TY - JOUR
T1 - Symbiotic Structure Learning Algorithm for Feedforward Neural-Network-Aided Grey Model and Prediction Applications
AU - Yang, Shih Hung
AU - Huang, Wun Jhu
AU - Tsai, Jian Feng
AU - Chen, Yon Ping
N1 - Funding Information:
This work was supported by the Ministry of Science and Technology of the Republic of China, Taiwan, under Contract MOST 105-2221-E-035-014-, Contract MOST 105-2218-E-150-003-, and Contract NSC 102-2511-S-009-011-MY3.
Publisher Copyright:
© 2013 IEEE.
PY - 2017
Y1 - 2017
N2 - The learning ability of neural networks (NNs) enables them to solve time series prediction problems. Off-line training can be applied to design the structure and weights of NNs when sufficient training data are available. However, this may be inadequate for applications that operate in real time, possess limited memory size, or require online adaptation. Furthermore, the structural design of NNs (i.e., the number of hidden neurons and connected topology) is crucial. This paper presents a novel algorithm, called the symbiotic structure learning algorithm (SSLA), to enhance a feedforward neural-network-aided grey model (FNAGM) for real-time prediction problems. Through symbiotic evolution, the SSLA evolves neurons that cooperate well with each other, and constructs NNs from the neurons with hyperbolic tangent and linear activation functions. During construction, the hidden neurons with the linear activation function can be simplified to a few direct connections from the inputs to the output neuron, leading to a compact network topology. The NNs share the fitness value with participating neurons, which are further evolved through neuron crossover and mutation. The proposed SSLA was evaluated through three real-time prediction problems. Experimental results showed that the SSLA-derived FNAGM possesses a partially connected NN with few hidden neurons and a compact topology. The evolved FNAGM outperforms other methods in prediction accuracy and continuously adapts the NN to the dynamic changes of the time series for real-time applications.
AB - The learning ability of neural networks (NNs) enables them to solve time series prediction problems. Off-line training can be applied to design the structure and weights of NNs when sufficient training data are available. However, this may be inadequate for applications that operate in real time, possess limited memory size, or require online adaptation. Furthermore, the structural design of NNs (i.e., the number of hidden neurons and connected topology) is crucial. This paper presents a novel algorithm, called the symbiotic structure learning algorithm (SSLA), to enhance a feedforward neural-network-aided grey model (FNAGM) for real-time prediction problems. Through symbiotic evolution, the SSLA evolves neurons that cooperate well with each other, and constructs NNs from the neurons with hyperbolic tangent and linear activation functions. During construction, the hidden neurons with the linear activation function can be simplified to a few direct connections from the inputs to the output neuron, leading to a compact network topology. The NNs share the fitness value with participating neurons, which are further evolved through neuron crossover and mutation. The proposed SSLA was evaluated through three real-time prediction problems. Experimental results showed that the SSLA-derived FNAGM possesses a partially connected NN with few hidden neurons and a compact topology. The evolved FNAGM outperforms other methods in prediction accuracy and continuously adapts the NN to the dynamic changes of the time series for real-time applications.
UR - http://www.scopus.com/inward/record.url?scp=85028944220&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85028944220&partnerID=8YFLogxK
U2 - 10.1109/ACCESS.2017.2702340
DO - 10.1109/ACCESS.2017.2702340
M3 - Article
AN - SCOPUS:85028944220
SN - 2169-3536
VL - 5
SP - 9378
EP - 9388
JO - IEEE Access
JF - IEEE Access
M1 - 7921699
ER -