spearmint
spearmint copied to clipboard
JasperSnoek
Examine objective as a function of parameters for all experiments so far
Is there any way I can access the value of the objective function for all experiments run so far together with the parameters used for each? I would like to visualise how the objective varies over the parameter space. Thanks, John.
I'm starting to setup a web interface (actually Jeff Rose submitted it and I'm iterating on it) to display various statistics and a useful list of completed experiments, etc. If you run spearmint with the flag -w it will return a link to the web status. I'd love feedback on this - e.g. what would be useful to see on that page!
Jasper
Nice. The web interface makes it much easier to see what is going on. Does the spearmint algorithm estimate the noise level or is it fixed? I'd be particularly interested in seeing a current estimate as I imagine it is quite easy for the GP to model real variation in the objective function as noise (and vice versa I guess).
Could you make the port number of the web interface a command line option? It would be nice to keep it fixed to avoid too much SSH tunnelling.
Ok done and done. I need to think of a nicer way of visualizing the noise - but it's there.
Great. Thanks. I'm assuming that's the SD of the noise term rather than the variance. If you want a nicer way to visualise it perhaps you could put error bars on the plot of the best objective so far.
Yes, that value is the variance of the noise.
Jasper
On Wed, Oct 23, 2013 at 6:33 AM, JohnReid [email protected] wrote:
Now I'm beginning to suspect that it is the variance of the noise term.
— Reply to this email directly or view it on GitHubhttps://github.com/JasperSnoek/spearmint/issues/10#issuecomment-26894812 .