Machine Learning and Friends Lunch





home
past talks
resources
conferences

Predicting the Difficulty of Factored Multi-Agent Domains


Stephan Murtagh
UMASS

Abstract

When presented with a novel domain it can be difficult to determine whether a particular technique will be successful. This is even more challenging in the case of factored domains where basic information such as the size and connectivity of the state-space is not known. In this talk, I will present on-going work with two techniques for estimating the difficulty of a domain. The first is a direct technique that estimates the number of paths in relaxations of a domain to produce upper and lower bounds on the original domain. The second is an indirect method that looks at the degree of independence in the domain, ie. the effect actions have on the future applicability of other actions.

Back to ML Lunch home