A regeneratable dynamic differential evolution algorithm for neural networks with integer weights

Neural networks with integer weights are more suited for embedded systems and hardware implementations than those with real weights. However, many learning algorithms, which have been proposed for training neural networks with float weights, are inefficient and difficult to train for neural networks...

Full description

Saved in:
Bibliographic Details
Published in:Frontiers of information technology & electronic engineering 2010-12, Vol.11 (12), p.939-947
Main Authors: Bao, Jian, Chen, Yu, Yu, Jin-shou
Format: Article
Language:English
Subjects:
Algorithms
Communications Engineering
Computer Hardware
Computer Science
Computer Systems Organization and Communication Networks
Dynamics
Electrical Engineering
Electronics and Microelectronics
Embedded systems
Evolution
Evolutionary algorithms
Evolutionary computation
Instrumentation
Integers
Machine learning
Networks
Neural networks
Strategy
Training
Citations: Items that this one cites
Items that cite this one
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Neural networks with integer weights are more suited for embedded systems and hardware implementations than those with real weights. However, many learning algorithms, which have been proposed for training neural networks with float weights, are inefficient and difficult to train for neural networks with integer weights. In this paper, a novel regeneratable dynamic differential evolution algorithm （RDDE） is presented. This algorithm is efficient for training networks with integer weights. In comparison with the conventional differential evolution algorithm （DE）, RDDE has introduced three new strategies： （1） A regeneratable strategy is introduced to ensure further evolution, when all the individuals are the same after several iterations such that they cannot evolve further. In other words, there is an escape from the local minima. （2） A dynamic strategy is designed to speed up convergence and simplify the algorithm by updating its population dynamically. （3） A local greedy strategy is introduced to improve local searching ability when the population approaches the global optimal solution. In comparison with other gradient based algorithms, RDDE does not need the gradient information, which has been the main obstacle for training networks with integer weights. The experiment results show that RDDE can train integer-weight networks more efficiently.
ISSN:1869-1951
2095-9184
1869-196X
2095-9230
DOI:10.1631/jzus.C1000137