Overview

Three colonies of ants learn to forage using different optimization methods. Each ant is controlled by a tiny neural network. We compare evolutionary approaches (GA and PSO) with traditional gradient-based learning (Backpropagation).

Neural "Brain"

• Inputs (5): food angle (relative), normalized distance, velocity X, velocity Y, bias (1.0)
• Hidden layer: 4 neurons (tanh)
• Outputs (2): turn angle (−π/4…π/4), speed (0…max)

Episode and Fitness

• One generation length is adjustable (default 600 frames). All ants act simultaneously.
• Fitness increases by +1 whenever an ant collects a food item.
• At generation end, algorithms produce the next set of brains; ants respawn.

Genetic Algorithm (GA)

• Selection: Tournament of random candidates; pick the best.
• Crossover: Single-point on flattened weight vectors.
• Mutation: Adds Gaussian noise; elite (best) always survives.

Particle Swarm Optimization (PSO)

• Each ant = particle with position (weights) and velocity.
• Updates use inertia + pull toward personal and global best weights.
• Velocities and weights are clamped to keep the search stable.

Backpropagation (BP)

• Traditional gradient descent on neural network weights.
• Uses REINFORCE-style policy gradient: rewards good actions, penalizes bad ones.
• Learns from experience within each generation, adjusting weights based on fitness improvements.

Tips

• Increase inertia to explore; increase cognitive/social to exploit.
• Mutation rate affects GA diversity; too high can cause churn.
• Learning rate controls BP's step size; too high can destabilize training.
• Click the canvas to add food and make learning signals denser.