Agnostic Extension of Contaminated PAC Learning

Interactive exploration of PAC learning under iterative synthetic contamination in the agnostic setting. Adjust parameters to see how contamination and noise interact.

Contamination alpha

0.25

Label Noise Rate eta

0.10

Samples per Round

800

Rounds

opt_H (best-in-class)

Naive ERM Final

Weighted ERM Final

Regularized ERM Final

Algorithm Comparison Across Rounds

Error vs Contamination Fraction

Noise-Contamination Interaction

Sample Complexity Scaling

Conjectured Error Bound

err(h_T) ≤ opt_H + C·sqrt(VC(H)·log(1/δ) / n_eff) + Πα_t · (0.5 - opt_H), where n_eff = n·Π(1-α_t). The three terms decompose into irreducible approximation error, statistical estimation error with effective sample size, and contamination amplification of the initial excess.