Week 11 - Training LSTM-4 for modeled samples

1 minute read

To try to improve the latest results obtained with the recurrent network, I have tried two different ways: increase of neurons and increase of layers

Increase the number of neurons

The increase in neurons is done by doubling this value, the network goes from having 25 neurons to 50. This new network is trained and tested with the dataset that produced the worst results in the simplest structure.

  • Sinusoidal dataset, 4 DOF, 80000 training samples, 10000 test samples
More units-Sinusoidal 4 DOF

รง

The results obtained are better than those of the 25-neuron LSTM, but a good predictive capacity is still not obtained wuith this dataset.

Increase the number of layers

For the number of layers I have done a study gradually increasing the number of LSTM layers in the network.

  • Sinusoidal dataset, 4 DOF, 80000 training samples, 10000 test samples
More layers-Sinusoidal 4 DOF

Very good results are obtained that improve until reaching 4 layers, from there the curve rises again. This fact causes the new structure to be defined with 4 layers since it offers the minimum of relative error.

Proposed net

after the results obtained the new defined network is as follows:

LSTM-4 Net

Linear dataset results

  • 2 DOF, 80000 training samples, 10000 test samples
LSTM-4-Linear 2 DOF

Parabolic dataset results

LSTM-4-Parabolic 3 DOF

Sinusoidal dataset results

  • 4 DOF, 80000 training samples, 10000 test samples
LSTM-4-Sinusoidal 4 DOF

Combined dataset results

By getting good results in each type of movement separately we have created a new combined dataset that uses 33% of linear examples, another 33% of parabolic and 34% of sinusoidal.

  • Combined 33%-33%-34%, 80000 training samples, 10000 test samples
LSTM-4-Combined

Conclusions

With the new more complex structure, very good prediction results are achieved, dominating all dynamics in all their DOFs. This fact shows that a more complex structure, in its proper measure, can make the network better predict.