RL: self-play PPO trening siete

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>

rl/policy_player.py

@@ -0,0 +1,42 @@
+"""Natrenovana siet ako hrac so standardnym rozhranim guess/play.
+
+Rovnake rozhranie ako rl/players.py, takze funguje v rl/evaluate.py aj ako
+boti "mozog" v api/bots.py. Hrac vidi len observaciu + masku z rl/encoding.py
+-- z principu nemoze podvadzat (do cudzich ruk sa nedostane).
+"""
+
+import torch
+
+from rl.encoding import encode_observation, guess_mask, play_mask
+from rl.model import BridzikNet, mask_tensor, masked_categorical, obs_tensor
+
+
+class NeuralPlayer:
+    def __init__(self, net: BridzikNet, greedy: bool = True):
+        self.net = net
+        self.greedy = greedy  # argmax pri evaluacii; sampling pre pestrost
+
+    def _act(self, rnd, seat: int, use_play_head: bool) -> int:
+        obs = obs_tensor(encode_observation(rnd, seat)).unsqueeze(0)
+        mask = mask_tensor(
+            play_mask(rnd, seat) if use_play_head else guess_mask(rnd)
+        ).unsqueeze(0)
+        self.net.eval()
+        with torch.no_grad():
+            guess_logits, play_logits, _ = self.net(obs)
+            logits = play_logits if use_play_head else guess_logits
+            logits = logits.masked_fill(~mask, float('-inf'))
+            if self.greedy:
+                return int(logits.argmax(dim=-1).item())
+            return int(masked_categorical(logits, mask).sample().item())
+
+    def guess(self, rnd, seat: int) -> int:
+        return self._act(rnd, seat, use_play_head=False)
+
+    def play(self, rnd, seat: int) -> int:
+        return self._act(rnd, seat, use_play_head=True)
+
+
+def load_player(checkpoint_path: str, greedy: bool = True) -> NeuralPlayer:
+    from rl.train import load_checkpoint
+    return NeuralPlayer(load_checkpoint(checkpoint_path), greedy=greedy)