Canada Spends Logo

Explore our Budget 2026 analysis

Grant: Discovery Grants Program - Individual

McGill University

$69,000.00

Responsible Decision Making in Dynamic Environments

FY 2023-24

Summary
Algorithmic systems based on AI are increasingly used in sensitive applications, such as credit lending, advertisement, education, healthcare, autonomous driving and others. The long-term promise of these systems is to accelerate, augment, improve and/or automate human decision-making, by leveraging data and machine learning algorithms. However there is mounting evidence that AI systems currently deployed or under development can be unsafe or unstable, some are vulnerable to privacy or security breaches, and they can also reproduce or exacerbate biases against certain groups. There is growing awareness of these problems, and this has led to many contributions on building socially responsible machine learning systems. However, most of this literature is focused on a static scenario, where the machine learning system is trained on a given finite dataset, and performance is assessed on a fixed hold-out test set. In practice, many real-world systems are dynamic, involving a sequence of observations and decisions, which lead to distribution shifts, feedback loops, and long-term delayed effects. These sequential processes are typically studied through the reinforcement learning (RL) paradigm. We have a rich literature on models and algorithms for RL, but much of our results to date are focused on learning efficiently and we have few methods to address the broader requirements for sensitive applications, such as safety, fairness, privacy, security, robustness. The long-term vision is to revisit the RL paradigm (models, algorithms, theoretical & empirical results) and develop new methods that address the requirements of sensitive applications. We will enhance the RL framework with the ability to jointly optimize for a standard reward function, as well as other desired properties. To achieve this, we must go beyond the classical explore-exploit trade-off (which balances efficiency and performance), to a multi-dimensional trade-off incorporating other dimensions of responsibility. Our methodological approach will build on the Constrained Markov Decision Process framework, which provides mathematical flexibility to incorporate a variety of complementary objectives, in addition to a core reward function. We will explore both exact and approximate solution methods and seek to provide theoretical guarantees in the form of bounds on performance across desired properties. Throughout the work, we will consider a variety of relevant applications. While it is beyond the scope of this grant to provide key-in-hand solutions for deployed systems, by working on concrete applications, and whenever possible with real dynamic systems, considering concrete applications serves to motivate and verify the validity of our assumptions and the feasibility of our proposed solutions. We will focus on applications in healthcare, where use of AI-assisted dynamic decision-making is increasing, and having stronger guarantees for safety, robustness and fairness is highly desirable.
Program
Discovery Grants Program - Individual
Location
Quebec, Canada
Recipient Type
Academia
Type
Grant
Source
NSERC's Awards Data

As published

Application id
RGPIN-2023-03524
Name nom
Pineau, Joelle
Department département
School of Computer Science
Organization id
61
Institution établissement
McGill University
Province en
Québec
Province fr
Québec
Country en
CANADA
Country fr
CANADA
Fiscal year exercice financier
2023
Competition year année de concours
2023
Award amount
69000
Program id
RGPIN
Program name en
Discovery Grants Program - Individual
Program name fr
Programme de subventions à la découverte - individuelles
Group en
DISCOVERY RESEARCH
Group fr
RECHERCHE AXÉE SUR LA DÉCOUVERTE
Committee code
1507
Committee name en
Computer Science
Committee name fr
Sciences informatiques
Area of application code
800
Area of application group en
Information and communication services
Area of application group fr
Services d'information et de télécommunications
Area of application en
Information and communication services
Area of application fr
Services d'information et de télécommunications
Research subject code
9999
Research subject group en
Not available
Research subject group fr
Not available
Research subject en
Non disponible
Research subject fr
Non disponible
Application title
Responsible Decision Making in Dynamic Environments
Keyword
AI fairness; AI safety; Reinforcement learning; Responsible AI
Application summary
Algorithmic systems based on AI are increasingly used in sensitive applications, such as credit lending, advertisement, education, healthcare, autonomous driving and others. The long-term promise of these systems is to accelerate, augment, improve and/or automate human decision-making, by leveraging data and machine learning algorithms. However there is mounting evidence that AI systems currently deployed or under development can be unsafe or unstable, some are vulnerable to privacy or security breaches, and they can also reproduce or exacerbate biases against certain groups. There is growing awareness of these problems, and this has led to many contributions on building socially responsible machine learning systems. However, most of this literature is focused on a static scenario, where the machine learning system is trained on a given finite dataset, and performance is assessed on a fixed hold-out test set. In practice, many real-world systems are dynamic, involving a sequence of observations and decisions, which lead to distribution shifts, feedback loops, and long-term delayed effects. These sequential processes are typically studied through the reinforcement learning (RL) paradigm. We have a rich literature on models and algorithms for RL, but much of our results to date are focused on learning efficiently and we have few methods to address the broader requirements for sensitive applications, such as safety, fairness, privacy, security, robustness. The long-term vision is to revisit the RL paradigm (models, algorithms, theoretical & empirical results) and develop new methods that address the requirements of sensitive applications. We will enhance the RL framework with the ability to jointly optimize for a standard reward function, as well as other desired properties. To achieve this, we must go beyond the classical explore-exploit trade-off (which balances efficiency and performance), to a multi-dimensional trade-off incorporating other dimensions of responsibility. Our methodological approach will build on the Constrained Markov Decision Process framework, which provides mathematical flexibility to incorporate a variety of complementary objectives, in addition to a core reward function. We will explore both exact and approximate solution methods and seek to provide theoretical guarantees in the form of bounds on performance across desired properties. Throughout the work, we will consider a variety of relevant applications. While it is beyond the scope of this grant to provide key-in-hand solutions for deployed systems, by working on concrete applications, and whenever possible with real dynamic systems, considering concrete applications serves to motivate and verify the validity of our assumptions and the feasibility of our proposed solutions. We will focus on applications in healthcare, where use of AI-assisted dynamic decision-making is increasing, and having stronger guarantees for safety, robustness and fairness is highly desirable.