CAREER: A Unified Architecture for Learning of, and Reasoning with, Task Models: Theory and Applications

Lehigh University, Bethlehem PA

Investigators

Abstract

Proposal 0642882 "CAREER: A Unified Architecture for Learning of, and reasoning with, Task Models: Theory and Applications" PI: Hector Munoz-Avila Lehigh University The focus of this project is the use of Hierarchical Task Networks (HTN) in planning. Humans learn complex skills by incrementally acquiring and combining simpler skills into skills of increasing complexity. Consequentially, many cognitive architectures use hierarchical models to represent relations between skills of different complexity. Over the years, the problem of automatically learning such hierarchical models has taken a backstage. Frequently, it is assumed that these models are given and research has instead concentrated on their use. This project drops this assumption and studies ways to learn the hierarchical models and use them for automated planning. The centerpiece of this project is a unified architecture for (1) learning hierarchical models from STRIPS plans and for (2) a multi-phase combination of lazy and eager learning using a common hierarchical task network (HTN) plan representation. Lazy learning produces hierarchical knowledge that directly represents the input examples. Eager learning produces hierarchical knowledge that generalizes multiple input examples. The unified architecture learns and plans with constructs whose generalization varies between these two extremes. The project also incorporates educational activities based on the multi-phase learning principles formulated in this work. This includes a pilot study, aimed at G6-12 students, to provide hands-on experience on introductory AI. Other broader impacts of this work include possible uses of HTN planning techniques in a wide range of potential applications that are amenable to hierarchical modeling, including project management for public events, software engineering project planning, and developing strategies for automated players in computer games.