Genetic programming with incremental data inheritance

摘要

A data-driven method for accelerating genetic programming is presented. This method, called incremental data inheritance or IDI for short, evolves programs using program-specific subsets of given data which also evolve incrementally as generation goes on. The concept of data evolution in IDI is contrasted to conventional genetic programming in which all the given training data are used repeatedly. IDI is also distinguished from the previous subset selection methods in that each program in IDI evolves its own data set of incremental size rather than a common data set of fixed or arbitrary size for the whole population. The method has been applied to time series prediction. Compared to the conventional methods, IDI significantly reduced the evolution speed of genetic programming without loss of the generalization accuracy of evolved programs. We also provide a theoretical foundation of the IDI method from the Bayesian inference point of view.