Configuration

The v2 code exposes configuration through model constructors and option structs rather than one monolithic ElmConfig. The effective configuration surface is the same: feature stack, solver options, backend, and seed.

Common fields

Field	Type	Default	Applies to	Notes
numInputs	std::size_t	required	all models	Input dimension
numHiddenNodes	std::size_t	required	BatchElm, OsElm, OsCelm	Additive hidden-node count
hiddenNodesPerLayer	std::vector<std::size_t>	required	MlElm, HierarchicalOsElm	One ELM-AE layer per entry
ActivationFunction / ActivationKind	enum	kSigmoid	additive maps and ELM-AE	kSigmoid, kTanh, kRelu
Backend	enum	kCpu	all models and feature maps	kCpu or kGpu
seed	unsigned int	42u	random maps and ELM-AE	Reproducible initialization
ridgeAlpha	FloatT	1e-6	batch ridge and ELM-AE	Positive regularization

Batch solver options

SolverOptions<FloatT> controls BatchRidgeSolver.

Option	Default	Values	Meaning
ridgeAlpha	1e-6	> 0	Regularizes HᵀH + αI or HHᵀ + αI
path	kAuto	kAuto, kPrimal, kDual	Selects primal when samples dominate hidden nodes and dual otherwise
method	kCholesky	kCholesky, kHouseholderQr	CPU factorization method; Cholesky is the stable default

Toggle path = kPrimal for many samples and few hidden nodes. Toggle path = kDual for many hidden nodes and few samples. Use kHouseholderQr only for research comparisons because Cholesky is faster on the regularized SPD system.

Online solver options

RlsOptions<FloatT> controls RlsSolver.

Option	Default	Values	Meaning
regularization	1e-3	>= 0	ReOS-ELM covariance regularization
forgettingFactor	1	(0, 1]	FOS-ELM forgetting; 1 is ordinary OS-ELM
constraint	kNone	kNone, kClassDistance	OS-CELM class-distance constraint
constraintStrength	1e-2	>= 0	Strength for class-distance constraint

An omitted extension must reduce to the baseline: forgettingFactor = 1, regularization = 0, and constraint = kNone are the off-switches.

RBF configuration

RbfMap is configured by:

Parameter	Default	Notes
inputDim	required	Input dimension
numCenters	required	Output dimension
width	1.0	Shared RBF width σ
centerInit	kRandom	kRandom or kKMeans
seed	42	Random center initialization

Demo configuration

DemoConfig controls the demo server. Every field is read from an environment variable by configFromEnvironment(); the demo images set the paths and port for you.

Field	Environment variable	Default (binary)	Default (images)	Notes
useGpu	DEMO_USE_GPU	false	CPU 0, GPU 1	1/true/yes/on enable it; 0 disables it
host	DEMO_HOST	0.0.0.0	0.0.0.0	Bind address
port	DEMO_PORT, then PORT	8888	7860	7860 is the Hugging Face Spaces default
staticPath	DEMO_STATIC_PATH	demo/ui	/app/ui	Web UI directory
benchmarkPath	DEMO_BENCHMARK_PATH	data/benchmarks/latest	/app/benchmarks	Benchmark JSON directory
datasetPath	DEMO_DATASET_PATH	data/datasets/digits_8x8.csv	/app/data/digits_8x8.csv	Digits CSV
maxHiddenNodes	DEMO_MAX_HIDDEN	2048	CPU 1024, GPU 2048	Upper bound accepted by /api/evaluate

Drift stream configuration

DriftStream::Config controls synthetic non-stationary streams.

Field	Default	Notes
inputDim	2	Minimum is 2
numClasses	2	Binary drift stream
streamLength	1000	Samples before exhaustion
driftPoint	500	Sample where the regime changes
seed	42	Reproducible stream