search

BlueBrain / BluePyOpt

Blue Brain Python Optimisation Library
https://bluepyopt.readthedocs.io/en/latest/
Other
196 stars 95 forks source link

parent population #495

Open FrancescoCavarretta opened 3 weeks ago

FrancescoCavarretta commented 3 weeks ago

I am using BluePyOpt to optimize a model implemented in NEURON.

In the optimization, I choose a parent population generated a-priori. fitness Fitness

The average fitness of the individuals worsens over the generations rather than improving. Why does it happen? What is wrong?

AurelienJaquier commented 2 weeks ago

Hi @FrancescoCavarretta ! What is the optimiser that you use in BluePyOpt? Is it IBEADEAPOptimisation or DEAPOptimisationCMA? Also could you share a checkpoint with us so that we can inspect what is happening?

FrancescoCavarretta commented 2 weeks ago

Hi Aurelien: Yes, I can share the checkpoint. I could upload the file into my onedrive and share the link with you via e-mail, if you provide you it. If it does not work for you, let me know your preference.

AurelienJaquier commented 2 weeks ago

Sure! You can send the link to me at aurelien.jaquier@epfl.ch

AurelienJaquier commented 2 weeks ago

Thank you @FrancescoCavarretta . First of all, I have checked the evolution of the best individual over the optimisation, and we can see that it tends to go towards lower scores, so the optimisation is working. As for why the average score is going up, I think it proabably has to do with the initial population you gave to BluePyOpt. As I don't have your scoring function, I could not test my hypothesis, but I think that the following is happening: I think your initial population is located in a region wit a good average score. When BluePyOpt tries to optimise the population, it will search the parameter space, thus going partially outside of this region, increasing the number of individuals with higher scores, thus increasing the mean score. And then after some time, the mean score starts going down. Even if the mean score is not going back to the value it had at the start very fast, the best individual has improved.

best_ind_495