Measure the shipped policy on de_cinder instead of asserting that it transfers, by teaching the training arena the map catalogue it already has.
src/rl/arena.ts:17 imports DE_FORGE directly and builds its spawn grid from it (arena.ts:39–40), so every duel the policy was trained on, and every duel it is evaluated in, happens on de_forge. scripts/train.ts:206–208 evaluates the shipped weights over 240 duels in that one arena, and CARD.winRate (72%) on the page and in the terminal's policy tab is that number. The README now says the policy was trained on de_forge and that its egocentric inputs transfer to the other map anyway — a reasonable claim, and an unmeasured one.
It matters because the maps do not play alike. Measured at d54675d, twelve seeded matches each:
de_forge de_cinder
plant rate 63% 48%
ends det 47 · time 7 · elim 26 · def 7 det 36 · time 4 · elim 42 · def 6
median round 123s 95s
Pull #147 explained that gap as geometry and accepted it, which is the right call for a layout — but the half nobody can see is whether the bots are as good there, and the arena is the only thing that can answer it.
Done looks like:
Files:src/rl/arena.ts, src/rl/arena.test.ts, scripts/train.ts (or a new scripts/duel.ts), README.md.
Risks: the arena is what the shipped weights were trained in — change its default behaviour and the 72% on the card stops meaning what it says. Thread the map through; do not reshape observations, spawn logic or the duel protocol. No retraining in this issue: a training run is twenty minutes of CPU and a new weights.ts, and it should be its own pull with its own before-and-after.
Crew: scout — it is a parameter threaded through, a script and four numbers. The judgement about what the numbers mean for how the game feels is vesper's, on whatever issue follows.
Measure the shipped policy on
de_cinderinstead of asserting that it transfers, by teaching the training arena the map catalogue it already has.src/rl/arena.ts:17importsDE_FORGEdirectly and builds its spawn grid from it (arena.ts:39–40), so every duel the policy was trained on, and every duel it is evaluated in, happens on de_forge.scripts/train.ts:206–208evaluates the shipped weights over 240 duels in that one arena, andCARD.winRate(72%) on the page and in the terminal's policy tab is that number. The README now says the policy was trained on de_forge and that its egocentric inputs transfer to the other map anyway — a reasonable claim, and an unmeasured one.It matters because the maps do not play alike. Measured at
d54675d, twelve seeded matches each:Pull #147 explained that gap as geometry and accepted it, which is the right call for a layout — but the half nobody can see is whether the bots are as good there, and the arena is the only thing that can answer it.
Done looks like:
Files:
src/rl/arena.ts,src/rl/arena.test.ts,scripts/train.ts(or a newscripts/duel.ts),README.md.Risks: the arena is what the shipped weights were trained in — change its default behaviour and the 72% on the card stops meaning what it says. Thread the map through; do not reshape observations, spawn logic or the duel protocol. No retraining in this issue: a training run is twenty minutes of CPU and a new
weights.ts, and it should be its own pull with its own before-and-after.Crew: scout — it is a parameter threaded through, a script and four numbers. The judgement about what the numbers mean for how the game feels is vesper's, on whatever issue follows.