Application of Back-propagation Neural Network to Categorization of Physical Fitness Levels of Taiwanese Females

Abstract

The purpose of this study is to establish the feasibility of adopting the back-propagation neural network (BPNN) to predict fitness category. In this study, 2218 healthy Taiwanese females aged 20 to 65 participated. Data collected included five parameters required for the physical fitness (PF) passport: subject's age, body mass index (BMI), performance in the sit-and-reach test, 1-min bent-leg curl-ups, and cardiorespiratory endurance. The network structure of BPNN adopted here consisted of three layers: input layer (5 neurons), hidden layer (5 neurons), and output layer (4 neurons). To prove the ability of BPNN in categorizing PF accurately and speedily, its learning effect must be confirmed. To achieve this purpose, the samples were divided randomly into two parts: training samples (n=1218) and testing samples (n=1000). Thereafter, learning algorithms of the BPNN were executed. The learning rate was assumed to be 0.75, and 1000 learning cycles were run. The results demonstrated that the root mean square (RMS) for the training samples was 0.059, while the RMS for the testing samples was 0.065. Such small RMS is evidence that the BPNN converged well and had a good learning effect. On the other hand, the mean degree of accuracy of the BPNN was 96.83% in identifying body composition, 98.41% for muscular flexibility, 94.39% for muscular strength and endurance, and 97.25% for cardiorespiratory capacity. The mean degree of accuracy for these four items was as high as 96.72%. Overall, the mean relative error for categorizing PF was 3.28%, which was within the acceptable range. Therefore, the results confirmed the reliability of the BPNN for categorizing PF. BPNN can be converted into software to assess the subject's PF in a precise and speedy manner, thus eliminating the need to refer to a chart of the modular standard to decide the fitness category.