Just for fun I am seeing if I can train a model to predict the next step in a smoke sim. I figured if that would work, I could put it in a solver sop, and have a ml smoke solver. It works surprisingly well for predicting the first next step, but when I feed the result back to the MlRegressionInference the second time, nothing happens.
Maybe someone out there can tell me the fix? I have attached my file.
That you're not seeing a difference for the second inference is probably due the model being overfitted. Even though the width of the hidden layers is set to be small, the model is still very big compared to the number of examples, due to the fully connected input and output layers. If you apply the inference once, you generate an unseen input for the next input. That unseen input is different from the seen inputs. A severely overfitted model may not do much for such unseen inputs.
On my end, I was able to get a visible difference between the first and second application of inference, going from from 200 to 400 examples. (But with the MLP mode architecture, you may need many more examples than that. Or you may want to switch to using PCA components for input/output)
