.png){kind=icon}
返回
Decomposition based cross-parallel multiobjective genetic programming for symbolic regression
.png){kind=icon}
PRE
AI
DOI：10.1016/j.asoc.2024.112239.png)
.png){kind=icon}
原文求助.png){kind=icon}
分享.png){kind=icon}
收藏
摘要
En 中文
Genetic Programming (GP) based Symbolic Regression (SR) algorithms suffer from the ineluctable effects over model bloat, blind search and diversity loss when determining explicit symbolic models to best depict the concealed laws in historical data, which often make them time-consuming and unstable. Most efforts often dealt with one of these effects, such that the algorithms still suffer from other effects. To deal with these effects, we propose a cross-parallel SR algorithm framework based on problem decomposition and multiobjective GP in this paper. The decomposition method is proposed to distill simple subproblems named global and local regression, which can be fast solved to produce various high quality models. In the proposed framework, by expressing the SR problem as the multiobjective optimization model, a number of subproblems are automatically distilled and solved in parallel to reduce model bloat and accelerate the algorithm. Traditional regression methods are employed to produce high quality models to seed the evolving populations for each subtask to maintain population diversity and improve search efficiency. Elite models obtained by each subtasks will be collected and randomly sent to other subtasks to improve the model generalization. Ablation and comparison experiments are conducted to evaluate the performance of the proposed algorithms. The ablation results show that the developed algorithm framework plays a positive role in reducing above ineluctable effects, and can fast determine concise symbolic models for the benchmarks. Comparisons by SRBench demonstrate the effectiveness of the developed algorithm on wide range problems.
Keyword:
Symbolic regression
Problem decomposition
Global regression
Local regression
Cross-parallel multiobjective genetic
programming
期刊
IF:
6.6
论文数:
1.4W
被引数:
4.8W
机构
引用论文
Studying bloat control and maintenance of effective code in linear genetic programming for symbolic regression.png)
NEUROCOMPUTING
NEUROCOMPUTINGIF6.5
PREAI
分享收藏
查看更多内容