12.12.2023
Imitation Learning and Value Alignment Under Mismatch and Constraints
Short bio:
Dr. Sebastian Tschiatschek is assistant professor at the University of Vienna. He is part of the research group for Data Mining and Machine Learning, leading the work group for Probabilistic and Interactive Machine Learning. He is also an alumni of TU Graz.
Abstract:
Reinforcement learning has been successfully used for training AI agents when given clearly defined reward signals. However, reinforcement learning can suffer from high sample complexity and is challenging to use in settings in which the reward signal is hard to specify, e.g., for training personal agents that should maximize a human-defined reward function. We investigate imitation learning and reinforcement learning from human feedback which allows us to avoid the specification of a reward function and can drastically reduce sample complexity. Inspired by real-world applications, we particularly study settings in which there is a mismatch between the human providing demonstrations or feedback and the learning AI agent, e.g., in terms of observations, capabilities, or constraints. Our insights emphasize the need for adaptation of the human and the AI agent to achieve the best alignment regarding the human's goals. Furthermore, we propose practical algorithms to perform this adaptation and sample efficient approaches for learning about rewards and constraints.
Photo: ©University of Vienna
Dr. Sebastian Tschiatschek is assistant professor at the University of Vienna. He is part of the research group for Data Mining and Machine Learning, leading the work group for Probabilistic and Interactive Machine Learning. He is also an alumni of TU Graz.
Abstract:
Reinforcement learning has been successfully used for training AI agents when given clearly defined reward signals. However, reinforcement learning can suffer from high sample complexity and is challenging to use in settings in which the reward signal is hard to specify, e.g., for training personal agents that should maximize a human-defined reward function. We investigate imitation learning and reinforcement learning from human feedback which allows us to avoid the specification of a reward function and can drastically reduce sample complexity. Inspired by real-world applications, we particularly study settings in which there is a mismatch between the human providing demonstrations or feedback and the learning AI agent, e.g., in terms of observations, capabilities, or constraints. Our insights emphasize the need for adaptation of the human and the AI agent to achieve the best alignment regarding the human's goals. Furthermore, we propose practical algorithms to perform this adaptation and sample efficient approaches for learning about rewards and constraints.
Photo: ©University of Vienna