Events Calendar

PhD Defense - Liyu Chen

Title: Understanding Goal-oriented Reinforcement Learning

Defense committee: Haipeng Luo, David Kempe, Ashutosh Nayyar, Rahul Jain

Date: January 12, 2022. 1 p.m. - 2 p.m. PST

Zoom meeting link: https://usc.zoom.us/j/93602803008

Abstract:
Reinforcement Learning (RL) studies how an agent learns to behave optimally in an unknown environment. One challenge in applying RL in practice is task specification, that is, how to inform the algorithm of the task we want it to solve. My research focuses on solving the problem of task specification in goal-oriented reinforcement learning (GoRL), whose objective is to reach a goal state with the smallest possible cost. Unlike standard RL that focuses solely on cost minimization, GoRL has dual objectives: 1) reach the goal state and 2) minimize the cost.

Standard RL models such as the finite-horizon model or the discounted model often have difficulty in specifying tasks for GoRL, which leads to heavy engineering efforts in practice.

To tackle this issue, we study learning in a Markov Decision Process named stochastic shortest path (SSP), which exactly captures the dual objectives of GoRL. We focus on developing practical learning algorithms for SSP. Specifically, we study the PAC learning setting for SSP, and develop various reduction schemes that connect SSP to the simpler finite-horizon model. Our reduction schemes help to develop optimal and efficient online learning algorithms for SSP.

Audiences: Everyone Is Invited

Contact: Caitlin Dawson