What Matters in Hierarchical Search for Combinatorial Reasoning Problems?

TL;DR

This paper investigates hierarchical search strategies, especially subgoal methods, for solving complex combinatorial reasoning problems, identifying key attributes that influence their effectiveness and providing a standardized evaluation framework.

Contribution

It offers an in-depth analysis of subgoal-planning methods, identifies critical attributes for their success, and proposes a consistent evaluation methodology for comparing algorithms.

Findings

High-level search benefits depend on specific problem attributes.

Certain attributes like complex action spaces and dead ends enhance subgoal method effectiveness.

A standardized evaluation framework enables fair comparison of algorithms.

Abstract

Efficiently tackling combinatorial reasoning problems, particularly the notorious NP-hard tasks, remains a significant challenge for AI research. Recent efforts have sought to enhance planning by incorporating hierarchical high-level search strategies, known as subgoal methods. While promising, their performance against traditional low-level planners is inconsistent, raising questions about their application contexts. In this study, we conduct an in-depth exploration of subgoal-planning methods for combinatorial reasoning. We identify the attributes pivotal for leveraging the advantages of high-level search: hard-to-learn value functions, complex action spaces, presence of dead ends in the environment, or using data collected from diverse experts. We propose a consistent evaluation methodology to achieve meaningful comparisons between methods and reevaluate the state-of-the-art algorithms.