Use specific function decorators for different algorithm categories

[fishroot]

The algorithm decorator currently allows arbitrary attributes, and therefore different usage within models, datasets, networks and systems. Thereby the internal organisation is facilitated by the attribute 'category', which, however is used differently within different contexts.

Code redundancy and code scattering can be avoided by providing different specific decoratos, like: inference for mean values, correlations etc.
estimand for optimization
objective for scalar evaluation functions