L-System Genetic Encoding for Scalable Neural Network Evolution: A Comparison with Direct Matrix Encoding

Researchers Alexander Stuy and Nodin Weddington have introduced Lsys, a novel L-System-based genetic encoding for evolving neural networks, and a modeling tool called Wp1hgn. In a series of 24 experimental runs using an artificial world with barriers and food, Lsys encoding achieved a mean maximum food count of 3,802 ± 197 at generation 1,000, compared to just 1,388 ± 610 for traditional direct matrix encoding—a 2.74x performance advantage. The consistency improvement was even more dramatic, with Lsys showing a coefficient of variation of 5.2% versus 44.0% for matrix encoding, an 8.5-fold improvement. All 8 Lsys populations successfully learned to navigate the environment, while 4 of 8 matrix populations failed to achieve competitive performance at any point.

When tested on a novel maze environment, Lsys populations demonstrated robust generalization, achieving a mean maximum food count of 2,455 ± 176 compared to 422 ± 212 for matrix populations—a 5.82x advantage that actually exceeded the training world performance gap. A control condition (MatrixLSG) showed that Lsys's advantage comes from operating on the compressed symbolic alphabet throughout evolution, not just from initial population structure. The paper, available on arXiv (2604.22000), demonstrates that Lsys encoding provides faster convergence, higher peak performance, dramatically greater reliability, and superior generalization across all conditions tested.