ReOS-ELM and FOS-ELM¶
ReOS-ELM and FOS-ELM are not separate classes in v2. They are toggles on RlsOptions<FloatT>.
ReOS-ELM¶
ReOS-ELM regularizes the initial covariance used by the online solver.
feature_elm::RlsOptions<float> options;
options.regularization = 1e-2f;
options.forgettingFactor = 1.0f;
options.constraint = feature_elm::RlsConstraint::kNone;
Use it when:
- The initial chunk is small.
- The hidden feature matrix is ill-conditioned.
- Early online updates produce unstable weights.
The extension is off when regularization = 0.
FOS-ELM¶
FOS-ELM applies a forgetting factor to down-weight stale data.
feature_elm::RlsOptions<float> options;
options.regularization = 0.0f;
options.forgettingFactor = 0.98f;
options.constraint = feature_elm::RlsConstraint::kNone;
Use it when:
- The stream distribution changes over time.
- Recent labels are more informative than old labels.
- You need to compare against ordinary OS-ELM.
The extension is off when forgettingFactor = 1.
Drift comparison¶
xychart-beta
title "Expected drift behavior"
x-axis "Samples seen" [0, 250, 500, 750, 1000]
y-axis "Accuracy" 0 --> 1
line "OS-ELM" [0.92, 0.90, 0.72, 0.68, 0.66]
line "FOS-ELM" [0.91, 0.89, 0.74, 0.82, 0.86]
FOS-ELM should recover faster after the configured drift point because older chunks receive lower weight.
References¶
- Liang, Huang, Saratchandran, and Sundararajan. 2006. OS-ELM.
- Online sequential learning and forgetting-factor variants in the ELM literature.