Can these ES algorithm work for Resnet?

[ouyangzhuzhu]

Hi:
I tried some cases, and have a question, hope you can help solve some~~
I tried OpenAI-ES for Lnet5 on MNIST whose weights space is 44426 and it took 31 mins, 20 epochs to get 95% accuracy, and in the end the highest accuracy is 97%
I tried OpenAI-ES for Lnet5 on CIFAR-10 whose weights space is 62006 and it took 214 mins, 66 epochs to get 50% accuracy, and in the end the highest accuracy is 52%
I tried OpenAI-ES for Resnet18 on MNIST whose weights space is 11172810 and it took quite a long time but the accuracy kept around 11.35% that is no improvement....
I tried OpenAI-ES for Resnet18 on CIFAR-10 whose weights space is 11173962 and it took quite a long time but the accuracy kept around 10.00% that is no improvement....
I guess there are some reasons that the OpenAI-ES did not work for Resnet18:

1. the weights space is huge!! and my popsize is too small, I tried set the popsize to 100,200,500, the result is still pool like above
2. the weights space is huge!! and the learning-rate and mute_rate(that is the sigma) should be carefully set based upon a lot of experiences and I tried my best to adjust them, still not work......
   Can u help me，hope can get some suggestions from u !!
   :) : ) : )
   Waiting on line
   :( : ( : (