MFES-HB: Efficient Hyperband with Multi-Fidelity Quality Measurements

[nabenabe0928]

MFES-HB: Efficient Hyperband with Multi-Fidelity Quality Measurements

Main points

• Train $K$ surrogate models using observations on $K$ different fidelities.
• Combine these surrogate models using rank loss-based weighting as in TST-R or RGPE
• Basically, BOHB using lower-fidelities.

Visualization has nothing new. Normal comparisons only. Only self-developed benchmarks.