You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Currently, the MICE algorithm uses the old capabilities of GPs and hasn't been adapted for the analytical mean function. This means that the fitting won't work if using a mean function (if just fitting a zero mean it should still be correct).
Updating the algorithm to do MICE naively (i.e. refitting the GP at each step when considering a candidate point) should be straightforward to implement, but will be computationally slow as it requires inverting the covariance matrix for each candidate point. If doing a design with 100 or so candidates (a reasonable number), this can be a significant slowdown.
The old implementation exploits block matrix inversion to improve the fitting performance (this is implemented in the MICEFastGP class). This won't work out of the box on the analytical mean function, as the analytical formulae using the design matrix haven't been updated to exploit the block inversion. Additionally, the inverse of the covariance matrix is never explicitly cached in the current implementation, so an updated version will need to adapt the algorithm to do back-substitution in the solves (or form the inverse explicitly if you can't write the operations using back-substitution).
The text was updated successfully, but these errors were encountered:
Currently, the MICE algorithm uses the old capabilities of GPs and hasn't been adapted for the analytical mean function. This means that the fitting won't work if using a mean function (if just fitting a zero mean it should still be correct).
Updating the algorithm to do MICE naively (i.e. refitting the GP at each step when considering a candidate point) should be straightforward to implement, but will be computationally slow as it requires inverting the covariance matrix for each candidate point. If doing a design with 100 or so candidates (a reasonable number), this can be a significant slowdown.
The old implementation exploits block matrix inversion to improve the fitting performance (this is implemented in the
MICEFastGP
class). This won't work out of the box on the analytical mean function, as the analytical formulae using the design matrix haven't been updated to exploit the block inversion. Additionally, the inverse of the covariance matrix is never explicitly cached in the current implementation, so an updated version will need to adapt the algorithm to do back-substitution in the solves (or form the inverse explicitly if you can't write the operations using back-substitution).The text was updated successfully, but these errors were encountered: