Comparison matrix of models versus optimisers versus inference methods

[martinjrobins]

I'm writing a repo that simply takes all the toy models in pints, and all the methods (optimisers and inference), and then tests every method against every model. This will take a while, so its doing it all on arcus-b (there is a lot of machine-specific stuff in there, so it's not suitable to put into Pints itself)

When I'm comparing optimisers, I compare using the following criteria:

• final score
• time taken to reach the final score (this is the time taken using all the cores of a whole node on arcus-b)

I might also average these results over multiple runs of the optimiser, since some of them will be stochastic.

I'm less sure how to compare the inference methods, perhaps:

• Effective sample size (in a given time limit?)
• Rhat
• anything else?

What other criteria do you all think are necessary @MichaelClerx @ben18785 @sanmitraghosh @mirams @chonlei ? I'm hoping this will give a bunch of heat maps comparing the performance of all of our methods, and will go into the first paper

[mirams]

A direct count of the number of forward solves involved in getting to an optimum/converged posterior is nice to have.

[MichaelClerx]

A direct count of the number of forward solves involved in getting to an optimum/converged posterior is nice to have.

Good point! Relates to #203

[martinjrobins]

Here are a couple of heat maps I did using the optimisers, just one run each, and 1% noise:

score_with_noise_0.pdf
time_with_noise_0.pdf

[MichaelClerx]

Just discussing if, for optimisers, we want to show the mean score of multiple runs, or the best score.

@DavidGavaghan ?

[mirams]

I think you want to see the distribution of optimiser results really, will have a big impact on use if they always get the same answer versus have a wide distribution of results.

[martinjrobins]

Yea, I think you're right. I'm storing all the results from all the independent runs of the optimisers, and plotting the mean and minimum scores and execution_time. But all the data is there, so we can post-process any other statistic you might wish for.

…

On 21 February 2018 at 15:08, Gary Mirams ***@***.***> wrote: I think you want to see the distribution of optimiser results really, will have a big impact on use if they always get the same answer versus have a wide distribution of results. — You are receiving this because you were assigned. Reply to this email directly, view it on GitHub <#229 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ABGF9LimpVDkUTm_rTjVrygsIQSkFPFTks5tXDFigaJpZM4SCZxF> .

[MichaelClerx]

@martinjrobins can this one be closed?