# Project roadmap

Publication venues:
- ICML (deadline January)
- [Climate Informatics](http://www.climateinformatics.org/conferences/) (deadline probably Feburary)

Proposed paper contents:
- Focus on algorithm that supports 3D grids but perhaps not time?
- Derivation of the algorithm (roughly following So's Overleaf)
- "Clever bits" of GPU implementation
- Demonstrations on simulated data
- Large scale application to real data


## Tasks
Each task has a priority:
P1 = high priority, P4 = low priority

I've then ordered the themes roughly in order of decreasing priority based on the tasks they contain.

#### Improve the INLA integration
- [P2] Why does it converge to something different to MP and 3D-Var?
    - Noise currently isn't set correctly (I think).
- [P3] Directly use solve functions suggested by INLA people rather than main INLA method, to reduce overheads
- [P4] Avoid writing the observations to disk when transferring between Python and R

#### Large scale applications for 2D GP
- [P3] Implement multi-GPU

#### Use MC method for the variance
Possibly following ["Efficient Covariance Approximations for Large Sparse Precision Matrices"](https://arxiv.org/abs/1705.08656)
- [P4] Can we use the output of this method as an input to the message passing to speed up convergece?
- [P4] Reimplement Marc's code from Slack in JAX
- [P4] Look at work of [Sharana Kumar Shivanand](https://www.turing.ac.uk/people/researchers/sharana-kumar-shivanand) who gave presentation on this to F-X's group meeting

#### Improve the convergence speed of the mean
- [P4] Try Aitken's and Steffenson’s method from [this paper](https://arxiv.org/abs/0810.1119)
- [P4] Try randomised message scheduling from [this paper](https://arxiv.org/pdf/1909.11469.pdf)

#### Investigate and write up why we can't choose good cs