Event

MSSE Thesis Defense: Yash Bansod, "Refinement Acting vs. Simple Execution Guided by Hierarchical.."

Monday, April 19, 2021
10:00 a.m.
Online presentation
https://umd.zoom.us/j/5584232830

Refinement Acting vs. Simple Execution Guided by Hierarchical Planning

Yash Bansod

Zoom meeting: https://umd.zoom.us/j/5584232830

Advisory Committee
Professor Dana Nau, Chair/Advisor
Professor Michael Otte
Professor Pratap Tokekar

Abstract
Humans have always reasoned about complex problems by organizing them into hierarchical structures. One approach to artificial intelligence planning is to design intelligent agents capable of breaking complex problems into multiple levels of abstraction so that at any one level, the problem becomes small and simple. However, for an agent to reason at multiple levels of abstraction, it needs knowledge at those abstraction levels. Hierarchical Task Network (HTN) planning allows us to do precisely that. This thesis presents a novel HTN planning algorithm that uses iterative tree traversal to refine HTNs. We also develop a purely reactive HTN acting algorithm using a similar procedure. Preserving the hierarchy in HTN plans can be helpful during execution. We make use of this fact to develop an algorithm for integrated HTN planning and acting. We show through experiments that our algorithm is an improvement over a widely used approach to planning and control.