What is training
Algorithm
Reward and league
Progress
Last hits per 5-minute game
training (self-play) eval vs passive eval vs unfair clone (4)
Outcomes per wave (live side's view)
wins draws losses · letter under the bar: s self, p pool, v Valve
Drift and exploration
KL to anchor entropy
Throughput
steps per wave (k) decision ms update s
Fleet
Evaluations
| it | vs passive: last hits · level · win · draw | vs unfair: last hits · level · win · draw | ||||||
|---|---|---|---|---|---|---|---|---|
Iterations (newest first)
| it | opponent | steps | trajectories | fort / cap | W–D–L | last hits | return | loss π | loss V | entropy | KL anchor | clip | speed | decision ms | update s | Lua t/o |
|---|