PhD Thesis Proposal - Robby Costales

Title: The Three-Tiered Exploration Problem in Open-Ended Adaptive Learning      
 
Committee members: Stefanos Nikolaidis (chair), Erdem Biyik, Stephen Tu, Willie Neiswanger, Daniel Seita      
 
Abstract:  A central challenge in training adaptive decision-making agents via meta-reinforcement learning (meta-RL) is meta-exploration—the search for an efficient exploration strategy that generalize to new, unseen tasks. Another bottleneck is the significant expense of manually designing training task distributions. While autocurricula methods—which automatically generate appropriately challenging training tasks for the learning agent—are well-studied in the standard RL setting, their application to meta-RL has been underexplored. These autocurricula approaches are a promising route for both (1) reducing the difficulty of meta-exploration and (2) removing the need for hand-designing tasks for meta-RL training, but the emergent training dynamics are complex—with each component mutually exacerbating each others' separate instabilities. In this talk, I outline a preliminary framework for understanding this combined learning problem, and present a research trajectory for addressing the associated challenges, building on my ongoing PhD work.        

Location: Ginsburg Hall (GCS) - 502C - 5th Floor

Audiences: Everyone Is Invited

Contact: Ellecia Williams

This event is open to all eligible individuals. USC Viterbi operates all of its activities consistent with the University's Notice of Non-Discrimination. Eligibility is not determined based on race, sex, ethnicity, sexual orientation, or any other prohibited factor.

Add to Google Calendar