In repeated games, strategies are often evaluated by their ability to guarantee the performance of the single best action that is selected in hindsight. Yet, the efficacy of the single best action as a benchmark is limited, as static actions may perform poorly in common dynamic settings. We propose the notion of dynamic benchmark (DB) consistency and we characterize the possible empirical joint distributions of play that may emerge when all players are relying on DB consistent strategies.

TBD

We present a framework for optimizing the selection of which content to promote on digital distribution channels, consisting of: (1) A model of the decision-problem, (2) An off-policy identification and evaluation method, (3) the design of ranking policies, and (4) a causal machine learning procedure. We partner with an international newspaper and show that our optimal policy improves the newspapers outcome by 18 percent. Our work contributes by supporting the curation of digital channels.

Tree-based models have interpretability but are not able to capture complex relationships while other ML models often lack interpretability. We propose neural-informed decision trees (NIDTs) to combine predictive power with interpretability. We evaluate NIDTs on over 20 UCI datasets and show they outperform multiple ML benchmarks. We demonstrate interpretability by extracting an explainable warfarin prescription policy and show how they can be used on a pricing problem.

Platforms run many A/B tests, but testing all combinations is impractical. DeDL, our debiased deep learning framework, estimates causal effects and identifies optimal treatments. It combines deep learning with double machine learning, yielding consistent estimators. Results on a video-sharing platform show accurate estimation and efficient iteration. DeDL enables platforms to iterate operations effectively.

Deep learning systems use huge training data sets to meet desired performances, but over/under-collecting training data can incur unnecessary costs and workflow delays. We propose and then solve an optimal data collection problem incorporating performance targets, collection costs, a time horizon, and penalties. Experiments on six deep learning tasks show that we reduce the risks of failing to meet performance targets by over 2x compared to existing estimation-based heuristics.

Learning optimal policies from historical data enables personalization in many applications. Adaptive data collection is becoming more common for allowing to improve inferential efficiency and optimize operational performance, but adaptivity complicates policy learning ex post. Our work complements the literature by learning policies with adaptively collected data. We propose an algorithm with proven finite-sample regret bound, which is minimax optimal and meets our established lower bound.

We study the performance implications of quality and quantity of data in contextual decision-making. We focus on the Newsvendor loss and consider a data-driven model in which outcomes observed in similar contexts have similar distributions. We characterize exactly the worst-case regret of a classical class of kernel policies. Our exact analysis unveils new structural insights on the learning behavior of these policies that cannot be observed through state-of-the-art general purpose bounds.

Counter-trafficking efforts are often conducted with a local scope. However, sex traffickers commonly move individuals geographically. Thus, there is a need for organizations in different geographical regions to collaborate in counter-trafficking efforts. This research leverages a large dataset of commercial sex ads and techniques from the areas of machine learning and network science to identify prominent geographical circuits and individuals operating on these circuits.

Community corrections provide alternatives for incarcerations, which can reduce jail overcrowding and recidivism rate, particularly for individuals with substance use disorder. We study the placement decisions for community corrections and relevant capacity planning via an MDP model and prove structural properties for policy insights. To address the complex dependence between optimal placement and system congestion, we leverage a two-timescale approach to develop algorithmic solutions.

This paper presents a data-driven approach to mitigate industrial plant air pollution on nearby cities by linking operational decisions with weather conditions. Our method combines predictive and prescriptive machine learning models to forecast short-term wind patterns and recommend production adjustments. Implemented at a chemical plant in Morocco, our algorithm improves weather forecasts by 40-50% and offers valuable trade-offs, reducing emissions by 33% and costs by 63%.

TBD

TBD

We consider the dynamic scheduling problem of a classical single-server multi-class queueing system where the system manager does not have full knowledge about the cost/reward of the customers. One of the key results is that the Whittle index policy is optimal for a two-class queue if the system manager knows the distribution of the reward of one class and dynamically learns the distribution parameters of the other class.

TBD

TBD

Our work provides new results on the arm-sampling behavior of the celebrated UCB family of multi-armed bandit algorithms, leading to several important insights. Among these, it is shown that arm-sampling rates under UCB are asymptotically deterministic, regardless of the problem complexity. This discovery facilitates new sharp asymptotic characterizations revealing profound distinctions between UCB and Thompson Sampling such as an "incomplete learning" phenomenon characteristic of the latter.

We study an off-policy contextual pricing problem where the seller has access to samples of prices that customers were previously offered, whether they purchased at that price, and auxiliary features. This is in contrast to the well-studied setting in which samples of the customer's valuation are observed. We focus on convex loss functions for pricing in this setting, prove expected revenue bounds when the valuation distribution is log-concave, and provide generalization bounds.

We study the design of stochastic optimization methods with a focus on guaranteed out-of-sample performance when data is corrupted. We design a novel robust approach that offers protection against corruption while ensuring strong generalization. Our approach is based on distributionally robust optimization with a combination of Kullback-Leibler and Levy-Prokhorov ambiguity sets. Our method is applied to training neural networks, resulting in robust networks with state-of-the-art performance.

The paper presents dynamic pricing with demand learning in presence of covariates and develops safe UCB pricing algorithms. The pricing algorithm minimizes total regret where the expected regret of the pricing policy should be at least as good as a fraction of that of the baseline policy. We extend it to the case that the cumulative reward of the algorithm is at least as high as a fraction of the cumulative reward of the baseline policy with known and unknown baseline expected reward.

We consider a problem where customers repeatedly interact with a platform. The probability that a customer engages depends on past purchase history. The platform maximizes the total revenue obtained from each customer over the horizon. We study the dynamic program when consumer preferences are known and prove structural properties. We provide a formulation in a contextual episodic RL setting and prove a regret bound. We evaluate effectiveness on simulations, using real data from Expedia.

In a feedback collection process, a company sequentially and adaptively shows display sets to a population of customers and collects their choices. The objective is to identify the most preferred item at a high confidence level with the least number of samples. We propose an elimination-based algorithm, namely Nested Elimination (NE). NE is intuitive, simple in structure, easy to implement, and has a strong theoretical/numerical performance for sample complexity.

Our study focuses on a preference learning setting where customers select k most preferred items from a personalized set. Our new Mallows-type ranking model offers simple closed-form (ranked) choice probabilities that can be learned through MLE with theoretical guarantees. We demonstrate the model's excellent performance with real data sets. We use our model to study how feedback structure is related to the efficiency of the feedback collection and find that a little favor goes a long way.

This paper applies a machine learning lens to the problem of choice estimation. We (1) establish the first truly realistic-scale benchmark for practical choice estimation, (2) use this benchmark to run the largest evaluation of choice models to date, and (3) propose and prescribe the use of a simple, irrational choice model, which we dub the Sparse Halo-MNL.

When allocating products to brick-and-mortar stores, retailers face product availability constraints that force them to balance product offerings across stores. We model this problem under multinomial logit demand and show that the problem is NP-complete. We develop a tractable continuous relaxation of the problem which has a unique local maximum and allows us to build near-optimal solution algorithms. We use data from a large retailer and identify improvements of better product-store matching.

We consider a decision-maker (DM) who can sample from multiple information sources to learn a state before making an earning decision. The DM optimizes his sampling and earning decisions to maximize his payoffs. The problem is motivated by financial and healthcare applications with multiple information sources. We develop a Bayesian control framework for this problem and solve it in the estimation and testing settings. We also develop an efficient algorithm for the general control setting.

TBD