Noddybear/18xx-rl

18xx RL checkpoints

1846

Native PufferLib flat-f32 checkpoints (obs 832). Load with 18xx_training tools/bin_to_pt.py (--hidden H --layers 2 --obs-n 832).

• 1846/h2048_final.bin — current champion (h2048 l2). Net2Wider era on Modal, 2026-09-12: 30 legs (last 10 on unique deals via seed_base); direct playoff vs league1_gen30 0.691 mean decided-WR over 1,800 games (0.707/0.665/0.702 per seed, 0% DNF); final gauntlet 11/11.
• 1846/league1_gen30.bin — h1024 league-era champion, 2026-09-09: 30 legs, 27 promotions; playoff vs the prior gen-22 champion 0.591 mean decided-WR over 1,800 games (0% DNF).

1870

• 1870/h2048_gen1_league.bin — current champion (h2048 l2, obs 1152, act 192, native PufferLib flat-f32 format). Load with 18xx_training tools/bin_to_pt.py (--hidden 2048 --layers 2 --obs-n 1152).

  Lineage: SFT h1024 (test top1 64%) -> RL warm-start -> Net2Wider h2048 staged capacity ladder (gen-14) -> FULL AlphaStar league (MAIN vs whole pool of ladder gens + main-exploiters, PFSP, floor-raising promotion). The league gen-1 leg (600M steps vs the whole pool) beats every ladder ancestor AND its own hard-counters, with pool mean win-rate ~0.68.

  EXPLOITER-ROBUST: the earlier staged-ladder gen-14 looked exploitable — a dedicated exploiter found a 0.958 hard-counter — which suggested 1870 might be an intrinsically cyclic game. It is not. That was catastrophic forgetting from the ladder's too-narrow opponent window (gen-14 even LOST to its own earlier ancestors g003/g005 at ~0.15). A single deep league best-response vs the full pool (this gen-1) flipped those 0.95 counters to ~0.82 wins, and FOUR consecutive fresh 200Mx2 exploiters then failed to counter it (best 0.477, 0.506, 0.490, 0.489 — all < the 0.60 join bar). Residual: g004 and g005 are a mutually-exclusive ancestor pair no single deterministic net beats simultaneously (min ~0.28), a mild 2-strategy non-transitivity, not exploitability. Supersedes the removed gen-14 checkpoint.