ENSEMBLES OF HYBRID INTELLIGENT EXPERTS: EXTENDING THE POWER OF OPTIMAL LINEAR COMBINERS Georgios C. Anagnostopoulos Dept. of Electrical zyxwv & Computer Engineering University of Central Florida Orlando, FL 32816 Email: gca@ece.engr.ucf.edu Michael Georgiopoulos Dept. of Electrical & Computer Engineering University of Central Florida Orlando, FL 32816 Email: mng@ece.engr.ucf.edu David Nic kerson Dept. of Statistics University of Central Florida Orlando, FL 32816 Email: nickersn@pegasus.cc.ucf.edu George Bebis Dept. of Mathematics & Computer Science University of Missouri at St. Louis St. Louis, MI 63121-4499 Email: bebis@mayura.cs.umsl.edu zyx ABSTRACT 1. INTRODUCTION zyx In the present paper we generalize the idea of Optimal Linear Combiners that are used to aggregate information from different sources providing estimates about a specific quantity. Two linear models are introduced, along with their analysis, which combine related components of information when more than one variable is to be predicted. The models' purpose is to produce point estimates of better accuracy in terms of mean squared error. Experimental results dealing with a functional approximation problem demonstrate that the generalized Optimal Linear Combiners suggested yield higher accuracy when compared to other combiners such as the Simple Average, or the conventional Optimal Linear Combiners. 53-1/97/$10.00 @ 1997 IEEE Estimation of zyxwv variables of interest (VI) whether involving point estimation or the estimation of an entire distribution, is a fundamental problem with vast numbers of applications such as time series forecasting, pattern recognition and functional approximation, to name a few. Quite often, a decision maker (DM) has access to a collection of experts (an ensemble) and is faced with the task of obtaining an optimal decision based on the estimates about a particular VI, that these experts supply. Being presented with a plethora of expert opinions, that do not necessarily coincide in all occasions, complicates the process of decision making. Some questions that arise are which experts should be ignored, which should be taken into account, and finally how to end up with a single opinion. For many years the classical approach (the so called naive approach) to all these problems was first to choose an estimation performance criterion like the Mean Square Error zy 1350