Choosing a model

Choose the smallest model that matches the data and latency requirements.

Need	Recommended model	Feature map	Solver	Backend
Fast baseline on tabular data	Batch ELM	RandomAdditiveMap	BatchRidgeSolver	CPU or GPU
Nonlinear toy with local structure	RBF ELM	RbfMap	BatchRidgeSolver	CPU
Incremental data chunks	OS-ELM	RandomAdditiveMap or RbfMap	RlsSolver	CPU
Concept drift	FOS-ELM	RandomAdditiveMap	RlsSolver with forgettingFactor < 1	CPU
Ill-conditioned online initialization	ReOS-ELM	RandomAdditiveMap	RlsSolver with regularization > 0	CPU
Class-distance constraint	OS-CELM	RandomAdditiveMap	RlsSolver with constraint = kClassDistance	CPU
Learned feature extraction	ML-ELM	StackedFeatureMap<ElmAutoEncoderLayer>	BatchRidgeSolver	CPU or GPU
Online learned feature stack	H-OS-ELM	StackedFeatureMap<ElmAutoEncoderLayer>	RlsSolver	CPU or GPU

Decision guide

1. Start with BatchElm when you need a deterministic baseline.
2. Switch to MlElm when a single random additive layer underfits nonlinear structure.
3. Use OsElm when data arrives in chunks and retraining the full batch is too expensive.
4. Enable forgetting only when the stream distribution changes over time.
5. Use RbfMap when centers and width are meaningful for the problem geometry.
6. Keep IdentityMap for ablations and linear baselines.

Backend selection

• Backend::kCpu is the correctness reference and is required for all algorithms.
• Backend::kGpu is intended for batch feature transforms and ridge solves.
• GPU tests skip cleanly when no CUDA device is visible.