Evolutionary Steering

How Natural Selection Works

This simulation implements an agent-based model of Darwinian evolution. Agents (organisms) move through an environment containing food (green) and poison (red). Each agent has a neural-network-style genome encoding how it weights different types of food in its field of view. Agents that eat more food survive longer, reproduce, and pass on their genome — sometimes with random mutations.

Over successive generations, the population becomes adapted to its environment. You can observe the emergence of feeding strategies, avoidance behaviours, and even competitive dynamics — all from simple fitness rules and variation.

Genetics and Inheritance

Each agent carries a genome — a set of numerical parameters controlling its behaviour when it senses nearby food or poison. When an agent reproduces, its offspring inherits its genome with a small random mutation: a slight tweak to one or more parameters.

Key genetic concepts illustrated:

• Variation — mutations create diversity within the population
• Heritability — offspring resemble parents
• Selection pressure — environmental conditions favour certain traits
• Fitness landscape — not every mutation is beneficial

Evolution in Practice

Real-world applications of evolutionary computation include:

• Genetic algorithms — optimising engineering designs, scheduling, and routing
• Neural architecture search — evolving the structure of AI models
• Drug discovery — evolving molecules with desired pharmaceutical properties
• Robot locomotion — evolving walking gaits for physical robots
• Game AI — evolving strategies for complex board games

The field of evolutionary computation (EC) spans genetic algorithms, genetic programming, and evolution strategies — all inspired directly by the biological processes modelled here.

Experiments to Try

• Increase poison density — watch how quickly agents evolve avoidance strategies
• Crank up the mutation rate — observe rapid environmental adaptation vs. genetic instability
• Reduce food — create strong selection pressure and watch population fitness jump
• Watch the fittest agents — identify the behavioural patterns that emerge over time
• Reset and compare — run multiple experiments to see how evolution converges differently each time