New AI Design Method Is 34 Times Faster — Here's Why It Matters

A new paper tackles a key bottleneck in neuroevolution by rethinking how indirect encoding explores network topologies. Instead of recursively subdividing space with a quadtree—a process that resists parallelization—the proposed approach, EMR-HyperNEAT, evaluates all positions at all resolutions up front and then filters them with the same variance criterion. This makes every query independent and parallelizable across cores and population members, reducing complexity from O(4^D) to O(4^D/P). The paper reports 12–34× on-device GPU speedups on XOR at depths 5–7 and empirically higher solve rates across benchmarks. It also introduces a connection type taxonomy that makes recurrent substrate configurations feasible.