Improve scaling of inputs and optionally target values for the Gaussian process
|Reported by:||gkronber||Owned by:||gkronber|
There are some issues related to scaling of data in the Gaussian process model.
- input features are always scaled to the range [0..1]. This has the effect that the period length in the periodic covariance function cannot be set to a natural value. Assume that we have a time series with daily observations where we want to model a weekly periodic signal. In this case it would be great if we could set the period = 7 in the periodic covariance function. However, because of scaling the correct value depends on minimum and maximum observed day.
- All hyper-parameters are initialized randomly using a log-uniform distribution. This is a good assumption for length scales if all input variables have been scaled to the same range. But for other hyper-parameters (esp. for hyper-parameters for mean functions) this initialization might not be ideal. Especially, for the constant mean function and for the noise hyper-parameter of the Gaussian likelihood it might be better to initialize these hyperparameters to the empirical mean and variance.
Change History (7)
comment:4 Changed 18 months ago by gkronber
- Milestone changed from HeuristicLab 4.0.x Backlog to HeuristicLab 3.3.13
- Owner set to gkronber
- Status changed from new to accepted
comment:5 Changed 18 months ago by gkronber
- Owner changed from gkronber to mkommend
- Status changed from accepted to reviewing
comment:6 Changed 18 months ago by mkommend
- Owner changed from mkommend to gkronber
- Status changed from reviewing to readytorelease
Note: See TracTickets for help on using tickets.