Predator–Prey Simulator

Overview

This project is a C++ predator–prey simulation where agents are controlled by neural networks trained through evolutionary methods. Predators and prey interact in a simulated environment, and behaviors emerge through selection and mutation over generations.

The project is inspired by the final project of a course I took at IMSA called Computational Science, where the entire class wrote deterministic or MLP-based policies for lynx and hare and competed in a “Hunger Games” style arena to see whose hare would survive the longest.

How to Build & Run

cd build && cmake .. && cmake --build . && cd .. && ./build/PredatorPreySim

Output is saved to logs/simulation_<timestamp>.csv. Visualize with:

python3 scripts/visualize.py

Architecture

• Vision: Each agent casts 8 rays (full 360° FOV, range 10 units). Each ray returns normalized distance and entity type (predator/prey/nothing).
• Brain: MLP (18 → 16 → 12 → 2) takes ray inputs + normalized velocity, outputs velocity delta with tanh squashing. Weights initialized with He init.
• Spatial Grid: Uniform 10×10 cell grid for O(1) neighbor lookups used by both raycasting and collision detection.
• Evolution: Agents with higher fitness (predators that eat more, prey that survive) reproduce at higher rates each generation.

Status

Work in progress

To-Do

• Raycasting for agent "vision"
• Spatial partitioning for collisions/rays
• Neural Network for agent "brain"
• Evolutionary weight inheritance (crossover + mutation across generations)
• Visualization of agent trajectories and neural activations