CliMA / CalibrateEmulateSample.jl

Stochastic Optimization, Learning, Uncertainty and Sampling

https://clima.github.io/CalibrateEmulateSample.jl/dev

Apache License 2.0

85 stars 17 forks source link

Roadmap #3

Open simonbyrne opened 5 years ago

simonbyrne commented 5 years ago

Functionality

[x] EKI
[x] higher-level interface for NEKI
[x] Gaussian process interpolation
- Can we use https://github.com/STOR-i/GaussianProcesses.jl or do we need to write our own?
[x] MCMC functionality (@glwagner, @sandreza)
- [x] Random-walk MH
- [ ] Slice sampling
- [ ] HMC (https://github.com/tpapp/DynamicHMC.jl)
[ ] Experimental design?
[ ] Multi-fidelity (@odunbar)
[ ] Parameter handling?

Test/Examples

[x] Linear case (@agarbuno)
- check it converges to the correct distribution
[x] Lorenz examples (@dburov190)
[ ] Elliptic / Darcy flow PDE (@agarbuno)
[ ] Simple MCMC examples
- Gaussian
- heavy tail
- Gaussian mixture
- Double well
[ ] Single column model
- Can compare direct MCMC with NEKI/GP approximation

For next meeting:

Lorenz model / GP (@dburov190 )
NEKI/EKI with simple examples (@agarbuno @odunbar @jinlong83)
Single column model / MCMC (@glwagner @sandreza)

dburov190 commented 5 years ago

I've marked "Gaussian processes" done for now; the implementation is in GPR.jl. It uses ScikitLearn.jl, which is quite bulky and ugly, but good enough for now. Planning to switch to GaussianProcesses.jl in the future.