Likelihoods & Inference

Changing Likelihood & Inference

Suggestions of which likelihood and inference method to use is implicitly given by default,

  • GPR uses Gaussian likelihood and exact inference.
  • GPC uses Erf likelihood and EP inference.
  • FITC model uses same default with corresponding FITC inference.
  • GPMC calls GPC and thus uses the default setting of GPC

You can change to other likelihood or inference methods using:

model.useLikelihood(newLik)
model.useInference(newInf)
newLik and newInf are Strings. Currently the options are:
  1. Regression model
    • newLik: “Laplace”. Note this will force inference method to be EP.
    • newInf: “EP”, “Laplace”.
  2. Classification model (including GPMC)
    • newInf: “Laplace”

To be consistent with Gaussian Processes community, we use the name “Laplace” for both Laplace likelihood and Laplace inference. Please note the differences.