This chapter represents a first attempt tocharacterise the fitness landscapes of real-lifeGenetic Programming applications by means of apredictive algebraic difficulty indicator. Theindicator used is the Negative Slope Coefficient, whoseefficacy has been recently empirically demonstrated ona large set of hand-tailored theoretical test functionsand well known GP benchmarks. The real-life problemsstudied belong to the field of Biomedical applicationsand consist of automatically assessing a mathematicalrelationship between a set of molecular descriptorsfrom a given dataset of drugs and some importantpharmacokinetic parameters. The parameters consideredhere are Human Oral Bioavailability, Median Oral LethalDose, and Plasma Protein Binding levels. Theavailability of good prediction tools forpharmacokinetics parameters like these is critical foroptimising the efficiency of therapies, maximisingmedical success rate and minimizing toxic effects. Theexperimental results presented in this chapter showthat the Negative Slope Coefficient seems to be areasonable tool to characterise the difficulty of theseproblems, and can be used to choose the most effectiveGenetic Programming configuration (fitness function,representation, parameters' values) from a set of givenones.
CITATION STYLE
Vanneschi, L. (2007). Investigating Problem Hardness of Real Life Applications. In Genetic Programming Theory and Practice V (pp. 107–124). Springer US. https://doi.org/10.1007/978-0-387-76308-8_7
Mendeley helps you to discover research relevant for your work.