Periodic Reporting for period 1 - CONVEY (Conveying Agent Behavior to People: A User-Centered Approach to Explainable AI)
Okres sprawozdawczy: 2023-01-01 do 2025-06-30
From self-driving cars to agents recommending medical treatment, Artificial Intelligence (AI) agents are becoming increasingly prevalent. These agents have the potential to benefit society in areas such as transportation, healthcare and education. Importantly, they do not operate in a vacuum --- people interact with agents in a wide range of settings. To effectively interact with agents, people need to be able to anticipate and understand their behavior. For example, a driver of an autonomous vehicle will need to anticipate situations in which the car fails and hands over control, while a clinician will need to understand the treatment regime recommended by an agent to determine whether it aligns with the patient's preferences.
Explainable AI methods aim to support users by making the behavior of AI systems more transparent. However, the state-of-the-art in explainable AI is lacking in several key aspects. First, the majority of existing methods focus on providing "local" explanations to one-shot decisions of machine learning models. They are not adequate for conveying the behavior of agents that act over an extended time duration in large state spaces.
Second, most existing methods do not consider the context in which explanations are deployed, including to the specific needs and characteristics of users. Finally, most methods are not interactive, limiting users' ability to gain a thorough understanding of the agents.
The overarching objective of this proposal is to develop adaptive and interactive methods for conveying the behavior of agents and multi-agent teams operating in sequential decision-making settings.
To tackle this challenge, the proposed research will draw on insights and methodologies from AI and human-computer interaction. It will develop algorithms that determine what information about agents' behavior to share with users, tailored to users' needs and characteristics, and interfaces that allow users to proactively explore agents' capabilities.
Explainable AI methods aim to support users by making the behavior of AI systems more transparent. However, the state-of-the-art in explainable AI is lacking in several key aspects. First, the majority of existing methods focus on providing "local" explanations to one-shot decisions of machine learning models. They are not adequate for conveying the behavior of agents that act over an extended time duration in large state spaces.
Second, most existing methods do not consider the context in which explanations are deployed, including to the specific needs and characteristics of users. Finally, most methods are not interactive, limiting users' ability to gain a thorough understanding of the agents.
The overarching objective of this proposal is to develop adaptive and interactive methods for conveying the behavior of agents and multi-agent teams operating in sequential decision-making settings.
To tackle this challenge, the proposed research will draw on insights and methodologies from AI and human-computer interaction. It will develop algorithms that determine what information about agents' behavior to share with users, tailored to users' needs and characteristics, and interfaces that allow users to proactively explore agents' capabilities.
Aim 1: Adapting to users' tasks and reasoning
We have developed several new explainability methods: (1) Counterfactual summaries that demonstrate an agent's alternative actions, helping users better discern agent goals and reasoning; (2) Policy summaries that support human operators in determining when to intervene and take over an agent, and (3) Textual policy summaries utilizing large language models which describe main patterns in agent behavior. We conducted user studies to assess each of these approaches, showing promising results.
In addition, we began an interdisciplinary project exploring the use of large language models to support the training of scientists in science communication. Here, the model generates explanations to scientists that suggest ways to better communicate their research.
Aim 2: Designing interactive interfaces for exploration of agent behavior
We developed ASQ-IT – an interactive explanation system that presents video clips of the agent acting in its environment based on queries given by the user that describe temporal properties of behaviors of interest. User studies show that end-users can understand and formulate queries in ASQ-IT and that using ASQ-IT assists users in identifying faulty agent behaviors.
We are currently exploring ways to use large language models to allow users to pose queries in a more natural way, building on our work on textual summaries.
We have developed several new explainability methods: (1) Counterfactual summaries that demonstrate an agent's alternative actions, helping users better discern agent goals and reasoning; (2) Policy summaries that support human operators in determining when to intervene and take over an agent, and (3) Textual policy summaries utilizing large language models which describe main patterns in agent behavior. We conducted user studies to assess each of these approaches, showing promising results.
In addition, we began an interdisciplinary project exploring the use of large language models to support the training of scientists in science communication. Here, the model generates explanations to scientists that suggest ways to better communicate their research.
Aim 2: Designing interactive interfaces for exploration of agent behavior
We developed ASQ-IT – an interactive explanation system that presents video clips of the agent acting in its environment based on queries given by the user that describe temporal properties of behaviors of interest. User studies show that end-users can understand and formulate queries in ASQ-IT and that using ASQ-IT assists users in identifying faulty agent behaviors.
We are currently exploring ways to use large language models to allow users to pose queries in a more natural way, building on our work on textual summaries.
Our results so far demonstrate the benefit of tailoring explanation methods to users’ needs, and the potential to develop more interactive methods that allow users to explore agent behavior. With recent developments in AI and in particular, in large language models, it is expected that a growing number of increasingly complex tasks will be supported or completed by agents. Therefore, such explanation methods will be crucial and likely also required from a regulatory perspective at some point.
The introduction of large language models also poses new explainability challenges, as these are usually highly complex and often proprietary models.
The introduction of large language models also poses new explainability challenges, as these are usually highly complex and often proprietary models.