When resources are scarce, an allocation policy is needed to decide who receives a resource. This problem occurs, for instance, when allocating scarce medical resources and is often solved using modern ML methods. This paper introduces methods to evaluate index-based allocation policies -- that allocate a fixed number of resources to those who need them the most -- by using data from a randomized control trial. Such policies create dependencies between agents, which render the assumptions behind standard statistical tests invalid and limit the effectiveness of estimators. Addressing these challenges, we translate and extend recent ideas from the statistics literature to present an efficient estimator and methods for computing asymptotically correct confidence intervals. This enables us to effectively draw valid statistical conclusions, a critical gap in previous work. Our extensive experiments validate our methodology in practical settings, while also showcasing its statistical power. We conclude by proposing and empirically verifying extensions of our methodology that enable us to reevaluate a past randomized control trial to evaluate different ML allocation policies in the context of a mHealth program, drawing previously invisible conclusions.
Restless multi-armed bandits (RMABs) are used to model sequential resource allocation in public health intervention programs. In these settings, the underlying transition dynamics are often unknown a priori, requiring online reinforcement learning (RL). However, existing methods in online RL for RMABs cannot incorporate properties often present in real-world public health applications, such as contextual information and non-stationarity. We present Bayesian Learning for Contextual RMABs (BCoR), an online RL approach for RMABs that novelly combines techniques in Bayesian modeling with Thompson sampling to flexibly model a wide range of complex RMAB settings, such as contextual and non-stationary RMABs. A key contribution of our approach is its ability to leverage shared information within and between arms to learn unknown RMAB transition dynamics quickly in budget-constrained settings with relatively short time horizons. Empirically, we show that BCoR achieves substantially higher finite-sample performance than existing approaches over a range of experimental settings, including one constructed from a real-world public health campaign in India.
There is a great desire to use adaptive sampling methods, such as reinforcement learning (RL) and bandit algorithms, for the real-time personalization of interventions in digital applications like mobile health and education. A major obstacle preventing more widespread use of such algorithms in practice is the lack of assurance that the resulting adaptively collected data can be used to reliably answer inferential questions, including questions about time-varying causal effects. Current methods for statistical inference on such data are insufficient because they (a) make strong assumptions regarding the environment dynamics, e.g., assume a contextual bandit or Markovian environment, or (b) require data to be collected with one adaptive sampling algorithm per user, which excludes data collected by algorithms that learn to select actions by pooling the data of multiple users. In this work, we make initial progress by introducing the adaptive sandwich estimator to quantify uncertainty; this estimator (a) is valid even when user rewards and contexts are non-stationary and highly dependent over time, and (b) accommodates settings in which an online adaptive sampling algorithm learns using the data of all users. Furthermore, our inference method is robust to misspecification of the reward models used by the adaptive sampling algorithm. This work is motivated by our work designing experiments in which RL algorithms are used to select actions, yet reliable statistical inference is essential for conducting primary analyses after the trial is over.
The theory of reinforcement learning currently suffers from a mismatch between its empirical performance and the theoretical characterization of its performance, with consequences for, e.g., the understanding of sample efficiency, safety, and robustness. The linear quadratic regulator with unknown dynamics is a fundamental reinforcement learning setting with significant structure in its dynamics and cost function, yet even in this setting there is a gap between the best known regret lower-bound of $\Omega_p(\sqrt{T})$ and the best known upper-bound of $O_p(\sqrt{T}\,\text{polylog}(T))$. The contribution of this paper is to close that gap by establishing a novel regret upper-bound of $O_p(\sqrt{T})$. Our proof is constructive in that it analyzes the regret of a concrete algorithm, and simultaneously establishes an estimation error bound on the dynamics of $O_p(T^{-1/4})$ which is also the first to match the rate of a known lower-bound. The two keys to our improved proof technique are (1) a more precise upper- and lower-bound on the system Gram matrix and (2) a self-bounding argument for the expected estimation error of the optimal controller.
Conjoint analysis is a popular experimental design used to measure multidimensional preferences. Researchers examine how varying a factor of interest, while controlling for other relevant factors, influences decision-making. Currently, there exist two methodological approaches to analyzing data from a conjoint experiment. The first focuses on estimating the average marginal effects of each factor while averaging over the other factors. Although this allows for straightforward design-based estimation, the results critically depend on the distribution of other factors and how interaction effects are aggregated. An alternative model-based approach can compute various quantities of interest, but requires researchers to correctly specify the model, a challenging task for conjoint analysis with many factors and possible interactions. In addition, a commonly used logistic regression has poor statistical properties even with a moderate number of factors when incorporating interactions. We propose a new hypothesis testing approach based on the conditional randomization test to answer the most fundamental question of conjoint analysis: Does a factor of interest matter in any way given the other factors? Our methodology is solely based on the randomization of factors, and hence is free from assumptions. Yet, it allows researchers to use any test statistic, including those based on complex machine learning algorithms. As a result, we are able to combine the strengths of the existing design-based and model-based approaches. We illustrate the proposed methodology through conjoint analysis of immigration preferences and political candidate evaluation. We also extend the proposed approach to test for regularity assumptions commonly used in conjoint analysis.
