This issue is to summarise the next steps in the ACE + IPFitting game, comments and discussion are very welcome. This roadmap will be adjusted according to discussions.
ACE 1.x
Overall Goal: rock-solid and usable linear regression, some basic hyper parameter optimisation
High Priority:
[x] fix slow basis assembly since v0.6 (cf. #10 for further work)
[x] fix crashing multi-threaded code in IPFitting
[x] shift to DAG evaluation code as default (cf v0.7.0)
[x] general repulsive core (cf v0.7.0)
[x] tests for cross-version compatibility and performance regression (will enforce compatibility only from v1.0.0 onwards) (cf. v0.7.1)
[x] #52
[ ] general cleanup and revamp of IPFitting (@casv2)
Next steps, could be 1.x or beyond
[x] Create a "wrapper" module that gives a nicer user-interface that will ideally remain independent of developments on the ACE backend (to discuss with Cas)
[ ] Generate basis for fixed cost rather than for fixed degree; cf. #2
[ ] general sure-fire way to generate repulsive cores (Gabor, Cas, CO to discuss)
[ ] real basis (CO); this should give an extra factor 2
[ ] optimisation of distance transform (Andres)
[ ] optimisation of radial basis (Andres)
[ ] Sparsification of basis (IPFitting, sparse priors, Matthias, ...)
[ ] automated weight selection through iteratively reweighed least squares (IPFitting)
[ ] Orthogonal basis (= pure basis; to be discussed with Simon)
[ ] Revisit regularisation with different metrics
[ ] kernel and GP type regression using ACE basis as feature vectors
[ ] Incorporate aPIPs via code-generated DAG evaluation
ACE 2.x
ACE 2.x will focus on improved composability and nonlinear models
[x] arbitrary densities
[x] bond model (cylindrical symmetry) (CO, best to move to 2.x)
[x] neighbour counting models
[x] rewrite multi-species via discrete density
[x] nonlinear combination (e.g. Finnis Sinclair) - this could be shifted to 1.x
This issue is to summarise the next steps in the ACE + IPFitting game, comments and discussion are very welcome. This roadmap will be adjusted according to discussions.
ACE 1.x
Overall Goal: rock-solid and usable linear regression, some basic hyper parameter optimisation
High Priority:
IPFitting
IPFitting
(@casv2)Next steps, could be 1.x or beyond
ACE 2.x
ACE 2.x will focus on improved composability and nonlinear models
cc @gabor1 @casv2 @andresrossb