Deep CFR legacy 재현 config 기반 추가

configs/deep_cfr/pure_self_play_zero_pit_poc_full_depth_slot_aware_playability.yaml

@@ -0,0 +1,89 @@
+run:
+  experiment_name: lost_cities_deep_cfr_pure_self_play_zero_pit_poc_full_depth_slot_aware_playability
+  seed: 79
+  max_iterations: null
+  max_hours: 4
+  device: cuda
+  use_amp: false
+
+rules:
+  n_colors: 5
+  n_ranks: 9
+  min_rank: 2
+  n_handshakes: 3
+  hand_size: 8
+  expedition_penalty: -20
+  bonus_threshold: 8
+  bonus_amount: 20
+
+encoding:
+  derived_playability: true
+  slot_aware_playability: true
+
+network:
+  hidden_size: 256
+  num_layers: 3
+  activation: relu
+
+traversal:
+  traversals_per_player: 70
+  strategy_sample_interval: 1
+  store_strategy_on_opponent_nodes: false
+  store_strategy_on_traverser_nodes: true
+  max_depth: null
+  max_nodes_per_traversal: 1000
+  opponent_policy: self_play_league
+  cutoff_value_mode: score_diff
+  cutoff_rollouts: 0
+  cutoff_rollout_policy: random
+  cutoff_rollout_max_steps: 300
+  num_workers: 8
+  traversal_worker_chunk_size: 8
+  regret_matching_epsilon: 0.0001
+  outcome_sampling_epsilon: 0.2
+  outcome_sampling_value_clip: 500
+  outcome_unsampled_regret: zero
+  endpoint_depth_bucket_width: 100
+  endpoint_depth_bucket_max: 1000
+
+self_play:
+  current_weight: 0.5
+  recent_weight: 0.3
+  older_weight: 0.2
+  anchor_weight: 0.0
+  recent_window: 5
+  max_snapshots: 20
+  snapshot_every: 1
+
+optimization:
+  advantage_batch_size: 1024
+  strategy_batch_size: 1024
+  advantage_updates_per_iteration: 256
+  strategy_updates_per_iteration: 256
+  learning_rate: 0.00003
+  weight_decay: 0.0001
+  grad_clip: 1.0
+
+memory:
+  advantage_capacity: 2000000
+  strategy_capacity: 2000000
+
+evaluation:
+  eval_every: 5
+  games: 100
+  max_steps: 1000
+  on_max_steps: score_diff
+  opponents:
+    - random
+    - passive_discard
+    - safe_heuristic
+    - safe_heuristic_loose
+    - safe_heuristic_strict
+    - noisy_safe
+
+checkpoint:
+  directory: checkpoints/lost_cities_deep_cfr_pure_self_play_zero_pit_poc_full_depth_slot_aware_playability
+  save_every_iteration: false
+  save_iteration_interval: 10
+  save_latest_only: false
+  progress_interval_seconds: 20.0

src/coolrl_lost_cities/games/classic/deep_cfr/benchmark.py

@@ -4,7 +4,6 @@ import time
 
 from coolrl_lost_cities.games.classic.deep_cfr.config import DeepCFRConfig
 from coolrl_lost_cities.games.classic.deep_cfr.trainer import DeepCFRTrainer
-from coolrl_lost_cities.games.classic.game import classic_config
 
 
 def benchmark_traversal(
@@ -24,7 +23,7 @@ def benchmark_traversal(
     data["checkpoint"]["save_every_iteration"] = False
     data["evaluation"]["eval_every"] = 0
     cfg = DeepCFRConfig.model_validate(data)
-    trainer = DeepCFRTrainer(cfg, classic_config(seed=cfg.run.seed))
+    trainer = DeepCFRTrainer(cfg, cfg.rules.to_lost_cities_config(seed=cfg.run.seed))
     started = time.perf_counter()
     metrics = trainer.run_iteration(1)
     elapsed = time.perf_counter() - started

src/coolrl_lost_cities/games/classic/deep_cfr/cli.py

@@ -65,7 +65,7 @@ def train_command(args: argparse.Namespace) -> None:
     config = _with_overrides(config, overrides)
     trainer = DeepCFRTrainer(
         config,
-        classic_config(seed=config.run.seed),
+        config.rules.to_lost_cities_config(seed=config.run.seed),
         device=args.device or config.run.device,
     )
     if args.resume:

src/coolrl_lost_cities/games/classic/deep_cfr/config.py

@@ -9,25 +9,86 @@ from typing import Any
 import yaml
 from pydantic import BaseModel, ConfigDict, Field, field_validator
 
