pints
pints copied to clipboard
pints-team
Exact and approximate simulation methods for stochastic models
For models that we use to exemplify PINTS' ABC capabilities - in particular, stochastic reaction equations - there are sometimes exact (computationally expensive) and approximate (computationally cheap) ways to simulate from these series. For example, the Gillespie algorithm is an exact SSA whereas tau leaping is approximate. Some ABC methods, such as lazy ABC, use both cheap and expensive simulation methods to simulate and so, think it'd be good to include these.
To this end, I would propose making an abstract toy model class for stochastic reaction equations. @danielfridman98 has started this work by coding up the stochastic degradation algorithm in lightning fashion!
Hmmmm! I've seen some other methods that use fast&cheap versus slow&exact, but should we be investigating these? I certainly don't have any models where this is possible!
Yep, think we should -- in typical ABC problems they often exist. They're a class of ABC methods that are of growing importance.
On Fri, Aug 2, 2019 at 4:32 PM Michael Clerx [email protected] wrote:
Hmmmm! I've seen some other methods that use fast&cheap versus slow&exact, but should we be investigating these? I certainly don't have any models where this is possible!
— You are receiving this because you were assigned. Reply to this email directly, view it on GitHub https://github.com/pints-team/pints/issues/889?email_source=notifications&email_token=ABCILKEYVZ6T4L3HJWFG35TQCRHQXA5CNFSM4II6XJKKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOD3OCVXI#issuecomment-517745373, or mute the thread https://github.com/notifications/unsubscribe-auth/ABCILKCO6TLG6FALCKII2C3QCRHQXANCNFSM4II6XJKA .
We should maybe discuss this at the next PINTS meeting: don't want to be PINTSing for the sake of it
Happy to discuss! Definitely worth reading up on this though if you have time -- the lazy ABC paper and lots of Ruth Baker's papers are based on this idea.
Cheers,
Ben
On Fri, Aug 2, 2019 at 4:38 PM Michael Clerx [email protected] wrote:
We should maybe discuss this at the next PINTS meeting: don't want to be PINTSing for the sake of it
— You are receiving this because you were assigned. Reply to this email directly, view it on GitHub https://github.com/pints-team/pints/issues/889?email_source=notifications&email_token=ABCILKELDJ6MKMLURBL44K3QCRIIFA5CNFSM4II6XJKKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOD3ODG7A#issuecomment-517747580, or mute the thread https://github.com/notifications/unsubscribe-auth/ABCILKBY2PNUBBHN2OLCCFLQCRIIFANCNFSM4II6XJKA .
Just some added info on optimising/sampling approximate (cheaper) functions.
There are optimisation methods called Sequential Model-based Global Optimization methods that do this. In the context of ODEs gradient matching is another technique that uses a different score function to speed things up #22
Multi-fidelity methods used in cardiac activation time simulations https://doi.org/10.1016/j.jcp.2019.03.026