Algorithm Configuration Landscapes:
More Benign than Expected?


Automated algorithm configuration finds high quality parameter settings that can improve algorithm performance by several orders of magnitude. Formally, automated algorithm configuration can be defined as an optimization problem with algorithm A with parameters p that we want to optimize with respect to metric m on a set I of problem instances.

Existing algorithm configuration procedures are typically heavily based on powerful meta-heuristics originally designed for complex and challenging search landscapes. The use of these heuristics implicitly assumes that configuration landscapes are also challenging to optimize.

We hypothesize that algorithm configuration landscapes are typically much more benign than previously expected — perhaps even uni-modal or convex; a claim that we support with empirical evidence.