In this work, we analyze an efficient sampling-based algorithm for general-purpose reachability analysis, which remains a notoriously challenging problem with applications ranging from neural network verification to safety analysis of dynamical systems. By sampling inputs, evaluating their images in the true reachable set, and taking their $\epsilon$-padded convex hull as a set estimator, this algorithm applies to general problem settings and is simple to implement. Our main contribution is the derivation of asymptotic and finite-sample accuracy guarantees using random set theory. This analysis informs algorithmic design to obtain an $\epsilon$-close reachable set approximation with high probability, provides insights into which reachability problems are most challenging, and motivates safety-critical applications of the technique. On a neural network verification task, we show that this approach is more accurate and significantly faster than prior work. Informed by our analysis, we also design a robust model predictive controller that we demonstrate in hardware experiments.
Reward functions are at the heart of every reinforcement learning (RL) algorithm. In robotic grasping, rewards are often complex and manually engineered functions that do not rely on well-justified physical models from grasp analysis. This work demonstrates that analytic grasp stability metrics constitute powerful optimization objectives for RL algorithms that refine grasps on a three-fingered hand using only tactile and joint position information. We outperform a binary-reward baseline by 42.9% and find that a combination of geometric and force-agnostic grasp stability metrics yields the highest average success rates of 95.4% for cuboids, 93.1% for cylinders, and 62.3% for spheres across wrist position errors between 0 and 7 centimeters and rotational errors between 0 and 14 degrees. In a second experiment, we show that grasp refinement algorithms trained with contact feedback (contact positions, normals, and forces) perform up to 6.6% better than a baseline that receives no tactile information.
Bandit algorithms are increasingly used in real-world sequential decision-making problems. Associated with this is an increased desire to be able to use the resulting datasets to answer scientific questions like: Did one type of ad lead to more purchases? In which contexts is a mobile health intervention effective? However, classical statistical approaches fail to provide valid confidence intervals when used with data collected with bandit algorithms. Alternative methods have recently been developed for simple models (e.g., comparison of means). Yet there is a lack of general methods for conducting statistical inference using more complex models on data collected with (contextual) bandit algorithms; for example, current methods cannot be used for valid inference on parameters in a logistic regression model for a binary reward. In this work, we develop theory justifying the use of M-estimators -- which includes estimators based on empirical risk minimization as well as maximum likelihood -- on data collected with adaptive algorithms, including (contextual) bandit algorithms. Specifically, we show that M-estimators, modified with particular adaptive weights, can be used to construct asymptotically valid confidence regions for a variety of inferential targets.
Bandit algorithms are increasingly used in real world sequential decision making problems, from online advertising to mobile health. As a result, there are more datasets collected using bandit algorithms and with that an increased desire to be able to use these datasets to answer scientific questions like: Did one type of ad increase the click-through rate more or lead to more purchases? In which contexts is a mobile health intervention effective? However, it has been shown that classical statistical approaches, like those based on the ordinary least squares estimator, fail to provide reliable confidence intervals when used with bandit data. Recently methods have been developed to conduct statistical inference using simple models fit to data collected with multi-armed bandits. However there is a lack of general methods for conducting statistical inference using more complex models. In this work, we develop theory justifying the use of M-estimation (Van der Vaart, 2000), traditionally used with i.i.d data, to provide inferential methods for a large class of estimators -- including least squares and maximum likelihood estimators -- but now with data collected with (contextual) bandit algorithms. To do this we generalize the use of adaptive weights pioneered by Hadad et al. (2019) and Deshpande et al. (2018). Specifically, in settings in which the data is collected via a (contextual) bandit algorithm, we prove that certain adaptively weighted M-estimators are uniformly asymptotically normal and demonstrate empirically that we can use their asymptotic distribution to construct reliable confidence regions for a variety of inferential targets.
Recent progress in reinforcement learning has led to remarkable performance in a range of applications, but its deployment in high-stakes settings remains quite rare. One reason is a limited understanding of the behavior of reinforcement algorithms, both in terms of their regret and their ability to learn the underlying system dynamics---existing work is focused almost exclusively on characterizing rates, with little attention paid to the constants multiplying those rates that can be critically important in practice. To start to address this challenge, we study perhaps the simplest non-bandit reinforcement learning problem: linear quadratic adaptive control (LQAC). By carefully combining recent finite-sample performance bounds for the LQAC problem with a particular (less-recent) martingale central limit theorem, we are able to derive asymptotically-exact expressions for the regret, estimation error, and prediction error of a rate-optimal stepwise-updating LQAC algorithm. In simulations on both stable and unstable systems, we find that our asymptotic theory also describes the algorithm's finite-sample behavior remarkably well.