Beating the curse of dimensionality through manageable experiments

Modern computer and biological experiments require the simultaneous manipulation of hundreds of variables for characterization, modeling and –ultimately- optimization purposes. Most experimental design software packages are able to prescribe experimental arrays, however, the number of variables is usually limited to a few dozens. This work describes the first ideas on the generation of experimental designs involving a large number of variables with clustering techniques. The aim is to formulate a generation strategy that is implementable in a desktop computer with initial target applications topolymer computer simulations and microarray analysis for cancer characterization. A first example on the reduction of the exploration of 9 variables will be presented. At three levels per factor, a full factorial would require 19,683 experimental runs. Through the application of the proposed method, the number of runs would decrease to 60. Estimability of a second order regression model will be assessed in this first attempt.