Skyline Run AI Trainer AI
An outer NEAT population controls the reward coefficients used by an inner NEAT trainer. Each controller is scored on real level progress, with the hand-tuned rewards rerun as the control.
Play the game: Skyline Run Open the agent AI labMeta-training setup
Evaluation horizon per trainer candidate — not the runner's training length (the runner trains open-ended in the AI lab). Runs extend 25% past the horizon, and higher values multiply meta-training time.
How to read this dashboard
- An outer NEAT population evolves controllers that tune the game's reward multipliers. Each candidate is scored by running a full inner training session; candidate 1 of every generation runs the hand-tuned rewards as the baseline.
- Fitness is the tail-weighted average of inner-runner level progress — later generations count more — measured 25% beyond the configured horizon, plus a completion bonus. The course seed rotates every meta generation, so raw fitness is only comparable within a generation — watch the advantage over the baseline instead.
- Going well looks like: the advantage chart above zero, best level reach climbing toward 100%, and the "Level cleared" badge appearing.
- The reward multipliers chart shows the knobs the controller is turning — progress, checkpoint, and completion rewards plus the time and death penalties — relative to the 1× dashed line.
Meta generation 0
Advantage vs baseline — Champion — vs baseline —
Evaluating candidate — Candidate 1 is the hand-tuned baseline
Inner generation 0/50
Best level reach — How far the current candidate's runners get
Simulation speed 0 ticks/s
Metrics appear once training starts.
Metrics appear once training starts.
Metrics appear once training starts.
Metrics appear once training starts.
Metrics appear once training starts.