4 Algorithms

4.2. Random Search

The range of attributes and attribute values that are at all acceptable to the candidate preference states can be small and CDAP models composed of such states can easily predict that no state demands any of the brands in the competitive set. This is a degenerate solution. The GP algorithm rapidly selects out such models, but if their proportion in the initial population is high then the effective variation in the initial population is restricted to the extent that the success of the GP module will depend critically upon the small subset of viable genes in the initial population.

In order to make the GP search more robust, a preliminary operation was devised which directly affects only the generation of the initial population. This operation randomly generates genes tests them for non-degeneracy until an initial population of the required size has been found which contains only viable genes. The test for viability is quicker than a full evaluation of fitness so this is relatively inexpensive in terms of computation time and results in a more comprehensive sifting of possible solutions by the GP algorithm.

Artificially Intelligent Specification and Analysis of Context-Dependent Attribute Preferences - 03 NOV 97

[Next] [Previous] [Up] [Top] [Contents]

Generated with CERN WebMaker