A learned gene editor: can a model mutate NEAT better than chance?
Tests whether a trained edit scorer can beat random NEAT mutations, with fresh seeds and a protocol fixed before the first run.
Posts tagged “Reinforcement learning”
Lab notebook posts about “Reinforcement learning”.
Should a genome keep what its network learned? Three ways to mix gradients into evolution
Three hybrid trainers ask whether evolution should discard learned weights, inherit them, or pool experience in one gradient learner.
Learning to read the level: replacing lucky jumps with a policy gradient
A failed champion sets the test: can a policy read unseen Skyline Run levels instead of replaying one lucky route?
Procedural districts: making Skyline Run test the policy, not the route
A measured generator makes every Skyline Run district clearable, separates training from test levels, and exposes route memorization.
Can a deterministic game teach a network to run the rooftops?
A deterministic engine becomes a browser AI lab—and measurement catches the environment changing before training begins.
Teaching the trainer: evolving the reward instead of tuning it
Evolve better rewards automatically: AutoRL searches training settings; PBT, or population-based training, copies and perturbs the best runs.
The memorization trap: when a champion learned one district
The rebuilt Skyline AI lab learned under fixed budgets, but 0/630 generated-world clears show that broader training still did not generalize.