Improving Modularity in Genetic Programming Using Graph-Based Data Mining

We propose to improve the efficiency of genetic programming, a method to automatically evolve computer programs. We use graph-based data mining to identify common aspects of highly fit individuals and modularizing them by creating functions out of the subprograms identified. Empirical evaluation on the lawnmower problem shows that our approach is successful in reducing the number of generations needed to find target programs. Even though the graph-based data mining system requires additional processing time, the number of individuals required in a generation can also be greatly reduced, resulting in an overall speed-up.

[1]  Lawrence B. Holder,et al.  MDL-Based Context-Free Graph Grammar Induction , 2003, FLAIRS.

[2]  Lawrence B. Holder,et al.  Applying the Subdue Substructure Discovery System to the Chemical Toxicity Domain , 1999, FLAIRS Conference.

[3]  Leonardo Vanneschi,et al.  Theory and practice for efficient genetic programming , 2004 .

[4]  Victor Ciesielski,et al.  Comparison of the Effectiveness of Decimation and Automatically Defined Functions , 2005, KES.

[5]  Lawrence B. Holder,et al.  Graph-Based Data Mining , 2000, IEEE Intell. Syst..

[6]  Himanshu Agrawal,et al.  Vectorization of Structure to Index Graph Databases , 2002, ISDB.

[7]  Una-May O'Reilly,et al.  Genetic Programming II: Automatic Discovery of Reusable Programs. , 1994, Artificial Life.

[8]  D.J. Cook,et al.  Structural mining of molecular biology data , 2001, IEEE Engineering in Medicine and Biology Magazine.

[9]  Vidroha Debroy,et al.  Genetic Programming , 1998, Lecture Notes in Computer Science.

[10]  Ayahiko Niimi,et al.  Object oriented approach to combined learning of decision tree and ADF GP , 1999, IJCNN'99. International Joint Conference on Neural Networks. Proceedings (Cat. No.99CH36339).