CODEWITHSUNDEEP
javamachine-learningneural-networkencog
Ben
Ben

Reputation: 199

Encog Time Series Prediction just stick at the top and not improveing

I use encog for java to do a time series prediction, but it does not seem to work at all. I am pretty new to this and i dont know whats wrong.

enter image description here

The red line is the training data (~3600 data entrys) and the blue line is what the Neural Net predicts...

I use the last 250 data points to predict the next one.

Network Structure:

BasicNetwork net = new BasicNetwork(); net.addLayer(new BasicLayer(null, true, 250)); net.addLayer(new BasicLayer(new ActivationSigmoid(), true, 6)); net.addLayer(new BasicLayer(new ActivationSigmoid(), true, 1)); net.setLogic(new FeedforwardLogic()); net.getStructure().finalizeStructure(); net.reset(); final ManhattanPropagation train = new ManhattanPropagation(net, ndata, 0.5);

Also, it does not realy matter how many iterations i do, after the first like 10 iterations it is sticking to one error value constantly.

Upvotes: 0

Views: 515

Answers (2)

Ironluca
Ironluca

Reputation: 3762

I dont think your approach will work (my humble view only) in terms of time series prediction with a simple feed forward network, reason being time series prediction needs to understand a pattern of the data over time that occurs frequently and then predict the next data point. My guess is the network is not converging because of (probably) moving data. Use LSTM or Hierarchical Temporal Memory kind of networks for this probably you will get a better result hopefully.

Upvotes: 0

Yuriy Zaletskyy
Yuriy Zaletskyy

Reputation: 5151

First of all, your ratio of inputs/hidden nodes/output is very unlikely to work. In his books Heaton gives advice to use following rule: suppose that x stands for input nodes. Then hidden neurons number should be x * 2 / 3. For output neurons layer I ( Heaton doesn't give any advice except to try ) as usually use x / 15.

Second part is training algorithm. Manhattan propagation is not as good as rmsprop + backpropogation ( also according to Jeff Heaton books) .

Upvotes: 0

Related Questions