Anytime Analysis on BinVal: Adaptive Parameters Help

A new theoretical computer science paper from researchers Timo Kötzing and Jurek Sander provides rigorous mathematical proof that evolutionary algorithms with adaptive parameters significantly outperform their fixed-parameter counterparts. Published on arXiv and accepted to GECCO 2026, the work analyzes the "anytime performance" of evolutionary and estimation-of-distribution algorithms on the BinVal benchmark problem, measuring how quickly they can find progressively better solutions rather than just the final optimum.

For the standard (1+1) Evolutionary Algorithm with a fixed mutation rate of 1/n, the researchers proved the fixed-target runtime to optimize the most significant k bits is Θ(n log k). However, when they replaced the fixed rate with a self-adjusting mutation rate, the runtime dropped dramatically to O(k^(1+ε)) for any constant ε > 0. This means the adaptive algorithm's performance becomes essentially independent of the total problem size n, scaling only with the number of bits k being optimized.

The significance lies in the mathematical guarantee that holds simultaneously for all k ∈ o(n), meaning the adaptive algorithm maintains its advantage throughout the entire optimization process. This provides theoretical justification for what practitioners have observed empirically: that algorithms that can automatically adjust their parameters during execution tend to perform better across diverse problem domains without requiring manual tuning for each specific instance.