Application of Genetic Algorithm in determining optimal backpropagation neural network architecture for ground-level ozone forecasting model
Backpropagation neural networks have shown better prediction accuracy than linear regression in summertime ground-level ozone forecasting. It is highly desirable to have a distinct, optimal neural network model presenting the distinct properties of each region. Before normalizing values of inputs and output to be used in the neural networks, they may be transformed, such as taking logarithm or their reciprocals, to provide distinct perspectives of studied regions. In this paper, the authors propose a system consisting of a backpropagation neural network and a Genetic Algorithm unit to examine efficiently the effects of form transformation and different number of hidden nodes in ozone concentration forecasting model. The Genetic Algorithm unit first randomly chooses number of hidden nodes, and individual transformation form for each input variable, and output variable. These chosen parameters compose one specific backpropagation neural network architecture. These parameters and their prediction accuracy are recorded in the Genetic Algorithm unit. After all other neural network architectures in the same generation submit their parameters and prediction accuracy, the Genetic Algorithm unit then selects the best architectures to reproduce, to crossover, and to mutate into next generation. The process of computation of fitness values, reproduction, crossover, and mutation continues until at least one of the stopping criteria are met. The authors conduct a case study consisting of eight cities in the US. Their results show that combinations of four transformation forms original, logarithm, reciprocal, and squares, have moderately improved prediction accuracy over a backpropagation neural network using original form in 6 of the 8 cities, given the same number of hidden nodes. The number of hidden nodes, ranging from 2 to 16, has minor effects on the improvement of the prediction accuracy.
- Research Organization:
- Univ. of Alabama, Tuscaloosa, AL (US)
- OSTI ID:
- 20002119
- Report Number(s):
- CONF-990608-; TRN: IM200002%%119
- Resource Relation:
- Conference: Air and Waste 92nd Annual Meeting and Exhibition, St. Louis, MO (US), 06/20/1999--06/24/1999; Other Information: 1 CD-ROM. Operating Systems: Windows 3.1, '95, '98 and NT; Macintosh; and UNIX; PBD: 1999; Related Information: In: Air and Waste 92nd annual meeting and exhibition proceedings, [9500] pages.
- Country of Publication:
- United States
- Language:
- English
Similar Records
A dynamic node architecture scheme for backpropagation neural networks
A dynamic node architecture scheme for backpropagation neural networks