+from coolrl_lost_cities.games.classic.game import LostCitiesConfig
+
 
 class StrictModel(BaseModel):
     model_config = ConfigDict(extra="forbid")
 
 
 class RunConfig(StrictModel):
+    experiment_name: str = "deep_cfr"
     iterations: int = 1
+    max_iterations: int | None = None
+    max_hours: float | None = None
     seed: int = 1
     device: str = "cpu"
+    use_amp: bool = False
+
+    @field_validator("device")
+    @classmethod
+    def _normalize_device(cls, value: str) -> str:
+        token = value.strip().lower()
+        if token == "cuda":
+            return "cuda"
+        if token == "cpu":
+            return "cpu"
+        if token == "auto":
+            return "auto"
+        return token
+
+
+class RulesConfig(StrictModel):
+    n_colors: int = 5
+    n_ranks: int = 9
+    min_rank: int = 2
+    n_handshakes: int = 3
+    hand_size: int = 8
+    expedition_penalty: int = -20
+    bonus_threshold: int = 8
+    bonus_amount: int = 20
+
+    def to_lost_cities_config(self, seed: int | None = None) -> LostCitiesConfig:
+        config = LostCitiesConfig(
+            n_colors=self.n_colors,
+            n_ranks=self.n_ranks,
+            min_rank=self.min_rank,
+            n_handshakes=self.n_handshakes,
+            hand_size=self.hand_size,
+            expedition_penalty=self.expedition_penalty,
+            bonus_threshold=self.bonus_threshold,
+            bonus_amount=self.bonus_amount,
+            seed=seed,
+        )
+        config.validate()
+        return config
+
+
+class EncodingConfig(StrictModel):
+    derived_playability: bool = False
+    slot_aware_playability: bool = False
 
 
 class NetworkConfig(StrictModel):
     hidden_size: int = 64
+    num_layers: int = 2
+    activation: str = "relu"
+
+    @field_validator("activation")
+    @classmethod
+    def _validate_activation(cls, value: str) -> str:
+        token = value.strip().lower()
+        if token not in {"relu", "gelu"}:
+            raise ValueError("must be 'relu' or 'gelu'")
+        return token
 
 
 class TraversalConfig(StrictModel):
     traversals_per_iteration: int = 2
+    traversals_per_player: int | None = None
     max_depth: int | None = 8
     max_nodes: int | None = 10_000
+    max_nodes_per_traversal: int | None = None
     regret_matching_epsilon: float = 1.0e-8
     outcome_sampling_epsilon: float = 0.0
     outcome_sampling_value_clip: float | None = None
@@ -42,6 +103,9 @@ class TraversalConfig(StrictModel):
     store_strategy_on_opponent_nodes: bool = True
     num_workers: int | str = 0
     worker_chunk_size: int = 4
+    traversal_worker_chunk_size: int | None = None
+    endpoint_depth_bucket_width: int = 100
+    endpoint_depth_bucket_max: int = 1000
 
     @field_validator("outcome_unsampled_regret")
     @classmethod
@@ -80,6 +144,21 @@ class TraversalConfig(StrictModel):
             return max(0, int(token))
         return max(0, int(self.num_workers))
 
+    def resolved_traversals_per_player(self) -> int:
+        if self.traversals_per_player is not None:
+            return max(0, int(self.traversals_per_player))
+        return max(0, int(self.traversals_per_iteration))
+
+    def resolved_max_nodes(self) -> int | None:
+        if self.max_nodes_per_traversal is not None:
+            return self.max_nodes_per_traversal
+        return self.max_nodes
+
+    def resolved_worker_chunk_size(self) -> int:
+        if self.traversal_worker_chunk_size is not None:
+            return max(1, int(self.traversal_worker_chunk_size))
+        return max(1, int(self.worker_chunk_size))
+
 
 class SelfPlayLeagueConfig(StrictModel):
     snapshot_every: int = 1
@@ -96,7 +175,33 @@ class OptimizationConfig(StrictModel):
     advantage_train_steps: int = 1
     strategy_train_steps: int = 1
     batch_size: int = 32
+    advantage_batch_size: int | None = None
+    strategy_batch_size: int | None = None
+    advantage_updates_per_iteration: int | None = None
+    strategy_updates_per_iteration: int | None = None
     learning_rate: float = 1.0e-3
+    weight_decay: float = 0.0
+    grad_clip: float = 0.0
+
+    def resolved_advantage_batch_size(self) -> int:
+        if self.advantage_batch_size is not None:
+            return max(1, int(self.advantage_batch_size))
+        return max(1, int(self.batch_size))
+
+    def resolved_strategy_batch_size(self) -> int:
+        if self.strategy_batch_size is not None:
+            return max(1, int(self.strategy_batch_size))
+        return max(1, int(self.batch_size))
+
+    def resolved_advantage_train_steps(self) -> int:
+        if self.advantage_updates_per_iteration is not None:
+            return max(0, int(self.advantage_updates_per_iteration))
+        return max(0, int(self.advantage_train_steps))
+
+    def resolved_strategy_train_steps(self) -> int:
+        if self.strategy_updates_per_iteration is not None:
+            return max(0, int(self.strategy_updates_per_iteration))
+        return max(0, int(self.strategy_train_steps))
 
