Comment optimiser A* adaptatif
Lou Fedon and Antoine Cornuéjols
In: RFIA-08 (Reconnaissance des Formes et Intelligence Artificielle), 23-25 Jan 2008, Amiens, France.

Abstract

Efficient search of (quasi-)optimal paths in graphs remains a fundamental task in Artificial Intelligence. Recent works \cite{KLF:04,KL:05,Koenig-Likhachev:06b,KLS:07} have contributed to a new point of view on this problem whereby heuristics are learned from past solving experiences rather than derived through a static abstraction of the description of the problem. In this paper, we show how to improve this work by better exploiting information from past solving episodes. The experiments reported here confirm the significant reduction in search space achieved by our algorithm. In a second part, we show how to generalize these learning techniques to the case of changing goal states. Extensive experiments and their analysis show that the variations of the goal states must obey strict laws in order for these adaptive A$^{*}$ algorithms to be advantageous.

PDF - Requires Adobe Acrobat Reader or other PDF viewer.

[Edit]