When Do Intrinsic Rewards Lead to Exploration? | Scott Viteri (Stanford)

​This week, Cornell AI Alignment will join the Probability Seminar for a technical talk by Scott Viteri, a CS PhD candidate at Stanford University’s Center for AI Safety.

​Talk: When Do Intrinsic Rewards Lead to Exploration?

​Intrinsic rewards are widely used in reinforcement learning to encourage agents to explore, for example by rewarding prediction error, learning progress, novelty, or information gain. But does maximizing an intrinsic reward actually lead an agent to seek out the most informative experiences?

​Scott will present a formal framework for comparing exploration policies based on the information they acquire, and show that several common intrinsic-reward objectives can favor policies that are strictly worse at exploration. He’ll also discuss conditions under which existing approaches do work, and an objective that more directly rewards improved exploration.

​When: Monday, October 5, 4:00-5:00 PM
Where: Malott 406

​This is a technical talk and will be most useful for attendees with some background in probability, machine learning, or reinforcement learning.