One obstacle course, different ways to learn: genetic search, gradient descent, hybrid training and evolving brains.
11 labs. Start with Hybrid Neuroevolution.
Watch two ways of learning race on the same seeded obstacle course: one lane evolves runners by pure genetic search, the other alternates genetic splits with online backpropagation. Live network diagrams show what each brain is doing.
Watch birds learn to jump, duck and sidestep obstacles and platforms. Their networks evolve layer counts and sizes along with the weights; eight lineage splits per generation, each logged.
Race a genetic population against one bird trained by online backpropagation on the same seeded obstacle course, side by side, with stats and an event log.
Neural runners learn to jump and dive between two worlds. Tune the population, mutation, and network.
Evolutionary epigenetic runner AI with real-time fitness visualization.
Evolve a population of runners and track their fitness across generations.
Pits a genetic-algorithm population against a backprop-with-replay agent in identical seeded obstacle courses, accepting or reverting gradient updates by outcome.
Evolves a population of jumpers whose policy is a mixture of three experts behind a softmax router; watch which expert is routing, change speed or hyper-train.
A compact neural runner with editable population and mutation. Add spikes, adjust the brain, and watch each generation learn.
One hundred two-hidden-layer agents evolve to cross platforms with a triple jump; weights and connections mutate each generation. Set speed, watch the high score, reset.
100 agents with 10-hidden-neuron nets evolve to jump obstacles through weight and structural mutation; run up to 1000x with a live network graph. Offline.
