A Fresh Look at Lamarckian Evolution and the Baldwin Effect
Inès Benito, Johannes F. Lutzeyer, Benjamin Doerr
Read on arXiv →Key claim
Baldwinian and Lamarckian evolution outperform Darwinian evolution.
This paper revisits Baldwinian and Lamarckian evolution in evolutionary algorithms, showing they outperform traditional Darwinian methods in various scenarios. The authors provide a set of generalist parameters that can benefit practitioners, highlighting the practical implications of their findings.
In plain English
This paper revisits Baldwinian and Lamarckian evolution in evolutionary algorithms, showing they outperform traditional Darwinian methods in various scenarios. The authors provide a set of generalist parameters that can benefit practitioners, highlighting the practical implications of their findings.
The paper provides a significant new comparison of evolutionary algorithms, extending existing benchmarks and demonstrating superior performance.
The empirical results are based on comprehensive experiments across multiple datasets, supporting the claims made.
Deep reliability assessment
The methodology supports the claim that, on GraphBench random-graph MIS and Max-Cut instances under a fixed 40,000-fitness-evaluation budget, Baldwinian/Lamarckian local-search-augmented EAs outperform Darwinian EAs and the listed neural baselines. Broader claims about practical dominance or solver-level performance are less supported because evaluation is limited to two graph optimization tasks, mostly synthetic/random graph families, and comparisons depend strongly on tuning budget, implementation cost of local search, and the chosen specialized solvers.
Reproducibility
Yes: the paper states that complete source code is available at https://github.com/Hypatia-II/polerina, and experiments use the public GraphBench benchmark with described train/test splits, parameter grid search, and fixed evaluation budget.
Discussion questions
- 1.Does the apparent advantage of Baldwinian evolution come from a genuinely better evolutionary search dynamic, or mainly from giving the algorithm access to a strong local-search oracle during fitness evaluation?
- 2.For builders solving routing, scheduling, or graph optimization problems, when is a memetic EA preferable to directly running a specialized heuristic or exact solver with the same wall-clock budget?
- 3.What result would falsify the paper’s main takeaway: poor performance on non-random real-world graph distributions, loss of advantage under equal wall-clock time instead of equal fitness evaluations, or failure to beat well-tuned domain heuristics?
Key figure
No Figure 1 or key architectural diagram is included in the provided excerpt; the central setup is a comparison of Darwinian, Baldwinian, Lamarckian, and mixed L-B evolutionary algorithms with or without local-search-based offspring evaluation/replacement.