Learning Chaotic Fluctuation of Japanese Vowel using Simple Recurrent Network with Short-term Memory

摘要

The concept of short-term prediction of deterministic chaos makes an impact on the engineering science for real-world time series. Elman's recurrent neural network is such a nonlinear prediction method. Recently, we have proposed a structure which directly uses past histories in Elman's network. It has been shown that learning is accelerated using the structure. However, data in the previous works are periodic, so past histories are important for learning time series. In this work, it is shown that the structure is also useful for chaotic data. Moreover, we employ the cross entropy method for fast learning. In this paper, fluctuation of Japanese vowel is treated as an example of chaotic time series. Using our method, it is shown that learning speed attains about 100 times faster without making generalization ability worse compared with Elman's method. Moreover, generalization ability of multi-step forecasting is improved by introducing the buffer.