We design a new family of estimators for off-policy evaluation in contextual bandits. Our estimators are based on the asymptotically optimal approach of doubly robust estimation, but they shrink importance weights to obtain a better bias-variance tradeoff in finite samples. Our approach adapts importance weights to the quality of a reward predictor, interpolating between doubly robust estimation and direct modeling. When the reward predictor is poor, we recover previously studied weight clipping, but when the reward predictor is good, we obtain a new form of shrinkage. To navigate between these regimes and tune the shrinkage coefficient, we design a model selection procedure, which we prove is never worse than the doubly robust estimator. Extensive experiments on bandit benchmark problems show that our estimators are highly adaptive and typically outperform state-of-the-art methods.

We design a new algorithm for batch active learning with deep neural network models. Our algorithm, Batch Active learning by Diverse Gradient Embeddings (BADGE), samples groups of points that are disparate and high-magnitude when represented in a hallucinated gradient space, a strategy designed to incorporate both predictive uncertainty and sample diversity into every selected batch. Crucially, BADGE trades off between diversity and uncertainty without requiring any hand-tuned hyperparameters. We show that while other approaches sometimes succeed for particular batch sizes or architectures, BADGE consistently performs as well or better, making it a versatile option for practical active learning problems.

We introduce the problem of model selection for contextual bandits, wherein a learner must adapt to the complexity of the optimal policy while balancing exploration and exploitation. Our main result is a new model selection guarantee for linear contextual bandits. We work in the stochastic realizable setting with a sequence of nested linear policy classes of dimension $d_1 < d_2 < \ldots$, where the $m^\star$-th class contains the optimal policy, and we design an algorithm that achieves $\tilde{O}(T^{2/3}d^{1/3}_{m^\star})$ regret with no prior knowledge of the optimal dimension $d_{m^\star}$. The algorithm also achieves regret $\tilde{O}(T^{3/4} + \sqrt{Td_{m^\star}})$, which is optimal for $d_{m^{\star}}\geq{}\sqrt{T}$. This is the first contextual bandit model selection result with non-vacuous regret for all values of $d_{m^\star}$ and, to the best of our knowledge, is the first guarantee of its type in any contextual bandit setting. The core of the algorithm is a new estimator for the gap in best loss achievable by two linear policy classes, which we show admits a convergence rate faster than what is required to learn either class.

We design an active learning algorithm for cost-sensitive multiclass classification: problems where different errors have different costs. Our algorithm, COAL, makes predictions by regressing to each label's cost and predicting the smallest. On a new example, it uses a set of regressors that perform well on past data to estimate possible costs for each label. It queries only the labels that could be the best, ignoring the sure losers. We prove COAL can be efficiently implemented for any regression family that admits squared loss optimization; it also enjoys strong guarantees with respect to predictive performance and labeling effort. We empirically compare COAL to passive learning and several active learning baselines, showing significant improvements in labeling effort and test cost on real-world datasets.

We consider contextual bandits: a setting in which a learner repeatedly makes an action on the basis of contextual information and observes a loss for the action, with the goal of minimizing cumulative loss over a series of rounds. Contextual bandit learning has received much attention, and has seen substantial success in practice (e.g., Auer et al., 2002; Langford and Zhang, 2007; Agarwal et al., 2014, 2017). This line of work mostly considers small, finite action sets, yet in many real-world problems actions are chosen from from an interval, so the set is continuous and infinite. How can we learn to make actions from continuous spaces based on loss-only feedback? We could assume that nearby actions have similar losses, for example that the losses are Lipschitz continuous as a function of the action (following Agrawal, 1995, and a long line of subsequent work).

We study the exploration problem in episodic MDPs with rich observations generated from a small number of latent states. Under certain identifiability assumptions, we demonstrate how to estimate a mapping from the observations to latent states inductively through a sequence of regression and clustering steps---where previously decoded latent states provide labels for later regression problems---and use it to construct good exploration policies. We provide finite-sample guarantees on the quality of the learned state decoding function and exploration policies, and complement our theory with an empirical evaluation on a class of hard exploration problems. Our method exponentially improves over $Q$-learning with na\"ive exploration, even when $Q$-learning has cheating access to latent states.

We use surrogate losses to obtain several new regret bounds and new algorithms for contextual bandit learning. Using the ramp loss, we derive a new margin-based regret bound in terms of standard sequential complexity measures of a benchmark class of real-valued regression functions. Using the hinge loss, we derive an efficient algorithm with a $\sqrt{dT}$-type mistake bound against benchmark policies induced by $d$-dimensional regressors. Under realizability assumptions, our results also yield classical regret bounds.

We study the computational tractability of PAC reinforcement learning with rich observations. We present new provably sample-efficient algorithms for environments with deterministic hidden state dynamics and stochastic rich observations. These methods operate in an oracle model of computation -- accessing policy and value function classes exclusively through standard optimization primitives -- and therefore represent computationally efficient alternatives to prior algorithms that require enumeration. With stochastic hidden state dynamics, we prove that the only known sample-efficient algorithm, OLIVE, cannot be implemented in the oracle model. We also present several examples that illustrate fundamental challenges of tractable PAC reinforcement learning in such general settings.

We use surrogate losses to obtain several new regret bounds and new algorithms for contextual bandit learning. Using the ramp loss, we derive a new margin-based regret bound in terms of standard sequential complexity measures of a benchmark class of real-valued regression functions. Using the hinge loss, we derive an efficient algorithm with a $\sqrt{dT}$-type mistake bound against benchmark policies induced by $d$-dimensional regressors. Under realizability assumptions, our results also yield classical regret bounds.

We study the computational tractability of PAC reinforcement learning with rich observations. We present new provably sample-efficient algorithms for environments with deterministic hidden state dynamics and stochastic rich observations. These methods operate in an oracle model of computation -- accessing policy and value function classes exclusively through standard optimization primitives -- and therefore represent computationally efficient alternatives to prior algorithms that require enumeration. With stochastic hidden state dynamics, we prove that the only known sample-efficient algorithm, OLIVE, cannot be implemented in the oracle model. We also present several examples that illustrate fundamental challenges of tractable PAC reinforcement learning in such general settings.