The prediction results of kmean are all the same & cause the svm error

[cysun0226]

Hi,

Recently I am trying to use LA-MCTS to optimize my own task, but some errors occurred during the execution.

My task has 9 dimensions, and the HP setting is:

• Cp = 10 ( follow the suggestion from the paper, ~= 10% of max f(x) )
• leaf_size = 10
• ninits = 40
• kernel_type = "rbf"

Here is the error log:

Traceback (most recent call last):
  File "mcts-exec.py", line 30, in <module>
    agent.search(iterations = args.iterations)
  File "/root/workspace/lamcts/MCTS.py", line 239, in search
    self.dynamic_treeify()
  File "/root/workspace/lamcts/MCTS.py", line 116, in dynamic_treeify
    bad_kid.update_bag(  bad_kid_data  )
  File "/root/workspace/lamcts/Node.py", line 83, in update_bag
    self.is_svm_splittable = self.classifier.is_splittable_svm()
  File "/root/workspace/lamcts/Classifier.py", line 55, in is_splittable_svm
    self.learn_boundary(plabel)
  File "/root/workspace/lamcts/Classifier.py", line 410, in learn_boundary
    self.svm.fit(self.X, plabel)
  File "/opt/app-root/lib/python3.6/site-packages/sklearn/svm/_base.py", line 173, in fit
    y = self._validate_targets(y)
  File "/opt/app-root/lib/python3.6/site-packages/sklearn/svm/_base.py", line 560, in _validate_targets
    " class" % len(cls))
ValueError: The number of classes has to be greater than one; got 1 class

Since it occurs at Classifier.learn_clusters(), I print the result of self.kmean.predict(tmp) and find some clues:

• The error will occur if plabel (the results of self.kmean.predict(tmp)) are all the same

# normal
plabel: [0 1 1 0 0 0 1 0 0 1 1]
# exception
plabel: [0 0 0 0 0]

I temporarily avoid this exception by making is_splittable_svm return False when plabel only contains a single class.

However, I would like to know that is it possible to happen in the general case? Or it may be caused by my own function?

Thank you for the work & sharing your code!

[linnanwang]

Hello,

Thank you for submitting this issue, we're aware of this problem. For most of the problems, the current code should be fine.

Here are suggested solutions,

1. normalize your f(x) inside your function.
2. change the leaf size to a larger size.
3. instead of using Kmean, we can learn a linear or non-linear regressor to separate the samples into two groups, rather using K-mean. We will release this feature after NeurIPS of this year.

[cysun0226]

Thanks for the reply and suggestions!
Look forward to your future works & update 🙌

[linnanwang]

Thank you. Before I will close this issue, can I ask if normalization or increasing the leaf size help you resolve the issue? Thanks.

please note if each elements of x and f(x) in a very small range, [-0.01, 0.01], or entires of [x, f(x)] are very similar this scenario can happen.