tutorial

README

GoGP Tutorial

• barebones --- A 'bare bones' use of Gaussian process.
• hyperpriors --- Specifying priors on hyperparameters.
• warpedtime --- Warping input to model non-stationarity.
• anynoise --- Handling non-Gaussian noise.
• events --- Accounting for scheduled events.

Documentation

Index

• Variables
• func Evaluate(gp *gp.GP, m model.Model, theta []float64, rdr io.Reader, wtr io.Writer) error

Variables

var (
	OPTINP      = false
	MINOPT      = 0
	ALG         = "lbfgs"
	PARALLEL    = false
	ITERS       = 1000 // major iterations
	MINITERS    = 10   // minimum iterations to accept in lbfgs
	THRESHOLD   = 1e-6 // gradient threshold
	RATE        = 0.01 // learning rate (for Adam)
	NTASKS      = 0
	NONORMALIZE = false
	OUTOFSAMPLE = false
)

Functions

func Evaluate

func Evaluate(
	gp *gp.GP,
	m model.Model,
	theta []float64,
	rdr io.Reader,
	wtr io.Writer,
) error

Evaluate evaluates Gaussian process on CSV data. One step out of sample forecast is recorded for each time point, along with the hyperparameters. This function is called by all case studies in the tutorial. For optimization, LBFGS from the gonum library (http://gonum.org) is used for faster execution. In general though, LBFGS is a bit of hit-or-miss, failing to optimize occasionally, so in real applications a different optimization/inference algorithm may be a better choice.