Canada Spends Logo

Explore our Budget 2026 analysis

Grant: Discovery Launch Supplement

University of Toronto

$12,500.00

Statistical learning for decision making: instance-dependent optimality and efficient algorithms

FY 2024-25

Summary
Data-driven methodology is a pillar of real-world decision-making. When applying statistical learning methods, puzzling phenomena have arisen in choosing estimators, tuning their parameters, and characterizing bias-variance trade-offs. There are various settings in which asymptotic and/or worst-case theory fails to provide the relevant guidance, so that a more refined approach, both non-asymptotic and instance-optimal, is required. The proposed research focuses on addressing key statistical challenges in machine learning for data-driven decision-making, covering modern approaches to reinforcement learning, causal effect estimation, bandit problems, as well as game theory. In particular, we aim at developing theories and methodologies from the following three aspects: 1. Instance-optimal statistical learning in decision-making. A popular approach in data-driven decision-making is function approximation, i.e., first estimate some functions, and then compute the target value or optimal policy from the learned model. This framework raises two important questions: how to train the learning model, and how to leverage it in decision-making. If we treat the two stages separately and deploy existing methods, the resulting method will lead to substantial waste of data. In the proposed research, we will bring statistical optimality consideration on the agenda, and develop methods that are information-theoretically optimal for the end-goal in decision-making systems. Furthermore, we will develop estimators that are not only worst-case optimal, but also optimal in an instance-dependent sense, making them adaptive to the difficulty of any given problem instances. 2. Efficient computational algorithms with large-scale data. Stochastic optimization have been the workhorse for large-scale statistical learning, while several new puzzles arise from decision-making applications. The proposed research aims at developing novel algorithms for these applications. Apart from existing challenges such as non-convexity and non-smoothness, we will investigate the key structures behind the new phenomena and puzzles, such as geometry of fixed-point equations and non-iid data. We will focus on the interplay between computation and statistics -- instead of just minimizing the empirical loss, we want the algorithm to generate outputs achieving statistical optimality in downstream applications. 3. Flexible and robust framework for statistical inference. Building upon optimal procedures, we also aim at developing frameworks for valid and powerful statistical inference. When learning models are in use, the asymptotics can be mis-leading for a finite sample size. So we aim at providing non-asymptotic guarantees. Furthermore, inference in decision-making problems can be challenging due to the non-standard geometric structures, including non-smooth functionals, and slow semi-parametric rates. We aim at addressing these challenges via novel theoretical tools.
Program
Discovery Launch Supplement
Location
Ontario, Canada
Recipient Type
Academia
Type
Grant
Source
NSERC's Awards Data

As published

Application id
DGECR-2024-00306
Name nom
Mou, Wenlong
Department département
Department of Statistical Sciences
Organization id
31
Institution établissement
University of Toronto
Province en
Ontario
Province fr
Ontario
Country en
CANADA
Country fr
CANADA
Fiscal year exercice financier
2024
Competition year année de concours
2024
Award amount
12500
Program id
DGECR
Program name en
Discovery Launch Supplement
Program name fr
Supplément Tremplin vers la découverte
Group en
DISCOVERY RESEARCH
Group fr
RECHERCHE AXÉE SUR LA DÉCOUVERTE
Committee code
1508
Committee name en
Mathematics and Statistics
Committee name fr
Mathématiques et statistique
Area of application code
1200
Area of application group en
Advancement of knowledge
Area of application group fr
Avancement des connaissances
Area of application en
Advancement of knowledge
Area of application fr
Avancement des connaissances
Research subject code
3001
Research subject group en
Statistics and probability
Research subject group fr
Statistique et probabilité
Research subject en
Statistical theory
Research subject fr
Théorie de statistique
Field of research list names en
Not available
Field of research list names fr
Non disponible
Socioeconomic objective list names en
Not available
Socioeconomic objective list names fr
Non disponible
Application title
Statistical learning for decision making: instance-dependent optimality and efficient algorithms
Keywords
statistical learning theory; reinforcement learning and stochastic control; nonparametric and semiparametric estimation; data-driven decision making; causal inference; off-policy estimation; high-dimensional statistics; statistical minimax theory; stochastic approximation and optimization; bandit algorithms and active learning
Application summary
Data-driven methodology is a pillar of real-world decision-making. When applying statistical learning methods, puzzling phenomena have arisen in choosing estimators, tuning their parameters, and characterizing bias-variance trade-offs. There are various settings in which asymptotic and/or worst-case theory fails to provide the relevant guidance, so that a more refined approach, both non-asymptotic and instance-optimal, is required. The proposed research focuses on addressing key statistical challenges in machine learning for data-driven decision-making, covering modern approaches to reinforcement learning, causal effect estimation, bandit problems, as well as game theory. In particular, we aim at developing theories and methodologies from the following three aspects: 1. Instance-optimal statistical learning in decision-making. A popular approach in data-driven decision-making is function approximation, i.e., first estimate some functions, and then compute the target value or optimal policy from the learned model. This framework raises two important questions: how to train the learning model, and how to leverage it in decision-making. If we treat the two stages separately and deploy existing methods, the resulting method will lead to substantial waste of data. In the proposed research, we will bring statistical optimality consideration on the agenda, and develop methods that are information-theoretically optimal for the end-goal in decision-making systems. Furthermore, we will develop estimators that are not only worst-case optimal, but also optimal in an instance-dependent sense, making them adaptive to the difficulty of any given problem instances. 2. Efficient computational algorithms with large-scale data. Stochastic optimization have been the workhorse for large-scale statistical learning, while several new puzzles arise from decision-making applications. The proposed research aims at developing novel algorithms for these applications. Apart from existing challenges such as non-convexity and non-smoothness, we will investigate the key structures behind the new phenomena and puzzles, such as geometry of fixed-point equations and non-iid data. We will focus on the interplay between computation and statistics -- instead of just minimizing the empirical loss, we want the algorithm to generate outputs achieving statistical optimality in downstream applications. 3. Flexible and robust framework for statistical inference. Building upon optimal procedures, we also aim at developing frameworks for valid and powerful statistical inference. When learning models are in use, the asymptotics can be mis-leading for a finite sample size. So we aim at providing non-asymptotic guarantees. Furthermore, inference in decision-making problems can be challenging due to the non-standard geometric structures, including non-smooth functionals, and slow semi-parametric rates. We aim at addressing these challenges via novel theoretical tools.