Regression output for LSTMs?

Hello,

I am trying to use LSTMlayers to properly predict my current 6 Degree of freedom orientation/position, using my Last Position and acceleration/angular momentum as inputs. My first important question: Is there any way to use LSTM layers with a regression output as of 2017b? Second:If so or if not, are there general implementation ideas for my problem?

thanks in advance.

[Shounak Mitra]

Hi David,

Regression Layers are not supported with 17b release. There are two options that I can think of:

1. Add your own custom regression layer: I understand that it's not convenient but it's not super difficult as well. See the heading 'Define Deep Layers': at https://www.mathworks.com/solutions/deep-learning/examples.html

2. Have you tried using Kalman Filters?

[David Kuske]

Hello Shounak,

Thanks for the answer, 1) So it is possible to use the provided layers, including LSTM if i just write a custom regression output layer?

2) Its a possibility I might use, but not my first choice, calibration would always need to be perfect.

[William Smith]

Creating my own custom regression layer resulted in the same error: "Error using trainNetwork (line 140) Regression is not supported for networks with LSTM layers."