AutoGP API

The purpose of this package is to automatically infer Gaussian process (GP) models of time series data, hence the name AutoGP.

At a high level, a Gaussian process $\{X(t) \mid t \in T \}$ is a family of random variables indexed by some set $T$. For time series data, the index set $T = \mathbb{R}$ is typically (a subset) of the real numbers. For any list of $n$ time points $[t_1, \dots, t_n]$, the prior distribution of the random vector $[X(t_1), \dots, X(t_n)]$ is a multivariate Gaussian,

\[\begin{aligned} \begin{bmatrix} X(t_1)\\ \vdots \\ X(t_n)\\ \end{bmatrix} \sim \mathrm{MultivariteNormal} \left( \begin{bmatrix} m(t_1)\\ \vdots \\ m(t_n)\\ \end{bmatrix}, \begin{bmatrix} k_\theta(t_1, t_1) & \dots & k_\theta(t_1, t_n) \\ \vdots & \ddots & \vdots \\ k_\theta(t_n, t_1) & \dots & k_\theta(t_n, t_n) \\ \end{bmatrix} \right). \end{aligned}\]

In the equation above $m : T \to \mathbb{R}$ is the mean function and $k_\theta: T \times T \to \mathbb{R}_{\ge 0}$ is the covariance (kernel) function parameterized by $\theta$. We typically assume (without loss of generality) that $m(t) = 0$ for all $t$ and focus our modeling efforts on the covariance kernel $k_\theta$. The structure of the kernel dictates the qualitative properties of $X$, such as the existence of linear trends, seasonal components, changepoints, etc.; refer to the kernel cookbook for an overview of these concepts.

Gaussian processes can be used as priors in Bayesian nonparametric models of time series data $[Y(t_1), \dots, Y(t_n)]$ as follows. We assume that each $Y(t_i) = g(X(t_i), \epsilon(t_i))$ for some known function $g$ and noise process $\epsilon(t_i)$. A common choice is to let $Y$ be a copy of $X$ corrupted with i.i.d. Gaussian innovations, which gives $Y(t_i) = X(t_i) + \epsilon(t_i)$ where $\epsilon(t_i) \sim \mathrm{Normal}(0, \eta)$ for some variance $\eta > 0$.

Writing out the full model

\[\begin{aligned} k &\sim \mathrm{PCFG} &&\textrm{(prior over covariance kernel)} \\ [\theta_1, \dots, \theta_{d(k)}] &\sim p_{k} && \textrm{(parameter prior)}\\ \eta &\sim p_{\eta} &&\textrm{(noise prior)} \\ [X(t_1), \dots, X(t_n)] &\sim \mathrm{MultivariteNormal}(\mathbf{0}, [k_\theta(t_i,t_j)]_{i,j=1}^n) &&\textrm{(Gaussian process)} \\ Y(t_i) &\sim \mathrm{Normal}(X(t_i), \eta), i=1,\dots,n &&\textrm{(noisy observations)} \end{aligned}\]

where PCFG denotes a probabilistic-context free grammar that defines a language of covariance kernel expressions $k$,

\[\begin{aligned} B &:= \texttt{Linear} \mid \texttt{Periodic} \mid \texttt{GammaExponential} \mid \dots \\ \oplus &:= \texttt{+} \mid \texttt{*} \mid \texttt{ChangePoint} \\ k &:= B \mid \texttt{(}k_1 \oplus k_2\texttt{)}. \end{aligned}\]

and $d(k)$ is the number of parameters in expression $k$. Given data $\{(t_i,y_i)\}_{i=1}^n$, AutoGP infers likely values of the symbolic structure of the covariance kernel $k$, the kernel parameters $\theta$, and the observation noise variance $\eta$.

The ability to automatically synthesize covariance kernels in AutoGP is in contrast to existing Gaussian process libraries such as GaussianProcess.jl sklearn.gaussian_process, GPy, and GPyTorch, which all require users to manually specify $k$.

After model fitting, users can query the models to generate forecasts, compute probabilities, and inspect qualitative structure.

AutoGP — Module

Main module.

Exports

AutoGP API

Model Initialization

End-to-End Model Fitting

Incremental Model Fitting

Model Querying

Serialization