A Fresh Look at Lamarckian Evolution and the Baldwin Effect

Researchers demonstrate that Baldwinian and Lamarckian evolutionary algorithms significantly outperform traditional Darwinian evolution on complex optimization problems like Maximum Independent Set and Maximum Cut. The study provides both empirical validation across multiple datasets and theoretical runtime analysis, showing that local search-augmented evolutionary algorithms offer practical advantages for solving NP-hard graph problems.

This research addresses a long-overlooked gap in evolutionary algorithm literature by rigorously comparing inheritance mechanisms that have remained peripheral to mainstream academic discussion. The work moves beyond theoretical speculation by combining empirical benchmarking on real GraphBench datasets with formal runtime analysis, establishing that Baldwinian evolution (where acquired traits improve fitness but aren't inherited) consistently outperforms pure Darwinian approaches. This matters because it validates decades-old evolutionary concepts that computer scientists largely abandoned, suggesting the field prematurely discounted mechanisms that could improve algorithm performance.

The findings have substantial implications for optimization research and practical applications. Graph problems like Maximum Independent Set and Maximum Cut appear in network analysis, circuit design, and computational biology—domains where current solutions often rely on expensive specialized solvers. The research demonstrates that well-tuned evolutionary algorithms can bridge the gap between general-purpose methods and domain-specific heuristics, while outpacing contemporary deep learning baselines. This is particularly significant because deep learning dominates modern algorithm research; showing EAs with local search mechanisms remain competitive challenges prevailing assumptions about which paradigms should drive future optimization research.

The theoretical contributions extend the Deceptive Leading Block benchmark to arbitrary block lengths, providing formal proof that Baldwinian evolution achieves better asymptotic runtime than Lamarckian, which beats Darwinian evolution. The ordering remains consistent even when accounting for local search costs, explaining why Baldwinian evolution achieved strongest empirical results. For practitioners, the authors provide generalist parameters across evolution types, reducing hyperparameter tuning burden. This work suggests the optimization community should revisit hybrid evolutionary-local-search approaches as a legitimate research direction alongside neural and symbolic methods.

• →Baldwinian and Lamarckian evolution consistently outperform standard Darwinian evolution on NP-hard graph optimization problems
• →Local search-augmented evolutionary algorithms match or exceed performance of recent deep learning baselines on Maximum Independent Set and Maximum Cut benchmarks
• →Theoretical analysis proves Baldwinian evolution achieves superior asymptotic runtime compared to Lamarckian and Darwinian approaches
• →Published generalist hyperparameters enable practitioners to apply these evolutionary approaches without extensive tuning
• →Findings challenge the field's decades-long dismissal of inheritance mechanisms in evolutionary computing