Authors:
Thibaut Bellanger
1
;
2
;
Matthieu Berre
1
;
Manuel Clergue
1
and
Jin-Kao Hao
2
Affiliations:
1
LDR, ESIEA, 38 rue des Docteurs Calmette et Guérin, 53000 Laval, France
;
2
LERIA, Université d’Angers, 2 Boulevard Lavoisier, 49045 Angers, France
Keyword(s):
Genetic Programming, Classification, Tangled Program Graph, Ensemble Learning, Evolutionary Machine Learning.
Abstract:
We propose an approach to improve the classification performance of the Tangled Programs Graph (TPG). TPG is a genetic programming method that aims to discover Directed Acyclic Graphs (DAGs) through an evolutionary process, where the edges carry programs that allow nodes to create a route from the root to a leaf, and the leaves represent actions or labels in classification. Despite notable successes in reinforcement learning tasks, TPG’s performance in classification appears to be limited in its basic version, as evidenced by the scores obtained on the MNIST dataset. However, the advantage of TPG compared to neural networks is to obtain, like decision trees, a global decision that is decomposable into simple atomic decisions and thus more easily explainable. Compared to decision trees, TPG has the advantage that atomic decisions benefit from the expressiveness of a pseudo register-based programming language, and the graph evolutionary construction prevents the emergence of overfittin
g. Our approach consists of decomposing the multi-class problem into a set of one-vs-one binary problems, training a set of TPG for each of them, and then combining the results of the TPGs to obtain a global decision, after selecting the best ones by a genetic algorithm. We test our approach on several benchmark datasets, and the results obtained are promising and tend to validate the proposed method.
(More)