{"title":"政策学习的 Bandit 算法:方法、实施和福利绩效","authors":"Toru Kitagawa, Jeff Rowley","doi":"arxiv-2409.00379","DOIUrl":null,"url":null,"abstract":"Static supervised learning-in which experimental data serves as a training\nsample for the estimation of an optimal treatment assignment policy-is a\ncommonly assumed framework of policy learning. An arguably more realistic but\nchallenging scenario is a dynamic setting in which the planner performs\nexperimentation and exploitation simultaneously with subjects that arrive\nsequentially. This paper studies bandit algorithms for learning an optimal\nindividualised treatment assignment policy. Specifically, we study\napplicability of the EXP4.P (Exponential weighting for Exploration and\nExploitation with Experts) algorithm developed by Beygelzimer et al. (2011) to\npolicy learning. Assuming that the class of policies has a finite\nVapnik-Chervonenkis dimension and that the number of subjects to be allocated\nis known, we present a high probability welfare-regret bound of the algorithm.\nTo implement the algorithm, we use an incremental enumeration algorithm for\nhyperplane arrangements. We perform extensive numerical analysis to assess the\nalgorithm's sensitivity to its tuning parameters and its welfare-regret\nperformance. Further simulation exercises are calibrated to the National Job\nTraining Partnership Act (JTPA) Study sample to determine how the algorithm\nperforms when applied to economic data. Our findings highlight various\ncomputational challenges and suggest that the limited welfare gain from the\nalgorithm is due to substantial heterogeneity in causal effects in the JTPA\ndata.","PeriodicalId":501293,"journal":{"name":"arXiv - ECON - Econometrics","volume":"22 1","pages":""},"PeriodicalIF":0.0000,"publicationDate":"2024-08-31","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":"{\"title\":\"Bandit Algorithms for Policy Learning: Methods, Implementation, and Welfare-performance\",\"authors\":\"Toru Kitagawa, Jeff Rowley\",\"doi\":\"arxiv-2409.00379\",\"DOIUrl\":null,\"url\":null,\"abstract\":\"Static supervised learning-in which experimental data serves as a training\\nsample for the estimation of an optimal treatment assignment policy-is a\\ncommonly assumed framework of policy learning. An arguably more realistic but\\nchallenging scenario is a dynamic setting in which the planner performs\\nexperimentation and exploitation simultaneously with subjects that arrive\\nsequentially. This paper studies bandit algorithms for learning an optimal\\nindividualised treatment assignment policy. Specifically, we study\\napplicability of the EXP4.P (Exponential weighting for Exploration and\\nExploitation with Experts) algorithm developed by Beygelzimer et al. (2011) to\\npolicy learning. Assuming that the class of policies has a finite\\nVapnik-Chervonenkis dimension and that the number of subjects to be allocated\\nis known, we present a high probability welfare-regret bound of the algorithm.\\nTo implement the algorithm, we use an incremental enumeration algorithm for\\nhyperplane arrangements. We perform extensive numerical analysis to assess the\\nalgorithm's sensitivity to its tuning parameters and its welfare-regret\\nperformance. Further simulation exercises are calibrated to the National Job\\nTraining Partnership Act (JTPA) Study sample to determine how the algorithm\\nperforms when applied to economic data. Our findings highlight various\\ncomputational challenges and suggest that the limited welfare gain from the\\nalgorithm is due to substantial heterogeneity in causal effects in the JTPA\\ndata.\",\"PeriodicalId\":501293,\"journal\":{\"name\":\"arXiv - ECON - Econometrics\",\"volume\":\"22 1\",\"pages\":\"\"},\"PeriodicalIF\":0.0000,\"publicationDate\":\"2024-08-31\",\"publicationTypes\":\"Journal Article\",\"fieldsOfStudy\":null,\"isOpenAccess\":false,\"openAccessPdf\":\"\",\"citationCount\":\"0\",\"resultStr\":null,\"platform\":\"Semanticscholar\",\"paperid\":null,\"PeriodicalName\":\"arXiv - ECON - Econometrics\",\"FirstCategoryId\":\"1085\",\"ListUrlMain\":\"https://doi.org/arxiv-2409.00379\",\"RegionNum\":0,\"RegionCategory\":null,\"ArticlePicture\":[],\"TitleCN\":null,\"AbstractTextCN\":null,\"PMCID\":null,\"EPubDate\":\"\",\"PubModel\":\"\",\"JCR\":\"\",\"JCRName\":\"\",\"Score\":null,\"Total\":0}","platform":"Semanticscholar","paperid":null,"PeriodicalName":"arXiv - ECON - Econometrics","FirstCategoryId":"1085","ListUrlMain":"https://doi.org/arxiv-2409.00379","RegionNum":0,"RegionCategory":null,"ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"","JCRName":"","Score":null,"Total":0}
引用次数: 0
摘要
静态监督学习--即实验数据作为估计最优治疗分配政策的训练样本--是一种常见的政策学习假设框架。一个可以说更现实但更具挑战性的场景是动态环境,在这种环境中,规划者同时对按顺序到达的受试者进行实验和开发。本文研究了学习最优个体化治疗分配政策的匪徒算法。具体来说,我们研究了 Beygelzimer 等人(2011 年)开发的 EXP4.P(Exponential weighting for Exploration andExploitation with Experts)算法在政策学习中的适用性。假定政策类别具有有限的 Vapnik-Chervonenkis 维度,且待分配的研究对象数量已知,我们提出了该算法的高概率福利-遗憾约束。为了实现该算法,我们使用了一种针对超平面安排的增量枚举算法。我们进行了大量的数值分析,以评估该算法对其调整参数的敏感性及其福利-遗憾表现。此外,我们还根据《国家就业培训合作法案》(JTPA)研究样本进行了进一步的模拟练习,以确定该算法在应用于经济数据时的表现。我们的研究结果凸显了各种计算挑战,并表明该算法的福利收益有限是由于 JTPA 数据中因果效应的巨大异质性造成的。
Bandit Algorithms for Policy Learning: Methods, Implementation, and Welfare-performance
Static supervised learning-in which experimental data serves as a training
sample for the estimation of an optimal treatment assignment policy-is a
commonly assumed framework of policy learning. An arguably more realistic but
challenging scenario is a dynamic setting in which the planner performs
experimentation and exploitation simultaneously with subjects that arrive
sequentially. This paper studies bandit algorithms for learning an optimal
individualised treatment assignment policy. Specifically, we study
applicability of the EXP4.P (Exponential weighting for Exploration and
Exploitation with Experts) algorithm developed by Beygelzimer et al. (2011) to
policy learning. Assuming that the class of policies has a finite
Vapnik-Chervonenkis dimension and that the number of subjects to be allocated
is known, we present a high probability welfare-regret bound of the algorithm.
To implement the algorithm, we use an incremental enumeration algorithm for
hyperplane arrangements. We perform extensive numerical analysis to assess the
algorithm's sensitivity to its tuning parameters and its welfare-regret
performance. Further simulation exercises are calibrated to the National Job
Training Partnership Act (JTPA) Study sample to determine how the algorithm
performs when applied to economic data. Our findings highlight various
computational challenges and suggest that the limited welfare gain from the
algorithm is due to substantial heterogeneity in causal effects in the JTPA
data.