http://zorro-trader.com/manual/en/training.htm says
The following cases are possible:
Rules depend on a parameter. ... Set the RULES and PARAMETERS flags at the same time. The training run will now repeat through many cycles. At first the parameter will be optimized and for every parameter step a new set of rules will be generated. ...
I find that when combining PARAMETERS with RULES with an adviseLong(NEURAL, ...), neural(mode=NEURAL_SAVE, ...) is not called as expected at the end of each optimize step. But, neural(model=NEURAL_LOAD,.. ) is called. Therefore Zorro dies when it tries to load an non-existing ml file since it wasn't produced, or it happily re-uses one from some old experiment.
