Funsearch Parameters

[jack-powers]

Hi there,

I am trying to replicate my own Funsearch experiments on a much smaller scale compared to that presented in the original paper, and am trying to tune the hyperparameters.

As detailed in the supplementary information document, a single funsearch experiment took approximately 48 hours generated 2 million prompts, and used the initial temperature and temperature period for cluster probabilities to be 0.1 and 30,000 respectively. For smaller experiments, could the temperature period be scaled proportionally?

Thank you

[timneumann1]

I think the true power of FunSearch lies in the immense number of iterations it runs through. It seems reasonable to rescale some parameters if the number of iterations is smaller, but I think we shouldn't expect great success from smaller experiments..