Machine Learning and Natural Language Processing Lab

Diskriminatives... |--| Logische Markov... |--| SeqLog und Theo... |--| Logical Decisio... |--| Implementierung... |--| Learning from G... |--| Learning Patter... |--| SCGEM - A Fast ... |--| Informationsext... |--| Tranformation-B... |--| EM and Gradient... |--| KDD Cup 2002 - ... |--| Analysis of Res... |--| Objektorientier... |--|

Student's Project

SCGEM - A Fast Acceleration of EM

Jörg Fischer, 2003

Parameter estimation of Bayesian netwoks is a fundamental task to learn Bayesian netwoks. Traditionally, EM- and gradient-based algrithms are used. The Em algortihm is known to converge slowly when close to a sulution in contrast to the gradient ascent algorithm. Therefore, it is not surprising that the EM has been accelerated mainly basd on well-known gradient techniques. However, almost all EM-accelerations use a time consuming line search. In the present paper, we introduce a new EM-acceleration called SCGEM (Scaled Conjugate Generalised EM) which avoids the linesearch. Moreover, we provide experiments confirming that the SCGEM is superior to the EM in many cases.