|
I want to do some time series learning/predictions using neural networks. I want to use matlab functions. Which are some of the prominent papers that can be looked at for references at the way that this should be done? |
|
I haven't personally used matlab to solve this problem before but the techniques are usually the same across APIs. If I understand your question, you are interested in figuring out how to structure time series data for use within a neural network. In general you will want to take a time series of data and turn it into a training set of data. For instance, the series: {1,2,3,4,5} could become: 1 2 or even {1,2} 3 and so forth. As for papers- there are quite a few of these out there so it is difficult to recommend a paper without more context. For example, what type of time series is it, how much data do you have, how much noise is in the data, etc. @Ryan Kirk , I have quite a bit of data, it is wind speed to wave size data. So I have two time series really. I want to predict the wave only, not the wind. So do you think that I should put half the inputs as wind past values and half the values from the wave heights? So that the output is the wave heights predicted?
(Feb 12 '12 at 17:59)
VassMan
Sure, I think there is reason to believe that both of these factors could be predictive. From your description, it may also be helpful to preprocess your continuous values so that they are easier to classify.
(Feb 12 '12 at 20:42)
Ryan Kirk
|
|
Here is a paper that explains about the time delay neural network. May be this paper can be useful for you. This is the link for ist MATLAB toolbox. @Kuri-kuri are time delay neural networks better than other types for time series? What alternatives exist?
(Feb 12 '12 at 13:01)
VassMan
@VassMan, other alternative can be convolution network.I don't know about which is more better. I think it differs from task to task.
(Feb 12 '12 at 13:28)
Kuri_kuri
@Kuri-kuri, thanks. The convolution network I have been told is better for image problems.
(Feb 12 '12 at 13:58)
VassMan
|
|
Standard Neural Networks are not that good a fit for time series data. Since Neural Networks were first conceived as a classification tool for batches of data rather than for time series data. I do not know about any implementations, but here, you can read about recursive neural networks, which indeed are designed to deal with time series data. (Or dynamic behavior) The thing with dynamic data, is that you cannot assume they are iid as easy as with batch data, so you have dynamic models like Hidden Markov Models where the probability of the next input depends on the latest input. If you are starting, I recommend you to look for toolboxes that implement recursive neural networks. Actually, I think neural networks are really good for time series modelling. As long as you don't have any long term dependencies, taking a neural autoregressive approach (aka neural network with time windows of inputs) work really well.
(Feb 14 '12 at 13:51)
Justin Bayer
I'm not a big fan of NN for time series. I prefer to use more intuitive models, like HMM. I know they work well, I just think there are better models.
(Feb 14 '12 at 18:42)
Leon Palafox ♦
|
To get useful advice, you will need to put a lot more information about your data in your post. Do you have wind speed /wave height data at many different measurement points? What kind of problem are you solving? Do you think relevant information for prediction exists at all previous points in time or only in a short window of time? Are the relationships you want to model highly nonlinear? Have you tried a simple linear autoregressive model yet? Also, papers are typically not about what language the source code the authors used is in.
@gdahl, I have around 4K data points for each series. They are not perfectly aligned though; is that a problem? I am trying to predict wave heights. They are not supposed to be very non-linear. Another guy did auroregressive techniques which worked alright. I am using matlab's ANN toolbox which seems powerful but complicated commands. Would the netlab which can work in matlab be easier?