// Academic Project · Year 4
Coevolutionary Boids: Predators and Prey on the GPU
Evolved 1,000 flocking prey against 50 predators on the GPU, and found prey spread out rather than bunch up.
A group project for Bio-Inspired Computation. We built a 3D flocking simulation on the GPU where 1,000 prey and 50 predators evolve against each other, to see how simple flocking rules adapt when something is hunting the flock. I wrote most of the code. The prey didn't do what theory predicted.
The question
- Module
- COMP5400 Bio-Inspired Computation
- Year
- 2025/26 · Year 4
- Team
- 3 students
- Weighting
- 60% of the module
Reynolds' boids model (1987) gets lifelike flocking from three local rules: keep your distance, match your neighbours' heading, and move towards the group. The weights on those rules are normally fixed. Real prey adapt to their predators, so we asked: what happens to those rules when they evolve under a predator that is evolving too?
How it works
Evolution
- • Prey evolve their three flocking weights and how far they can see.
- • Predators evolve how sharply they can turn and how far they can see. They are 1.5 times faster than the prey, so the prey can't simply outrun them.
- • Prey that survive breed. Predators are picked by kill count in tournaments of three.
- • Children mix their parents' traits, with an 8% chance of a random mutation per trait.
- • 20 generations without predators set a baseline, then 500 generations of predators and prey evolving together.
Engineering
- • Physics in NumPy, with a CuPy path that computes every pairwise interaction on the GPU. It ran at about 60 ticks a second with 1,050 agents on an RTX 5090.
- • On the CPU, a spatial grid limits each boid's neighbour search.
- • 3D rendering with ModernGL and GLSL shaders, with a 2D Pygame fallback and a sidebar for tuning live.
- • A headless mode for the long runs, saving a checkpoint every generation and producing all the plots automatically.
What evolved
When the predators arrived, prey survival crashed to about 30% by generation 50. The prey recovered to around 48% by generation 100 and then held between 45% and 50% for the remaining 400 generations, while both sides kept evolving. Neither side pulled ahead. That fits the Red Queen hypothesis: species have to keep adapting just to stay where they are.
The surprise was how the prey survived. Hamilton's selfish herd theory predicts prey should pack tighter, each trying to get to the middle. Ours spread out instead. The pull towards the group fell from 1.0 to 0.35, alignment fell from 1.0 to 0.5, and the average distance to a prey's nearest neighbour grew from 52 to 80 units. At the same time, how far prey could see more than doubled, from 170 to about 390, and then stopped growing: past that point, seeing further gave no extra protection. A dense flock gave fast predators lots of targets. A loose, less predictable flock that sees them coming early left each predator pass with at most one kill.
The predators' vision told a story too. It grew quickly at first, from 300 to around 800, then shrank back towards 400 once the flock was too sparse to be worth tracking from far away.
