Publication Type

Book Chapter

Version

publishedVersion

Publication Date

10-2008

Abstract

We propose, theorize and implement the Recursive Pattern-based Hybrid Supervised (RPHS) learning algorithm. The algorithm makes use of the concept of pseudo global optimal solutions to evolve a set of neural networks, each of which can solve correctly a subset of patterns. The pattern-based algorithm uses the topology of training and validation data patterns to find a set of pseudo-optima, each learning a subset of patterns. It is therefore well adapted to the pattern set provided. We begin by showing that finding a set of local optimal solutions is theoretically equivalent, and more efficient, to finding a single global optimum in terms of generalization accuracy and training time. We also highlight that, as each local optimum is found by using a decreasing number of samples, the efficiency of the training algorithm is increased. We then compare our algorithm, both theoretically and empirically, with different recursive and subset based algorithms. On average, the RPHS algorithm shows better generalization accuracy, with improvement of up to 60% when compared to traditional methods. Moreover, certain versions of the RPHS algorithm also exhibit shorter training time when compared to other recent algorithms in the same domain. In order to increase the relevance of this paper to practitioners, we have added pseudo code, remarks, parameter and algorithmic considerations where appropriate.

Keywords

Genetic Algorithm, Gradient Descent, Training Time, Task Decomposition, Generalization Accuracy

Discipline

Artificial Intelligence and Robotics | Databases and Information Systems

Research Areas

Data Science and Engineering

Publication

Engineering Evolutionary Intelligent Systems

Editor

ABRAHAM, Ajith; GROSAN, Crina; PEDRYCZ, Witold

First Page

129

Last Page

126

ISBN

9783540753957

Identifier

10.1007/978-3-540-75396-4_5

Publisher

Springer

City or Country

Germany

Additional URL

http://doi.org/10.1007/978-3-540-75396-4_5

Download

Share

COinS