Meno:  Marek


Priezvisko:  Petrik


Názov:  Learning Parallel Portfolios of Algorithms


Vedúci:  RNDr. Mikular Popper


Rok:  2005


Blok:  UI


Kµúčové slová:  AI, Algorithms, Machine Learning, VapnikChervonenkis, Markov Decision Process, Portfolios


Abstrakt:  I present Parallel Portfolios of Algorithms (PPA) as an approach to enhance the performance of algorithms. The main idea is to execute several algorithms for the same problem concurrently. I show how to determine the optimal execution schedules of PPA with regard to a sample set of instances. I also provide theoretical distributionfree bounds for generalization probability, and a practical evaluation on SAT problem. In the SAT domain, PPA turned out to be generally 2 times as fast as the fastest algorithm in the portfolio.

