How do you determine which is the best next point where to evaluate the model?
Simulated annealing is blind, what you do with simulated annealing you just move locally around the current point.
If you had a partial picture of the landscape of the function, how would you locate the most interesting unvisited point in the landscape?
I do have to admit that what I'm seeing here is only the application of Bayesian sampling algorithms, there are startups doing similar things for A/B testing and similar.
The significant change in Bayesian methods is moving away from Metropolis Hastings to an approach based on Hamiltonian dynamics, implemented widely in the STAN package. It would be interesting to see if this is what is being used.
My understanding of simulated annealing is that the "fudge factor" is chosen and reduced more or less "naively". By which I mean it's initial value and reduction across iterations is hard wired into the execution path before it runs and...the parameters for iteration \i\ are determined before iteration \1\.
My understanding of the article is that the new algorithm determines the parameters of iteration \i\ statistically from the results of iterations \1\ through \i-1\.
From a cursory reading of the article it sounds like they optimize for the best annealling schedule which of course is what every manufacturer of steel has been doing with varying degrees of success for at least a thousand years.
From a cursory reading of the article this looks like it's solving an extremely similar problem to Gaussian Process Bayesian Optimization (GBPO [1]) in a fairly similar way. I wonder what a head-to-head comparison between these two would look like.
[1] J. Gardner, M. Kusner, Z. Xu, K. Weinberger, and J. Cunningham,
in Proceedings of the 31st International Conference on Machine
Learning (ICML-14) (JMLR Workshop and Conference Proceedings,
2014), p. 937
I don't understand this paper yet, but Marzouk is an awesome fella--a great explainer, a brilliant researcher, etc. I almost took a job with him a few years ago.
If I get through this paper quick enough, I'll post back any significant thoughts I may have (if any).
9 comments
[ 5.1 ms ] story [ 27.6 ms ] threadSimulated annealing is blind, what you do with simulated annealing you just move locally around the current point.
If you had a partial picture of the landscape of the function, how would you locate the most interesting unvisited point in the landscape?
I do have to admit that what I'm seeing here is only the application of Bayesian sampling algorithms, there are startups doing similar things for A/B testing and similar.
Although I'm vaguely familiar with MCMC and Metropolis, it's not nearly enough to pass comment on this advance.
[1] http://arxiv.org/pdf/1402.1694v4.pdf
My understanding of the article is that the new algorithm determines the parameters of iteration \i\ statistically from the results of iterations \1\ through \i-1\.
[1] J. Gardner, M. Kusner, Z. Xu, K. Weinberger, and J. Cunningham, in Proceedings of the 31st International Conference on Machine Learning (ICML-14) (JMLR Workshop and Conference Proceedings, 2014), p. 937
[1] https://github.com/sigopt/MOE
If I get through this paper quick enough, I'll post back any significant thoughts I may have (if any).