Improve the cost function learner towards full automation

[sekruse]

There are few impediments that inhibit a fully automated use of the GeneticOptimizer:

• [x] the cost function templates must specify "type":"mathex" for backwards compatibility reasons
• [x] the final cost functions should be writable to a file rather than just on the standard output, mangled with other information
• [x] the key for platform initializations are not set properly in the cost function outputs
• [x] a time limit would be a useful convergence criterion