search

AIworx-Labs / chocolate

A fully decentralized hyperparameter optimization framework
http://chocolate.readthedocs.io
BSD 3-Clause "New" or "Revised" License
120 stars 41 forks source link

Bayes parallel samples not different. #13

Open knuds opened 7 years ago

knuds commented 7 years ago

I run bayes with kappa=1.2

The problem is that I get the same suggestion 2 times. Experiment 6 and 8 is the same. I ask for 8 before 6 was run. And 11 and 12 is also the same.

id NClusters NParents R_cut _loss
0 6 4 9.7 -0.50 1 12 2 6.1 -0.04 2 3 1 6.2 -0.01 3 6 1 7.4 -0.28 4 9 1 11.9 -0.11 5 5 2 11.9 -0.59 6 3 3 11.9 -0.01 7 5 3 11.9 -0.65 8 3 3 11.9 -0.01 9 12 4 2.0 0.00 10 12 4 2.8 0.00 11 5 4 11.9 NaN 12 5 4 11.9 NaN

knuds commented 7 years ago

I think that a solution could be that instead of using the GP estimate for pending items just use the maximum. So if the GP estimate for item 11 is around 0.7 then it will keep suggestion it. but if you set it to the maximum 0.0 then the next suggestion will be something completely different.