Machine Learning and Friends Lunch





home
past talks
resources
conferences

Agent Space vs. Problem Space: Knowledge and Skill Transfer in Reinforcement Learning


George Konidaris
UMass

Abstract

One aspect of human intelligence that remains poorly understood is our ability to appropriately generalise knowledge and skills gained in one task to improve performance in another. One way to formalise this is to consider an agent that faces of a series of related but distinct tasks, and ask what kinds of mechanisms the agent can employ to use its experience in earlier tasks to improve its performance in later tasks. Unfortunately, even the of a "related task" is vague and a difficult one to formalise.

I will present an evolving framework for thinking about the problem of transfer in reinforcement learning agents that can be used to semi-formalise the notions of related and reward-linked tasks. I will demonstrate how this can be used to provide better estimates of a new task's value function and build portable skills that can be used in later tasks, and present some preliminary results.

Back to ML Lunch home