BBOB Landscape Atlas

The BBOB Landscape Atlas is a separate public viewer of the 24 functions of the bbob suite. It shows each function as a rotating heat map or an elevation view, drawn from exact function values of instance 1 on a 201 by 201 grid, with a classification label for each function. The recorded runs on the How it works page use the same instance.

Links hidden by default

The links name the author’s account, so they appear only when the build shows author details (AMM_SHOW_AUTHOR=1). They are hidden by default.

Why different functions need different engines

Which classical algorithm is best for a black-box problem depends on the budget and on the landscape. Around 10 evaluations per dimension, model-based trust-region methods lead. On smooth functions, quasi-Newton methods with finite-difference gradients converge fast. With thousands of evaluations per dimension, restarted CMA-ES with growing populations finds global structure that local methods miss. The Atlas makes the landscape side of that statement visible: a smooth valley, a field of regularly placed local optima and a set of basins of different depth call for different tools.

The user usually knows the budget before the run, but not the landscape, so AMM reads the landscape with probes. The paper draft gives one example: f3 and f15 are the same Rastrigin function without and with a rotation. A router that reads only the dimension and the budget must send both to the same engine, while AMM’s pair test, at a cost of 3D+1 evaluations, sends f3 to STEP restarts and leaves f15 on the general path. The How it works page lists every probe and the engine it selects.