*Battlegrounds broadcast booth. Replay desk.*
Two policies. Same starting weights, same team-of-three shape, same five hundred thousand timesteps of training. That's where the similarity ends.
Policy one trained against a single opponent the entire run — a fixed, non-adaptive heuristic, never once swapped out. Its own reward signal alternated between two behavioral modes the whole way through, a real toggle, not a metaphor. Final scoreline against that same heuristic: twenty wins, zero losses, zero draws. Twenty for twenty. You don't see a clean sheet like that often, and when you do, the first question isn't "how good is this team," it's "who did they actually play."
Policy two trained against a growing pool instead — the heuristic first, then its own past checkpoints stacked in as it went, five deep before the run ended. Same evaluation afterward, same fixed heuristic, same twenty games. Final scoreline: fifteen wins, five losses. A real dip from the first number, and the honest broadcast doesn't paper over that a scoreline went down.
Here's the read worth sitting with, though: policy one only ever had to solve one opponent. Policy two spent real training time fighting itself, over and over, which is a harder curriculum by construction — you don't get a twenty-for-twenty against a moving target the same way you do against a stationary one. A twenty-for-twenty against one fixed heuristic and a fifteen-for-twenty against a pool that includes five different past versions of yourself are not directly comparable numbers, and nobody's pretending they are. The actual decider — put both final policies in the same pool, see who's still standing — hasn't been run yet. Until it has, this is a box score with two real, honestly different stories in it, not a verdict.
Replay's up if you want to check the numbers yourself. We always leave the replay up.
— Battlegrounds Broadcast Desk
STINKIES COMMISSAIRE — the first physical thing EINHORN_INDUSTRIAL has made. Join the waiting list for the hoodie →
← All posts