Claes Fransson comments

Results 52 comments of


                                            Claes Fransson

The landscapes flatten during the iterations after resume with 2 games per iteration.

Okay, thanks, yes it seems like it does the "_apply_noise_vector" on line 602 (bayesgpr.py) when it calls "fit" in the resume process...

The landscapes flatten during the iterations after resume with 2 games per iteration.

How come Chess Tuning Tools follow the Bask optimizer behavior to not call `fit`, but only `sample` in the main optimization loop of the tuner (cli.py, lines 472-479)? You seem...

The landscapes flatten during the iterations after resume with 2 games per iteration.

Okay, I also made an experiment of letting `bask/optimizer.py` always call `fit` by changing line 344 to `self.gp.fit(`. The flattening behavior was still there, confirming what you say about a...

The landscapes flatten during the iterations after resume with 2 games per iteration.

What affects the y-scales and the y-values of the partial dependence plots seems to be this line https://github.com/kiudee/chess-tuning-tools/blob/01ec0b8967caa131ed9c04315feeca35c13b268f/tune/plots.py#L134 When I step through the code with Pdb, that line calls `predict`...

The landscapes flatten during the iterations after resume with 2 games per iteration.

That was a lot of theory to try to grasp! :) > One possible source of errors could be that the hyperparameters of the warping functions and the kernel are...

The landscapes flatten during the iterations after resume with 2 games per iteration.

Here are some logged data from an experiment with the public version of Fat Fritz 2: normalize_y=True, input-warping enabled, gp-initial-burnin (on resume)=gp-burnin=100, Acquisition function: ts, 2 games/iteration Iteration 174, after...

The landscapes flatten during the iterations after resume with 2 games per iteration.

I try to work around this issue by resetting the optimizer each iteration (see my [fork](https://github.com/Claes1981/chess-tuning-tools/blob/master/tune/cli.py)). Now the y-scales of the 6 parameters are still very short after only 228...

The landscapes flatten during the iterations after resume with 2 games per iteration.

I just noticed something that appears a bit strange to me: When I set `normalize_y` to `true` (default), `opt.gp.alpha` actually contains quite large values, much larger than the values of...

The landscapes flatten during the iterations after resume with 2 games per iteration.

I see, thanks for explaining. It makes me wonder, why the `alpha` parameter is not scaled when `normalize_y=true` in the upstream Scikit-learn `GaussianProcessRegressor`? Besides, how come we multiply by the...

The landscapes flatten during the iterations after resume with 2 games per iteration.

> It makes me wonder, why the `alpha` parameter is not scaled when `normalize_y=true` in the upstream Scikit-learn `GaussianProcessRegressor`? Actually `alpha` is [here](https://github.com/scikit-learn/scikit-learn/blob/36915ae390fab4742f98c82dc6802f072c4effa5/sklearn/gaussian_process/_gpr.py#L268) and [here](https://github.com/kiudee/bayes-skopt/blob/7f3e7af1d6b8f3b4f850224909d292b0d8deb387/bask/bayesgpr.py#L205) only applied (added) on `K[np.diag_indices_from(K)]`,...