Laser Evade — Evolution Lab

Simulation

Width Height Agent radius Speed Max concurrent lasers Max concurrent (late cap) Spawn chance / step (early) Spawn chance / step (late) Warning (steps, early) Warning (steps, late) Beam alive (early) Beam alive (late) Beam max length Laser thickness Dodge score weight Rays (sensors) Max ray length Max steps / episode

Fitness shaping

Training only: rewards open space, dodging, and moving away from beams in the hazard band; penalizes wall camping, lingering in that band, and dying. The playback HUD uses the same total.

Center bonus weight Wall hug penalty / step Hazard margin (arena) Hazard proximity penalty Hazard escape bonus Hazard exit bonus Death penalty

Wall spikes

Stay near a wall for too many steps in a row and lethal spikes jut inward (same randomness as training).

Enable wall spikes Hug margin Hug steps → spikes Spike depth Spike thickness Spikes count min Spikes count max Spike alive (steps) Spike cooldown (steps)

Evolution

Everyone is scored each generation. How the next batch is built depends on the mode below.

Battle royale — highest-scoring agent teaches the next generation (elite slots are exact copies, the rest are mutated copies). Classic GA — turn this off for tournament parents and crossover.

Generations Population Hidden neurons Elite Tournament k Crossover Mutation rate Mutation σ Eval workers GA seed Arena seed Parallel fitness (multi-core)

Parallel evaluation scores the whole population each generation. Raise population or elite for a stronger “teacher” signal in battle royale.

Arena

Bright dot = round winner (highest fitness). Dim dots = everyone else in the same generation, each in a parallel episode. Beams/spikes follow the winner’s world. The view advances only after all dots finish; the server then builds the next generation from that winner (battle royale) or classic GA if you disabled it.

Generation — High score (evolution) — This run — This gen — —

Generations

Each dot is one generation (darker = lower best fitness). Click a dot or use the slider on the right.

Generation log

Per-generation deaths (lethal beam vs spike vs survived max steps), mean scores, and what the population struggled with. On this static copy (e.g. GitHub Pages), evolution runs in your browser; the Python server log file is not used.