An Anti-Noise Method for Function Mining Based on GEP

Mining functions from experimental data based on traditional gene expression programming （GEP） fitness mechanism falls short in handling noises, which may lead to anamorphic results The contributions of this paper include: （1）Proposing a new concept called weak adaptive model （WAM） based on GEP to break the limitation, which is enlightened by the biologic nature known as “seek advantage, avoid disadvantage”; （2）Presenting new concepts “In Band set” and “Out Band set” for partitioning the training data set; （3）Designing a new approach called Relative Error Fitness Algorithm （REFA） to mine functions in terms of WAM; and （4）By using extensive experiments demonstrating the effectiveness of REFA The results show that when mining functions in a dataset with 3 33% noise data, REFA increases the success probability by 3 times and decreases the average relative error from 7 899% to 2 320% compared with the traditional approach