Version 0.3¶

Accuracy improvements of the optimization of the acquisition function by pre-selecting good candidates as starting points when using acq_optimizer='lbfgs'.

Support a ask-and-tell interface. Check out the Optimizer class if you need fine grained control over the iterations.

Parallelize L-BFGS minimization runs over the acquisition function.

Implement weighted hamming distance kernel for problems with only categorical dimensions.

New acquisition function gp_hedge that probabilistically chooses one of EI, PI or LCB at every iteration depending upon the cumulative gain.

Warnings are now raised if a point is chosen as the candidate optimum multiple times.

Infinite gradients that were raised in the kernel gradient computation are now fixed.

Integer dimensions are now normalized to [0, 1] internally in gp_minimize.

The default acq_optimizer function has changed from "auto" to "lbfgs" in gp_minimize.