From Conway's Game of Life to Digital Darwinism: Building a GPU-Accelerated Neuroevolutionary Ecosystem

As I approach the second semester of my AI Master's program at the University of York, one elective, Evolutionary Intelligence, has particularly captured my interest. During this holiday period, I've taken the classic "Conway's Game of Life" to an entirely new dimension: a large-scale, neural network-driven, GPU-accelerated Artificial Life (ALife) evolutionary system.

Video Demo: YouTube

Core Architecture: When Neural Networks Have "Genes"

In my simulator, NeuroEvo-Life, digital organisms are no longer rigid pixels, but intelligent agents endowed with 12-dimensional genetic traits (12D Genome):

• Neural Fingerprint (8D): Statistical features extracted from neural network weights, determining the agent's decision-making style.
• Chemical Affinity (4D): Determines inter-species signal transmission, compatibility, and predatory relationships.

GPU Vectorization Optimization

Leveraging years of backend architecture experience, I realized that the bottleneck for large-scale simulations lies in computational efficiency.

• The system is built entirely upon PyTorch vectorized operations (such as einsum and convolution kernels), offloading the decisions of tens of thousands of agents and chemical field diffusion entirely to the GPU.
• This design allows us, even at extremely high population densities, to observe real-time species drift and evolutionary competition.

Evolutionary Results: 100% Ecological Dominance

By introducing a hybrid evolutionary paradigm (combining individual adaptation from reinforcement learning with population mutation from genetic algorithms), I observed a significant phenomenon of "Digital Darwinism".

After thousands of generations of optimization, one specific "super-lineage" demonstrated extreme environmental plasticity, occupying 100% of the ecological niche in competition with populations of randomly weighted agents—manifesting the absolute survival advantage that learned evolution confers.

Open Source & Archiving:

The project has been archived on Zenodo and received DOI certification: 10.5281/zenodo.18397035.

The source code has also been officially open-sourced: github.com/geyuxu/yuxus-life-of-game