Comparison of the effectiveness of decimation and automatically defined functions

Decimation and automatically defined functions are intended to improve the fitness of the generated programs and to increase the rate of convergence to the solution. Each method has an associated computational cost, the cost for automatically defined functions being considerably higher than for decimation. This paper compares the performance improvements in genetic programming provided by automatically defined functions with that of decimation on four common benchmark problems - the Santa Fe ant, the lawnmower, even 3-bit parity and a symbolic regression problem. The results indicate that decimation provides improvement in performance that justifies the additional computation but the added computational effort required for automatically defined functions is not justified by any performance improvements.