| Summary: | 碩士 === 長庚大學 === 資訊管理研究所 === 90 === Interest in algorithms which dynamically construct artificial neural networks has been growing in recent years. The traditional methods that have been used to construct near optimal artificial neural networks are to minimize the sequence of error functions associated with the growing network. But, constructing a near optimal artificial neural networks by considering the sequence of error unilaterally, will lead to more complicated network architecture.
This paper proposes an evolutionary system for constructing neural networks named “MeNet”. Combining the concepts of “Evolutionary Programming” and ”Multi-Objective Programming”, MeNet minimizes the sequence of error and the complexity of network architecture simultaneously.
MeNet has been tested on two benchmark problems, including Xor problem and 8-bit parity checking problem. The results show that MeNet can produce neural network architecture with lower complexity while keeping the acceptable accuracy. Such effects are clear, especially in testing more complicated problem.
|
|---|
