CISE Seminar: Jyotirmoy Deshmukh, USC

Date: Sept 30, 2022
Time: 3:00PM-4:00PM

Jyotirmoy Deshmukh
Assistant Professor,
Computer Science, USC

Symbolic Reward Machines or: How I Stopped Worrying and Started Loving Reinforcement Learning

Reinforcement learning (RL) is a popular technique in the AI and robotics domains to obtain control policies for autonomous agents operating in uncertain environments. The basic setup of RL is as follows: in any given state, the autonomous agent takes an action, and the environment stochastically decides the next state for the agent and also gives it a reward. Rewards are defined locally, and are associated with the agent transitions or states. RL algorithms attempt to find control policies that maximize the cumulative reward obtained by the agent. Recent work in these domains has focused on model-free RL, where a further challenge is that the model of the environment is unknown. A crucial element of RL is reward design: ill-designed rewards can make the agents learn undesirable or unsafe control policies – unacceptable in safety-critical systems. In the formal methods community, techniques like reactive synthesis have focused on correct-by-construction design of control policies given a model of the environment and formal specifications in a suitable temporal logic. In this talk, we will look at some recent work on how we can leverage powerful specification formalisms with model-free learning-based methods to get control policies that are safer, and in some cases, accompanied with safety guarantees.

Jyotirmoy V. Deshmukh (Jyo) is an Assistant Professor in the Department of Computer Science at the University of Southern California, and the co-Director of the Center for Autonomy and AI. Before joining USC, Jyo worked as a Principal Research Engineer at Toyota R&D. He got his Ph.D. from the University of Texas at Austin in 2010, and is the recipient of the 2021 NSF Career Award and the 2021 Amazon Research Award.

Faculty Host: Wenchao Li
Student Host: Andres Chavez Armijos