Reinforcement Learning: Objectives, Feedback, and Improvement

The Lab for Applied AI at Loyola University Chicago is pleased to be co-sponsoring this one-day workshop with IDEAL, bringing together researchers working across reinforcement learning, control, robotics, learning theory, and foundation models.

Reinforcement learning provides a general framework for improving behavior through interaction, but many contemporary applications begin with agents that already possess substantial capabilities and with objectives that are difficult to specify directly. A robot policy may first be learned from demonstrations before being improved through interaction, while a large language model is pretrained and instruction-tuned before reinforcement learning is used to further shape its behavior. These settings raise a common set of questions: Where does the objective come from when desired behavior is easier to demonstrate or evaluate than to specify? How should demonstrations, preferences, and other forms of feedback be incorporated into learning? When can interaction improve upon a policy learned from data, and how can that improvement remain reliable as the policy moves beyond the distribution on which it was trained? These questions connect longstanding ideas from optimal and inverse optimal control, bandit and online learning, and learning theory with contemporary work in reinforcement learning, robotics, and foundation models.