florianjehn / acme Goto Github PK

When creating the parents it should be somehow guaranteed that more complex models are created roughly equal as simple models. This might be accomplished by making the chance of the addition of more layers and connections higher, in regard how many storages and connections the model already has. This helps compensate for the additional tests that are used.

It is confusing if the input unit can be different. Therefore ACME should only accept certain units in the forcing data.

DREAM uses this weird objective function, can Spotpy still be used to calculate other objective functions in the progress? For example Klingt Gupta, so in the End one can better compare stuff with ACME?

Make sure all files have the same header (including description, license and stuff)

With networkx or https://pygraphviz.github.io/documentation/pygraphviz-1.4rc1/tutorial.html#layout-and-drawing da graph can be drawn. This could be automated by reading out the structural information from the genotype. This can then be saved as a png or something similar.

The construction of the genotypes is not very self explanatory and therefore needs a lot of documentation. To not overfill the source code this should be outsourced to markdown files (one for each model generator genotype).

With pprint you can create Python interpretable strings, which in turn can be saved to the hard drive as *.py files. This enables to import them easily again if needed.
For importing those files see:
https://stackoverflow.com/questions/17123719/importing-lists-from-multiple-files-using-a-for-loop-in-python
For general explanation see "Automate the boring stuff with python" page 185

For testing purposes it might be useful to have a predefined test parents, which can be used as a benchmark.

For more info see "Genetic algorithms with pythoN" p.94

Tobias thinks it might be possible to stop spotpy from giving you an output by simply not giving it a dbname and dbformat.
This would be quite handy so nothing is written to the harddrive

Fitness evaluation is very time consuming try to find an approximate way to calculate fitness.

It might be useful to see how active the genome is, meaning how many of its genes are expressed.

Connection building might be enhanced by using try, instead of if/else to control which connections are built.

ACME should, at leat in theory, be compatible with other modeling frameworks. It would be nice if someone could try to implement with other frameworks lice FUSE.

When starting a model with the full genome present, it does not calculate correctly and hangs somewhen in the first year. Strangely, a model build by hand with the same layout (its the same when looked at with cmf.describe), does not show this behaviour. Whats the problem?

The use of the template is way easier, if only those genes are forwarded to it, which really are expressed. For example a beta parameter can only be expressed if the matching connection exists. Or a connection can only exist when the storages exists which it is meant to connect.

This could probably done even before translating the model to code and thus save computational time
For this one would have to write a function that checks the status of the variables for the existens of the storages and the fluxes.
E.g.
If first_storage not present:
return ModelNotClosedError
if first storage has no connections
return ModelNotClosedError

and so forth

Create a first file in the layout that it should have, hash it and compare the hash to the hash of the output of the translation function.

Build keeps failing because Travis cannot find the right files.
How do I tell Travis where the files are to be found?
https://travis-ci.org/zutn/ACME/jobs/286938074

Make a dict or list of all already tested solution. Check new indiviuals after mutation. If they are already in the list, discard them.