 
 class MemoryConfig(StrictModel):
@@ -107,6 +212,9 @@ class MemoryConfig(StrictModel):
 class CheckpointConfig(StrictModel):
     directory: str = "runs/deep_cfr/default"
     save_every_iteration: bool = True
+    save_iteration_interval: int = 0
+    save_latest_only: bool = False
+    progress_interval_seconds: float = 20.0
 
     @property
     def path(self) -> Path:
@@ -118,10 +226,21 @@ class EvaluationConfig(StrictModel):
     games: int = 10
     opponents: tuple[str, ...] = ("random",)
     max_steps: int = 10_000
+    on_max_steps: str = "score_diff"
+
+    @field_validator("on_max_steps")
+    @classmethod
+    def _validate_on_max_steps(cls, value: str) -> str:
+        token = value.strip().lower()
+        if token not in {"score_diff", "loss", "draw"}:
+            raise ValueError("must be 'score_diff', 'loss', or 'draw'")
+        return token
 
 
 class DeepCFRConfig(StrictModel):
     run: RunConfig = Field(default_factory=RunConfig)
+    rules: RulesConfig = Field(default_factory=RulesConfig)
+    encoding: EncodingConfig = Field(default_factory=EncodingConfig)
     network: NetworkConfig = Field(default_factory=NetworkConfig)
     traversal: TraversalConfig = Field(default_factory=TraversalConfig)
     self_play: SelfPlayLeagueConfig = Field(default_factory=SelfPlayLeagueConfig)

src/coolrl_lost_cities/games/classic/deep_cfr/evaluate.py

@@ -94,10 +94,10 @@ def load_strategy_policy_from_checkpoint(
     payload = torch.load(checkpoint_path, map_location="cpu")
     cfg = config_from_dict(payload["config"])
     game_config = LostCitiesConfig(**payload["game_config"])
-    network = DeepCFRMLP(
+    network = DeepCFRMLP.from_config(
         int(payload["input_dim"]),
         int(payload["action_size"]),
-        cfg.network.hidden_size,
+        cfg.network,
     ).to(device)
     network.load_state_dict(payload["strategy_network"])
     network.eval()

src/coolrl_lost_cities/games/classic/deep_cfr/networks.py

@@ -3,16 +3,46 @@ from __future__ import annotations
 import torch
 from torch import nn
 
+from coolrl_lost_cities.games.classic.deep_cfr.config import NetworkConfig
+
+
+def _activation(name: str) -> nn.Module:
+    token = name.lower()
+    if token == "relu":
+        return nn.ReLU()
+    if token == "gelu":
+        return nn.GELU()
+    raise ValueError(f"unsupported activation: {name!r}")
+
 
 class DeepCFRMLP(nn.Module):
-    def __init__(self, input_dim: int, output_dim: int, hidden_size: int = 64) -> None:
+    def __init__(
+        self,
+        input_dim: int,
+        output_dim: int,
+        hidden_size: int = 64,
+        *,
+        num_layers: int = 2,
+        activation: str = "relu",
+    ) -> None:
         super().__init__()
-        self.net = nn.Sequential(
-            nn.Linear(input_dim, hidden_size),
-            nn.ReLU(),
-            nn.Linear(hidden_size, hidden_size),
-            nn.ReLU(),
-            nn.Linear(hidden_size, output_dim),
+        layers: list[nn.Module] = []
+        last_dim = input_dim
+        for _ in range(max(0, int(num_layers))):
+            layers.append(nn.Linear(last_dim, hidden_size))
+            layers.append(_activation(activation))
+            last_dim = hidden_size
+        layers.append(nn.Linear(last_dim, output_dim))
+        self.net = nn.Sequential(*layers)
+
+    @classmethod
+    def from_config(cls, input_dim: int, output_dim: int, config: NetworkConfig) -> DeepCFRMLP:
+        return cls(
+            input_dim,
+            output_dim,
+            config.hidden_size,
+            num_layers=config.num_layers,
+            activation=config.activation,
         )
 
     def forward(self, x: torch.Tensor) -> torch.Tensor:

src/coolrl_lost_cities/games/classic/deep_cfr/trainer.py

@@ -28,6 +28,13 @@ from coolrl_lost_cities.games.classic.deep_cfr.workers import (
 from coolrl_lost_cities.games.classic.game import GameState, LostCitiesConfig
 
