ChangYong-Oh

Hi Wenlung. https://www.google.com/url?sa=t&source=web&rct=j&url=http://www.or.uni-bonn.de/research/montreal.pdf&ved=2ahUKEwi7stuW_ebmAhXHfFAKHd-KAWUQFjAKegQIARAB&usg=AOvVaw2SzZaJ9dQJYcpvdzapH7VQ This may be relevant. Since many chip design problems can be formulated as combinatorial optimization problems, sample efficient combinatorial optimization method like COMBO may be useful for...

Hi. You can ignore that partition. It was intended for some extension of COMBO I tried. I haven't cleaned that up . On Sat, Mar 26, 2022 at 1:23 AM...

Hi. 'pestcontrol' will be a good example to start with. https://github.com/QUVA-Lab/COMBO/blob/e16358d706525874fb3b1a79c36b6200d4689ed9/COMBO/main.py#L28 In order to run COMBO, you need to run the function here with a properly defined 'objective' argument. In...

In pestcontrol, There are 25(PESTCONTROL_N_STAGES) categorical variables each of which has 5(PESTCONTROL_N_CHOICE) categories. Thus self.n_vertices = np.array(5, 5, ..., 5) where self.n_vertices.shape = (25,) In this example, the numbers of...

Hi, Natalie. If you want to set your first categorical variable to '1' in all your initial points and others random. for example in Ising problem. then you can simply...

In order to set all to '1', then you can just call self.suggested_init[:, 0] = 1 [making the first row have all '1' values] For example, here https://github.com/QUVA-Lab/COMBO/blob/e16358d706525874fb3b1a79c36b6200d4689ed9/COMBO/experiments/test_functions/binary_categorical.py#L100 you can...

Hi, Natalie You can ignore the comment which is about learning underlying graph structure, which is not supported. In your experiment, collecting 10 samples takes almost 40 minutes? In COMBO,...

The elementwise slice sampling is nothing but a Gibbs sampling using a sequence of one dimensional slice samplers. In order for sampling to be the Markov chain, one dimensional slice...

Sometimes, I multiply adj. mat. with some constant for numerical stability. The number is chosen by checking kernel values with some preliminary kernel computations before doing BO experiments. This is...

What I did is nothing special. Because I applied to COMBO to kind of high dimensional problems in which there are 20~60 variables, kernel computation becomes the product kernel of...