Prior for mechanistic parameters via penalty in loss function

[BernhardAhrens]

plus L2 and L1 regularization parameters, mechanistic and NN.

[mreichMPI-BGC]

And in the same vein, we should allow to put the latent variables in the cost function. Either to penalize some deviation from prior summary stats (e.g. mean), or to allow for soft constraints on some properties of the latents so that they obey some principles (or plausibilities).

It would be nice if the cost function could be defined in a similar generic way as the Hybrid model itself...

[lazarusA]

how do you envision the syntax of passing these to train ?

[BernhardAhrens]

My idea would be to pass all latents through the mechanistic model into the model output. That way would could treat all latents in a generic way - regardless if there are global parameter (latent parameter), a neural parameter (latent variable). We could pass these priors by extending the parameter tuple or passing something similar which gives for example a shape, scale and distribution.

The more bespoke losses like that the latents have to obey certain principles, could be just done through a cost function. The pass-through the mechanistic function maybe unnecessary. We have a step before the mechanistic model where all arguments for the mechanistic function are collected. We could save those. However, later when we want to include an NN into the mechanistic model when it receives state variables as inputs, we have to do the pass-through

[BernhardAhrens]

Closed with 14ab9c7