is it possible to train a Dynamic/Recurrent neural network with Ensemble Method?

i want implementation y(k)=a1 y(k-1)+ b1*u, and u is initiate value.

how learn a1 and b1. is it possible ?

[Greg Heath]

Do you mean you have y(k) for k = 0,1,...

and u(k) for k = 0 with u(k) = 0 for k > 0?

[Greg Heath]

What do you mean by "ensemble" method?

G.

[mohad]

when k=0 we have y(0)= b1* u my professor said me how learn a1 and b1 with ensmeble ?

i think we can train system with Dynomic Network, and i dont know how use Ensemble

[Greg Heath]

You didn't explain the term "ensemble". Does it mean the weighted average output of a number of nets that are trained from different random initial weights?

Answer 1

[Greg Heath]

The documentation in

help fitensemble

doc fitensemble

lists the methods of regression and classification for which this function is applicable. Neural Networks are not mentioned in the list.

However, this does not mean that you cannot form an ensemble by combining the outputs from a number of NNs.

In my experience, simple linear combinations of multiple designs have been very successful. The success results from the "strong learner" capability of NNs.

The differences between designs tend (but are not limited) to be

1. Design (Train + Validation) data subsets

2. Number of hidden nodes

3. Type of hidden node activations (e.g., sigmoidal and Gaussian)

4. Random initial weights

In addition, the function stagewisefit can be used to identify nets whose removable will not significantly decrease performance.

Hope this helps.

Greg

Answer 2

[mohad]

i want to just know is it possible work with neural network and Ensemble ? does Ensmeble such az MEthod of Neural ? and how work with neural and Ensemble?

example of matlab we have::

load carsmall

X = [Horsepower Weight];

ens = fitensemble(X,MPG,'LSBoost',100,'Tree');

mileage = ens.predict([150 2750])

mileage = 22.6735

[Greg Heath]

You still haven't explained what you mean by "Ensemble"

Answer 3

[mohad]

many tnx . i think ensemble is not Leaner just decide. at first i train system with Recurent or dynomic network such as narxnet then out put pass to Ensmeble and Ensmeble Vote best answer. is it true ?

i Read apeper "A new boosting algorithm for improved time-series forecasting with recurrent neural networks" that Work such this. tnxx

[Greg Heath]

Not true.

How does the ensemble know which answer is best?