search

colgreen / sharpneat

SharpNEAT - Evolution of Neural Networks. A C# .NET Framework.
https://sharpneat.sourceforge.io/
Other
379 stars 97 forks source link

Saved genome provides consistent result that don't match the ones it got during training #83

Open voiceofthedead opened 3 weeks ago

voiceofthedead commented 3 weeks ago

The problem is mostly covered in the title. I've trained a genome on a specific set of data, but when I connect the library into a different project and try to backtest it on the same set of data, the results are consistent from run to run, but don't match the one that occurred during the learning process. Any suggestions on how could that come about ?

colgreen commented 3 weeks ago

Is it a cyclic/recurrent network, or an acyclic net?

voiceofthedead commented 3 weeks ago

It's set as acyclic in the config file

colgreen commented 3 weeks ago

I wonder - does your fitness function have state that needs resetting?

voiceofthedead commented 3 weeks ago

It did at some point, but I figured out that it caused an issue, so now everything that's going to be changed in the process of calculating fitness is populated at the beginning of RunTrial and stored locally in the method, so not anymore

voiceofthedead commented 2 weeks ago

Any other ideas perchance ?

voiceofthedead commented 1 week ago

Switching to cyclic didn't influence the issue btw