

Supported Features

Summary

"If you're not sure whether DynaML fits your requirements, this list provides a semi-comprehensive overview of available features."

Models¶

Model Family	Supported	Notes
Generalized Linear Models	Yes	Supports regularized least squares based models for regression as well as logistic and probit models for classification.
Generalized Least Squares Models	Yes	-
Least Squares Support Vector Machines	Yes	Contains implementation of dual LS-SVM applied to classification and regression.
Gaussian Processes	Yes	Supports gaussian process inference models for regression and binary classification; the binary classification GP implementation uses the Laplace approximation for posterior mode computation. For regression problems, there are also multi-output and multi-task GP implementations.
Student T Processes	Yes	Supports student T process inference models for regression, there are also multi-output and multi-task STP implementations.
Multi-output Matrix T Process	Yes	_
Skew Gaussian Processes	Yes	Supports extended skew gaussian process inference models for regression.
Feed forward Neural Networks	Yes	Can build and learn feedforward neural nets of various sizes.
Committee/Meta Models	Yes	Supports creation of gating networks or committee models.

Solver	Supported	Notes
Regularized Least Squares	Yes	Solves the Tikhonov Regularization problem exactly (not suitable for large data sets)
Gradient Descent	Yes	Stochastic and batch gradient descent is implemented.
Quasi-Newton BFGS	Yes	Second order convex optimization (using Hessian).
Conjugate Gradient	Yes	Supports solving of linear systems of type $A.x = b$ where $A$ is a symmetric positive definite matrix.
Committee Model Solver	Yes	Solves any committee based model to calculate member model coefficients or confidences.
Back-propagation	Yes	Least squares based back-propagation with momentum and regularization.

Solver	Supported	Notes
Grid Search	Yes	Simple search over a grid of configurations.
Coupled Simulated Annealing	Yes	Supports vanilla (simulated annealing) along with variants of CSA such as CSA with variance (temperature) control.
ML-II	Yes	Gradient based optimization of log marginal likelihood in Gaussian Process regression models.