py -m rl.export vyexportuje checkpoint do rl/weights/neural-bot.json
(bit-exact float32, 1.3 MB) a rl/pure_net.py ho hra bez torch/numpy
(stdlib forward pass, ~16 ms/tah). Testy parity: logity aj akcie sa
zhoduju s torch, identicke trajektorie celych kol. Natrenovany model:
6.7-6.9 b/kolo proti vsetkym baseline-om (heuristika prekonana).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
BridzikNet (trup + guess/play/value hlavy s maskovanim), self-play
generator so zdielanou sietou na 4 sedadlach a opponent mixingom
(random/heuristicke sedadla pre robustnost), vlastny clipped-PPO
so skalovanim odmien a lr/entropy annealom. Torch je len trenovacia
zavislost na hoste (requirements-rl.txt); checkpointy a logy su
gitignorovane. Spustenie: py -m rl.train.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>