 
+def _resolve_torch_device(device: str) -> torch.device:
+    token = device.strip().lower()
+    if token == "auto":
+        token = "cuda" if torch.cuda.is_available() else "cpu"
+    return torch.device(token)
+
+
 @dataclass(frozen=True)
 class IterationMetrics:
     iteration: int
@@ -68,8 +75,10 @@ class DeepCFRTrainer:
         device: str = "cpu",
     ) -> None:
         self.config = config or DeepCFRConfig()
-        self.game_config = game_config or LostCitiesConfig(seed=self.config.run.seed)
-        self.device = torch.device(device)
+        self.game_config = game_config or self.config.rules.to_lost_cities_config(
+            seed=self.config.run.seed
+        )
+        self.device = _resolve_torch_device(device)
 
         probe = GameState.new_game(self.game_config, seed=self.config.run.seed)
         self.input_dim = input_dim(probe)
@@ -77,20 +86,26 @@ class DeepCFRTrainer:
 
         torch.manual_seed(self.config.run.seed)
         self.advantage_networks = [
-            DeepCFRMLP(self.input_dim, self.action_size, self.config.network.hidden_size).to(
+            DeepCFRMLP.from_config(self.input_dim, self.action_size, self.config.network).to(
                 self.device
             )
             for _ in range(2)
         ]
-        self.strategy_network = DeepCFRMLP(
-            self.input_dim, self.action_size, self.config.network.hidden_size
+        self.strategy_network = DeepCFRMLP.from_config(
+            self.input_dim, self.action_size, self.config.network
         ).to(self.device)
         self.advantage_optimizers = [
-            torch.optim.Adam(network.parameters(), lr=self.config.optimization.learning_rate)
+            torch.optim.Adam(
+                network.parameters(),
+                lr=self.config.optimization.learning_rate,
+                weight_decay=self.config.optimization.weight_decay,
+            )
             for network in self.advantage_networks
         ]
         self.strategy_optimizer = torch.optim.Adam(
-            self.strategy_network.parameters(), lr=self.config.optimization.learning_rate
+            self.strategy_network.parameters(),
+            lr=self.config.optimization.learning_rate,
+            weight_decay=self.config.optimization.weight_decay,
         )
         self.advantage_memory = ReservoirMemory(self.config.memory.advantage_capacity)
         self.strategy_memory = ReservoirMemory(self.config.memory.strategy_capacity)
@@ -175,7 +190,7 @@ class DeepCFRTrainer:
             store_strategy_on_traverser_nodes=self.config.traversal.store_strategy_on_traverser_nodes,
             store_strategy_on_opponent_nodes=self.config.traversal.store_strategy_on_opponent_nodes,
             max_depth=self.config.traversal.max_depth,
-            max_nodes=self.config.traversal.max_nodes,
+            max_nodes=self.config.traversal.resolved_max_nodes(),
             outcome_sampling_epsilon=self.config.traversal.outcome_sampling_epsilon,
             outcome_sampling_value_clip=self.config.traversal.outcome_sampling_value_clip,
             outcome_unsampled_regret=self.config.traversal.outcome_unsampled_regret,
@@ -195,7 +210,7 @@ class DeepCFRTrainer:
         )
         for network in self.advantage_networks:
             network.eval()
-        for traversal_index in range(self.config.traversal.traversals_per_iteration):
+        for traversal_index in range(self.config.traversal.resolved_traversals_per_player()):
             for player in range(2):
                 seed = self.config.run.seed + iteration * 10_000 + traversal_index * 10 + player
                 state = GameState.new_game(self.game_config, seed=seed)
@@ -227,12 +242,12 @@ class DeepCFRTrainer:
             {name: value.detach().cpu() for name, value in network.state_dict().items()}
             for network in self.advantage_networks
         ]
-        chunk_size = max(1, self.config.traversal.worker_chunk_size)
+        chunk_size = self.config.traversal.resolved_worker_chunk_size()
         batch_index = 0
         for player in range(2):
             seeds = [
                 self.config.run.seed + iteration * 10_000 + index * 10 + player
-                for index in range(self.config.traversal.traversals_per_iteration)
+                for index in range(self.config.traversal.resolved_traversals_per_player())
             ]
             for start in range(0, len(seeds), chunk_size):
                 chunk = seeds[start : start + chunk_size]
@@ -275,7 +290,7 @@ class DeepCFRTrainer:
         league: list[list[nn.Module]] = []
         for snapshot in self.self_play_league_snapshots:
             networks = [
-                DeepCFRMLP(self.input_dim, self.action_size, self.config.network.hidden_size).to(
+                DeepCFRMLP.from_config(self.input_dim, self.action_size, self.config.network).to(
                     self.device
                 )
                 for _ in range(2)
@@ -293,20 +308,48 @@ class DeepCFRTrainer:
         self._start_run_logging()
         metrics: list[IterationMetrics] = []
         start = self.iteration + 1
-        stop = self.iteration + self.config.run.iterations
-        for iteration in range(start, stop + 1):
+        stop = self._stop_iteration()
+        run_started = time.perf_counter()
+        iteration = start
+        while iteration <= stop:
             started = time.perf_counter()
             item = self.run_iteration(iteration)
             elapsed = time.perf_counter() - started
             metrics.append(item)
             self._append_metrics(item, elapsed)
             self._maybe_record_self_play_snapshot(iteration)
-            if self.config.checkpoint.save_every_iteration:
-                checkpoint_dir = self.run_dir
-                self.save_checkpoint(checkpoint_dir / f"iteration_{iteration:05d}.pt", item)
-                self.save_checkpoint(checkpoint_dir / "latest.pt", item)
+            if self._should_save_iteration(iteration):
+                self._save_iteration_checkpoints(iteration, item)
+            if self._time_limit_reached(run_started):
+                break
+            iteration += 1
         return metrics
 
