NSGA-III crossover boosts multi-objective optimization speed: New runtime analysis

A new theoretical analysis from researcher Andre Opris, accepted at IJCAI 2026, provides the first rigorous runtime comparison of NSGA-III with and without crossover on many-objective optimization problems. The study focuses on the classical m-objective OneJumpZeroJump (m-OJZJ) function, a standard benchmark for multi-objective evolutionary algorithms (MOEAs). The results prove that NSGA-III with crossover solves m-OJZJ asymptotically faster than its crossover-free variant for any number of objectives m, across a wide parameter regime. This holds even as the number of objectives grows, addressing a long-standing gap between theoretical understanding and practical success of crossover in many-objective optimization.

To solidify the finding, Opris also provides a lower runtime bound for the crossover-free NSGA-III on the 4-objective OJZJ function, confirming that the speedup is fundamental and not an artifact of specific parameter choices. The work is significant because while crossover is widely used in practice to improve MOEA performance, theoretical guarantees were limited to artificially designed benchmarks. This analysis offers the first general proof of crossover's benefit in the many-objective setting, helping algorithm designers decide when to include crossover operators. The paper is available on arXiv as arXiv:2605.11201.