Concurrent and Temporal Composition for Zero-Shot Transfer in Reinforcement Learning

Abstract

An agent can autonomously learn goal-oriented value functions that combine to solve new goal tasks specified as Boolean expressions, optimally and without further learning. This talk shows how these skills can also be composed over time, allowing the agent to satisfy complex temporal specifications, such as regular fragments of linear temporal logic, and achieve zero-shot transfer to unseen tasks.

Steven James

Associate Professor

My research focuses on artificial intelligence and reinforcement learning, in particular designing agents capable of learning transferable, high-level abstractions of their environment.