Hello,
I was testing a bit HEBOSearch - I want to compare with other searchers.
Do you think it is possible to implement batch suggest by having n_suggestions > 1?
Please see the speed difference bellow for finding the minima for pymoo.factory.get_problem(“rastrigin”, n_var=2)
Test nb (non-ray) is here
https://drive.google.com/file/d/10w-plbY4pdbLPgwiq-upStH2x_bX2adb/view?usp=sharing
Regards,
Viorel
Hey, that’s a great suggestion. I have confirmed it with HEBO authors that setting n_suggestions >1 can improve convergence speed, and that they used a value of 8 for the competition. I have created a PR implementing support for this feature: [Tune] Batch suggestions for HEBO by Yard1 · Pull Request #14246 · ray-project/ray · GitHub
Let me know if this works for you!
ok, thanks for letting me know.
One more thing: I did some testing (for rastrigin problem) of almost all functions in suggest and here is the result:
One iteration for the problem takes short time and this is why some algs itself take longer than one problem iteration