I am working on an implementation of Pattern Search for an auto-tuning-framework based on the pattern search algorithm of Hooke and Jeeves. By looking at the results I get with my implementation and comparing them with the ones from OpenTuner Pattern Search, I recognized that the results from OpenTuner are not as good as the ones one I got with the added pattern step, especially in big search spaces. I guess that's because the algorithm is trapped in local minima, not able to get out again to continue exploring, but I don't really know.
So I was wondering, why did you decide to go for this specific version of Pattern Search? Is there a specific reason?
The included pattern search is primarily a baseline and in most of our testing is worse than the other available techniques. Pull requests to improve it are welcome.
Hello all,
I am working on an implementation of Pattern Search for an auto-tuning-framework based on the pattern search algorithm of Hooke and Jeeves. By looking at the results I get with my implementation and comparing them with the ones from OpenTuner Pattern Search, I recognized that the results from OpenTuner are not as good as the ones one I got with the added pattern step, especially in big search spaces. I guess that's because the algorithm is trapped in local minima, not able to get out again to continue exploring, but I don't really know.
So I was wondering, why did you decide to go for this specific version of Pattern Search? Is there a specific reason?
Greets!