Ahmed Rida
Artificial Intelligence group
Centre Universitaire d'Informatique
Rue General Dufour, 24
Geneva, Switzerland, CH-1211
E-mail: Ahmed.Rida@cui.unige.ch
Phone: +(41-22) 705-76-30
Fax: +(41-22) 705-77-80Abderrahim Labbi
Artificial Intelligence group
Centre Universitaire d'Informatique
Rue General Dufour, 24
Geneva, Switzerland, CH-1211
E-mail: Abderrahim.Labbi@cui.unige.ch
Phone: +(41-22) 705-76-42
Fax: +(41-22) 705-77-80Christian Pellegrini
Artificial Intelligence group
Centre Universitaire d'Informatique
Rue General Dufour, 24
Geneva, Switzerland, CH-1211
E-mail: Christian.Pellegrini@cui.unige.ch
Phone: +(41-22) 705-76-60
Fax: +(41-22) 705-77-80
In this paper we describe a divide-and-combine strategy for decomposition of a complex prediction problem into simpler local sub-problems. We firstly sh ow how to perform a soft decomposition via clustering of input data. Such decomposition leads to a partition of the input space into several regions wh ich may overlap. Therefore, to each region is assigned a local predictor (or exp ert) which is trained only on local data. To construct a solution to the glob al prediction problem, we combine the local experts using two approaches: weighted averaging where the outputs of local experts are weighted by their p rior densities, and nonlinear adaptive combination where the pooling pa rameters are obtained through minimization of a global error. To illustrate the validity of our approach, we show simulation results for two classification task s, vowels and phonemes, using local experts which are Multi-Layer Perceptrons (MLP) and Support Vector Machines (SVM). We compare the results obta ined using the two local combination modes with the results obtained using a g lobal predictor and a linear combination of global predictors.
local experts, nonlinear adaptive combination of experts, stacking, neural networks, support vector machines.
Pos tScript (171Kb)
gzipped PostScript (63Kb)