IRIN O2 — Neuroevolution Lab WEB EDITION

💡 The big idea in plain words

This page runs evolution in your browser. Each candidate is a tiny neural "brain" — a list of numbers, the weights. Random brains drive the robot; the best-scoring ones breed (their weights mix like genes), their children get small random mutations, and the elite survive unchanged. Generation after generation the brains get better at the task. The fitness function is the score evolution optimises: go fast, go straight, don't hit walls.

🎓 The Project

IRIN O2 (UPM, 2019) — a Webots e-puck whose neural controller is evolved by a genetic algorithm instead of learned by gradient descent. The lab evolves layered distributed NN controllers; this page reproduces the pipeline in miniature: an 8-4-2 MLP (proximity rays → hidden → wheel speeds) in a 2D arena, evolved live with the same fitness function.

🧬 Genetic algorithm

Each genome is the weight vector of the controller. The population of \(P\) individuals is evaluated, then a new generation is built by

• Tournament selection — pick \(k\) genomes at random, keep the fittest, repeat to fill the mating pool;
• Crossover — recombine two parents' weight vectors;
• Gaussian mutation — perturb weights with \(\mathcal{N}(0, \sigma^2)\); \(\sigma\) anneals each generation (0.996^gen) to refine late in the run;
• Elitism — the best individuals survive unchanged.

🎯 Fitness (Nolfi & Floreano)

Each genome drives the robot for 120 steps; the per-step fitness rewards fast, straight, obstacle-avoiding motion:

\[ f = V\,(1 - \sqrt{\Delta v})\,(1 - i) \]

where \(V\) is the translational speed, \(\Delta v\) the rotational difference between wheels (penalizing turns) and \(i \in [0,1]\) the strongest proximity reading (penalizing walls). The chart plots the best fitness per generation; with the default settings the population reaches the task ceiling (\(f \approx 0.53\)) within ~100 generations.