Empowering Agents with Human Feedback- Integrating Implicit and Explicit Signals in Deep Reinforcement Learning
Abstract
The integration of autonomous agents in contemporary society has led to an increasing expectation for their operation within mixed human-AI domains. While the benefits of incorporating autonomous agents in our daily lives are evident, there are tasks that demand human interaction and collaborative decision-making. One of the main challenges in human-agent collaboration is the processing of large amounts of information under time pressure, which can significantly increase human task-load, ultimately affecting performance and decision-making. To address this issue, strategies such as task prioritization based on urgency, providing detailed explanations, and monitoring their emotion, workload reduction are crucial in maintaining a balance in mental workload. While current autonomous agents optimize their actions using traditional reinforcement learning models that focus on maximizing a singular objective score or reward, they do not consider important human factors such as emotional states, mental workload, and time constraints. To bridge this gap, our project proposes integrating explicit and implicit human feedback into reinforcement learning models. While explicit feedback signals have been extensively explored, implicit feedback and considering the task urgency and prioritization, has been relatively understudied. The core idea of our project is to develop a novel multi-objective framework that leverages both explicit and implicit human feedback, task urgency, and prioritization to enhance the learning and performance of autonomous agents in dynamic and complex environment. By considering these factors, the agents can adapt their behavior, offer explanations for task prioritization, and foster trust in human-agent interactions. Additionally, we propose the development of an Explainable AI framework that enables comparisons of multiple policies to evaluate the consequences of agent actions, providing detailed explanations of their decision-making process. By incorporating explicit and implicit human feedback, considering task urgency, and prioritization, and developing an Explainable AI.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Feb 16, 2024
- Source ID
- FA23862314087
Entities
People
- Bahareh Nakisa
Organizations
- Air Force Office of Scientific Research
- Deakin University
- United States Air Force