TY - GEN

T1 - Derivative operators for preference predicate evolution

AU - Lorenz, David H.

PY - 1994

Y1 - 1994

N2 - This work deals with the problem of function learning by genetic algorithms where the function is used as a preference predicate. In such a case, learning the exact function is not necessary since any function that preserves the order induced by the target function is sufficient. The paper presents a methodology for solving the problem with genetic algorithms. We first consider the representation issues involved in learning such a function, and conclude that canonical representation, relative coding, and search restrictions, are required. We then show that the traditional homologous genetic operators are not appropriate for such learning, and introduce a new configurable analogous genetic operator, named derivative crossover. This operator works on the derivative of the chromosomes and is therefore suitable for preference predicate learning where only the relative values of the functions are important. We support our methodology by a set of experiments performed in the domain of continuous function learning and in the domain of evaluation-function learning for game-playing. The experiments show that indeed using derivative operators increases the speed of learning significantly.

AB - This work deals with the problem of function learning by genetic algorithms where the function is used as a preference predicate. In such a case, learning the exact function is not necessary since any function that preserves the order induced by the target function is sufficient. The paper presents a methodology for solving the problem with genetic algorithms. We first consider the representation issues involved in learning such a function, and conclude that canonical representation, relative coding, and search restrictions, are required. We then show that the traditional homologous genetic operators are not appropriate for such learning, and introduce a new configurable analogous genetic operator, named derivative crossover. This operator works on the derivative of the chromosomes and is therefore suitable for preference predicate learning where only the relative values of the functions are important. We support our methodology by a set of experiments performed in the domain of continuous function learning and in the domain of evaluation-function learning for game-playing. The experiments show that indeed using derivative operators increases the speed of learning significantly.

UR - http://www.scopus.com/inward/record.url?scp=85027500233&partnerID=8YFLogxK

U2 - 10.1007/3-540-58484-6_266

DO - 10.1007/3-540-58484-6_266

M3 - ???researchoutput.researchoutputtypes.contributiontobookanthology.conference???

AN - SCOPUS:85027500233

SN - 9783540584841

VL - 866

T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)

SP - 219

EP - 228

BT - Proceedings of the International Conference on Parallel Problem Solving from Nature (PPSN III)

A2 - Davidor, Yuval

A2 - Schwefel, Hans-Paul

A2 - Männer, Reinhard

PB - Springer Verlag

T2 - 3rd International Conference on Parallel Problem Solving from Nature, PPSN III 1994

Y2 - 9 October 1994 through 14 October 1994

ER -