Create Wasserstein metric for ABC discrepancy measures

[ben18785]

[MichaelClerx]

Are you planning to sample some metric that compares 2 probability distributions?

[ben18785]

The idea is to compare the raw data themselves through this metric. I should have put a reference to the paper I was thinking of before. It's this one.

[MichaelClerx]

Cool! Are they fitting rather than sampling with ABC here? It sounds suspiciously like they want to fit a distribution to me:

propose to avoid the use of summaries and the ensuing loss of information by instead using the Wasserstein distance between the empirical distributions of the observed and synthetic data.

[ben18785]

Nope. They are comparing the empirical distribution of the actual and simulated data!

…

On 1 Aug 2019, at 00:09, Michael Clerx ***@***.***> wrote: Cool! Are they fitting rather than sampling with ABC here? It sounds suspiciously like they want to fit a distribution to me: propose to avoid the use of summaries and the ensuing loss of information by instead using the Wasserstein distance between the empirical distributions of the observed and synthetic data. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or mute the thread.