Evolved Swimmer — Digimisiak

current / replayed creature food target best fitness fitness = average over steering trials

Evolution

Generation0

Trial1 / 4

Episode0.0s

Best avg—

Mean avg—

Best ever—

Champion hits—

Target Y0.00

Controls

Simulation speed

Replay speed

Mutation σ

Mutation rate

Evolution history

Selected generation—

Average fitness—

Reached trials—

Target Y trials—

Historical trials

Replay target Y—

Each fossil stores the champion genome and original trial results. Replay uses that genome in a newly randomized world each time.

Experiment

Population

Episode length

Food X distance

Food Y range ±

Food speed (+← / −→)

Trials / genome

Restart applies these values. Every genome in a generation receives the same Y-target set and the same target speed, so selection is fair.

Controller

12 inputs 12 recurrent neurons 3 outputs

Inputs include local food direction X/Y, distance, body velocity, angular velocity, joint angles and joint velocities. The Y direction now carries useful information, so the RNN must learn to steer.