I have been trying to figure out how to to generate a time series dataset for feedforward neural network.
Let's say I want to simulate a time series dataset with lenght n and order m using feedforward neural network by following steps:
1- firstly I randomly initial m values for weights, biases and data using uniform distribution with the specific parametes as mean and variance
2- secondly I calculate the n elements of time series data using feedforward neural network function
3- thirdly I randomly generate white noise using normal distribution with the specific parametes as mean and variance
4- then I adding generated white noise to my n elements of time series data which calculated form second step
I have found resource related to adding white noise to series and I see at the first we should generate n time series data completly and then we should adding white noise to each elements of generated time series. So when we want generate first element (m+1 element) we should not adding white noise to it. Therfor the next data elements as m+2, m+3, m+4,... generate using feedforward neural network function whithout the effects of white noise of m+1, m+2, m+3 ,...
As I see the data generation using feedforward neural network by above steps led to same constant value when m is smal (equel to 1 or 2,3) and led to zero when m is large (m=6,7,..14) and my data series whithout white noise is same to a const line for smal value of m, and is a zero for large value of m! And finally by adding white noise to it my simulated series is same to generated white noise.
What I am trying to find out for simulate a time series dataset for feedforward neural network?
The empirical evaluation that increasingly noisy data series increasingly degrade the ability of the neural network to learn the underlying form of the data series and produce good forecasts. Feed-forward neural networks are extremely sensitive to architecture and parameter choices, and making such choices is currently more art than science, more trial-and-error than absolute, more practice than theory.
The field of neural networks is very diverse and opportunities for future research exist in many aspects, including data preprocessing and representation, architecture selection, and application. The future of neural network time series forecasting seems to be in more complex network types that merge other technologies with neural networks, such as wavelet networks. Nevertheless, a theoretic foundation on which to build is absolutely necessary, as more complex networks still suffer from basic problems such as data preprocessing, architecture selection, and parameterization.
Hi
did you train the FFANN? you used initial random values only. How did you train it? what is the correlation in your series?
Dear zakaria Masoud
Sorry for the late answer.
No I did not train the FFANN. I find the answer of the question. I did just a small mistake. With the specific order, when I generate each element (for example the mth), I should not include that into next generated data (m+1th ). The determined order makes the correlation in the series.
- Badges
- Science topic