Hello,
I am testing LA-MCTS together with TuRBO-1 and found it works well. I tried to increase the batch_size in TuRBO-1 to reduce the time spent on GP fitting. From what I got, the TuRBO paper indicates this is a legit choice for increasing overall computational efficiency without affecting the quality of the results.
To this end, I only modified collect_samples() in MCTS.py to handle the multiple values returned by each TuRBO-1 iteration. Nothing particularly ambiguous here.
And now on rare occasion I get an AssertionError thrown at line 47 in update_kids() in Node.py:
assert self.kids[0].classifier.get_mean() > self.kids[1].classifier.get_mean()
I'm not sure why just yet. And I don't understand why returning more samples from TuRBO-1 could trigger this.
Could you please indicate possible reasons for this?
I believe increasing batch_size in TuRBO-1 is something desirable and could benefit many.
Or am I misunderstanding something in the interplay between LA-MCTS and TuRBO-1?
Thanks for the great algo!
Hello,
I am testing LA-MCTS together with TuRBO-1 and found it works well. I tried to increase the
batch_sizein TuRBO-1 to reduce the time spent on GP fitting. From what I got, the TuRBO paper indicates this is a legit choice for increasing overall computational efficiency without affecting the quality of the results.To this end, I only modified
collect_samples()in MCTS.py to handle the multiple values returned by each TuRBO-1 iteration. Nothing particularly ambiguous here.And now on rare occasion I get an
AssertionErrorthrown at line 47 inupdate_kids()in Node.py:assert self.kids[0].classifier.get_mean() > self.kids[1].classifier.get_mean()I'm not sure why just yet. And I don't understand why returning more samples from TuRBO-1 could trigger this.
Could you please indicate possible reasons for this?
I believe increasing
batch_sizein TuRBO-1 is something desirable and could benefit many.Or am I misunderstanding something in the interplay between LA-MCTS and TuRBO-1?
Thanks for the great algo!