm, which can be insufficient in nonstationary settings where future optimal arms may differ substantially from current ones. In this paper, we propose Future Information-Directed Sampling (FIDS), a new algorithm for Bayesian nonstationary bandits that explicitly explores to gather information about future optimal arms. We show that FIDS achieves regret comparable to Thompson Sampling up to a small constant factor, while being able to exploit predictive information structures that conventional exploration objectives fail to capture. To address the practical difficulty of posterior inference, we further propose a supervised-learning-based approximation framework that learns the FIDS policy from offline data, and demonstrate its effectiveness on synthetic benchmarks.

Exhaustively evaluating every candidate LLM configuration on every benchmark item to identify a high-performing one is costly. We formulate configuration selection as a cost-aware Bayesian bandit problem and propose GittinsEval, which draws on the Bayesian-optimal Gittins policy to determine which configuration to evaluate next and when to stop. We extend the policy with an anytime recommendation rule over both fully and partially evaluated configurations, using an LCB-style score to account for posterior uncertainty. GittinsEval is computationally efficient, requiring only lightweight online updates after offline precomputation. Across GSM8K, PIQA, AlpacaEval, and MMLU response matrices, GittinsEval is consistently competitive, with particularly strong gains over configuration-level Bayesian optimization on large-example benchmarks and over cost-unaware bandit baselines on large-candidate tasks. Crucially, GittinsEval often attains near-zero simple regret using only 1% to 2% of the exhaustive-evaluation cost; it also offers an adaptive stopping rule that typically triggers at 1% to 10%.

This thesis studies policy learning in interactive systems where an agent observes a context, selects an action from a very large set, and receives partial feedback. The main framework is contextual bandits, with two paradigms: on-policy learning, where the agent interacts sequentially with the environment and minimizes regret, and off-policy learning, where it learns from logged data collected by a logging policy. In large action spaces, both settings face major challenges: inefficient exploration, sparse data coverage, high-variance importance weights, extrapolation bias, and difficult optimization landscapes. The first part develops structured Bayesian methods for on-policy learning. We introduce meTS, a mixed-effect extension of Thompson sampling, and dTS, which leverages diffusion-inspired priors to model dependencies between actions. These methods share information across actions and yield regret guarantees depending on an effective number of actions. The second part addresses off-policy learning. We propose sDM, a structured direct method based on latent variables, show that optimization error can dominate estimation error in large action spaces, and introduce concave, efficiently optimizable policy-weighted log-likelihood objectives. Finally, we develop differentiable pessimistic methods based on exponential smoothing and PAC-Bayesian bounds to control the bias-variance trade-off of regularized importance-sampling estimators.

This paper studies the problem of regret minimization in Markovian bandits with non-observable states and possibly constrained decision epochs. The focus is restricted to a ``pure'' regret benchmark, that compares the performance of the learning algorithm to the best pure policy which -- akin to optimal policies of stochastic bandits -- picks the optimal arm from start to finish without ever switching. We introduce a generalization of rested Markovian bandits, self-degrading Markovian bandits, for which pure policies are always asymptotically optimal.We show that without prior knowledge on the underlying bandit, the regret of algorithms that switch arms rarely necessarily scales super-logarithmically for every bandit, i.e., as $ω(\log(T))$, where $T$ is the learning horizon. Despite the unreachability of the logarithmic regime, we design UCB-NOM, an optimistic algorithm inspired by UCB, of which the regret is nearly logarithmic. Lastly, we show that given prior knowledge on the Markovian bandit in the form of a bound on the bias functions of its arm, a proper instantiation of UCB-NOM achieves $O(\log(T))$ regret. We further show that this prior knowledge allows for a $O(\sqrt{T \log(T)})$ worst-case regret bound for UCB-NOM. Notably, our regret bounds do not depend on the number of states of the underlying Markov chains. Our findings suggest that the non-observability of states is a mild inconvenience in self-degrading Markovian bandits.

Recent work shows that when contexts are drawn i.i.d., linear contextual bandits can be reduced to single-context linear bandits. This ``contexts are cheap'' perspective is highly advantageous, as it allows for sharper finite-time analyses and leverages mature techniques from the linear bandit literature, such as those for misspecification and adversarial corruption. However, this reduction crucially relies on the independence of contexts and does not extend to settings with temporally correlated (e.g., Markovian) contexts, which arise frequently in practice. Motivated by applications with temporally correlated availability, we extend this perspective to linear bandits with Markovian context processes, where the action set evolves via an exogenous Markov chain. Our main contribution is a reduction that applies under uniform geometric ergodicity. We construct a stationary surrogate action set to solve the problem using a standard linear bandit oracle, employing a delayed-update scheme to control the bias induced by the nonstationary conditional context distributions. We further provide a phased algorithm for unknown stationary distributions that learns the surrogate mapping online. In both settings, we obtain a high-probability worst-case regret bound matching that of the underlying linear bandit oracle in sufficiently fast mixing regimes. We then validate our results on a real-world instance, where we show practical gains over a LinUCB baseline.

Online model selection in Bayesian bandits raises a fundamental exploration challenge: When an environment instance is sampled from a prior distribution, how can we design an adaptive strategy that explores multiple bandit learners and competes with the best one in hindsight? We address this problem by introducing a new Bayesian algorithm for online model selection in stochastic bandits. We prove an oracle-style guarantee of $O\left( d^* M \sqrt{T} + \sqrt{(MT)} \right)$ on the Bayesian regret, where $M$ is the number of base learners, $d^*$ is the regret coefficient of the optimal base learner, and $T$ is the time horizon. We also validate our method empirically across a range of stochastic bandit settings, demonstrating performance that is competitive with the best base learner. Additionally, we study the effect of sharing data among base learners and its role in mitigating prior mis-specification.