Belder et al. slash runtime bound for (μ+1) EA on BinVal by orders of magnitude

A trio of researchers from TU Delft and ETH Zurich—Joris Belder, Johannes Lengler, and Raghu Raman Ravi—have published a significant theoretical improvement in the analysis of the (μ+1) Evolutionary Algorithm (EA) on the Binary Value (BinVal) function. BinVal is a classic benchmark in evolutionary computation known for its deceptive landscape: fitness is the integer value of the bitstring interpreted as a binary number, making it monotonic yet difficult for population-based algorithms because early decisions on high-order bits dominate. The previous best upper bound, due to Krejca, Neumann, and Witt, was O(μ^5 n log(n/μ^4))—a polynomial with a large exponent that made theoretical predictions imprecise.

Belder, Lengler, and Ravi have slashed that bound to O(μ log μ · n log n) under the condition μ = o(n/ log n). This is a dramatic improvement: for typical population sizes, the exponent drops from 5 to 1 in μ. Their proof relies on a refined potential function and a careful analysis of mutation effects, showing that the algorithm's runtime on BinVal is at most a factor O(log μ · log n) slower than on the simpler OneMax function. The result holds for standard bit mutation and several common mutation operators, confirming that the (μ+1) EA is surprisingly efficient even on deceptive problems. This work narrows the gap between theory and practice, providing tighter guarantees for an algorithm widely used in optimization.