Bloat-aware GP-based methods with bloat quantification

摘要

Genetic programming (GP) solves optimization problems by simulating the evolution procedure in nature. It has a serious problem termed as bloat, which can cost memory, hamper effective breeding and slow down the evolution process. However, there are only a limited number of works to quantify bloat directly, and existing techniques use the solution size/complexity as an indirect indicator for bloat control. Therefore, a new bloat quantification measure is designed in this work, based on which three bloat-aware GP methods are proposed. Specifically, the bloat quantification measure is incorporated with two parsimony pressure techniques and a multi-objective technique respectively, termed as GPLTSb (GP Lexicographic Tournament Selection bloat), GPPTSb (GP Proportional Tournament Selection bloat), and MOGPb (Multi-objective GP bloat). Unlike the existing bloat control methods, the bloat-aware methods apply the bloat values directly for bloat control. The proposed methods are tested on benchmark symbolic regression tasks, and are compared with GP, existing bloat control methods and four widely-used regression methods. Results show that MOGPb is effective for bloat control with the solution size reduced obviously; while GPLTSb and GPPTSb can also reduce bloat in GP with the solution size reduced slightly. In addition, compared with GP and existing bloat control methods, the proposed methods evolve solutions with similar/better regression performance. Moreover, the evolved solutions of proposed methods can outperform most reference regression methods for the given tasks consistently.