Sensitivity analysis for Dynamic solution
Sensitivity analysis for Dynamic solution#
This functionality is based on the backpropagation computations, and its general idea is:
At a set of timepoints of a chosen output trajectory compute the gradients with respect to all input trajectories via the backpropagation
For each input calculate the norm of its obtained gradients
Normalize the results, so that the sum of sensitivities to all inputs is equal to one
To perform the Sensitivity analysis:
Switch to Metrics panel
Choose the file in Evaluation files section:
If the file was supplied earlier, it is already listed in Evaluation files
To add new file for evaluation, press + in Additional section of Evaluation files
Choose the type of initialization (see Evaluate NeurEco Discrete Dynamic model with GUI and Metrics for the Discrete Dynamic model with GUI)
Click on one of the output neurons (representing output features) on the plot in the Network sensitivity section
Each input neuron (representing input features) becomes colored according to the sensitivity of the chosen output neurons with respect to this neuron and its sensitivity coefficient is showed in the plot
An example of the sensitivity analysis:
Note
By default, the Network sensitivity is performed with the last model available in the checkpoint. Use the checkpoint slider in the bottom to choose any other available model.