+    def _stop_iteration(self) -> int:
+        if self.config.run.max_iterations is not None:
+            return max(self.iteration, int(self.config.run.max_iterations))
+        if self.config.run.max_hours is not None:
+            return 2**31 - 1
+        return self.iteration + self.config.run.iterations
+
+    def _time_limit_reached(self, run_started: float) -> bool:
+        if self.config.run.max_hours is None:
+            return False
+        elapsed_hours = (time.perf_counter() - run_started) / 3600.0
+        return elapsed_hours >= self.config.run.max_hours
+
+    def _should_save_iteration(self, iteration: int) -> bool:
+        if self.config.checkpoint.save_every_iteration:
+            return True
+        interval = int(self.config.checkpoint.save_iteration_interval)
+        return interval > 0 and iteration % interval == 0
+
+    def _save_iteration_checkpoints(self, iteration: int, item: IterationMetrics) -> None:
+        checkpoint_dir = self.run_dir
+        if not self.config.checkpoint.save_latest_only:
+            self.save_checkpoint(checkpoint_dir / f"iteration_{iteration:05d}.pt", item)
+        self.save_checkpoint(checkpoint_dir / "latest.pt", item)
+
     def _start_run_logging(self) -> None:
         self.run_dir.mkdir(parents=True, exist_ok=True)
         config_path = self.run_dir / "config.json"
@@ -371,14 +414,6 @@ class DeepCFRTrainer:
             return 0.0
         return self._train_strategy(self.strategy_network, self.strategy_optimizer, samples)
 
-    def _batch(self, samples: list[TrainingSample], step: int) -> list[TrainingSample]:
-        batch_size = min(self.config.optimization.batch_size, len(samples))
-        offset = (step * batch_size) % len(samples)
-        batch = samples[offset : offset + batch_size]
-        if len(batch) < batch_size:
-            batch = batch + samples[0 : batch_size - len(batch)]
-        return batch
-
     def _batch_tensors(
         self,
         batch: list[TrainingSample],
@@ -408,10 +443,12 @@ class DeepCFRTrainer:
     ) -> float:
         last_loss = 0.0
         network.train()
-        for _step in range(max(self.config.optimization.advantage_train_steps, 0)):
+        for _step in range(self.config.optimization.resolved_advantage_train_steps()):
             x, y, legal = self._batch_tensors(
                 self.advantage_memory.sample(
-                    self.config.optimization.batch_size, self.rng, player=player
+                    self.config.optimization.resolved_advantage_batch_size(),
+                    self.rng,
+                    player=player,
                 )
             )
             pred = network(x)
@@ -419,6 +456,10 @@ class DeepCFRTrainer:
             loss = diff.square().sum() / legal.sum().clamp_min(1)
             optimizer.zero_grad(set_to_none=True)
             loss.backward()
+            if self.config.optimization.grad_clip > 0.0:
+                torch.nn.utils.clip_grad_norm_(
+                    network.parameters(), self.config.optimization.grad_clip
+                )
             optimizer.step()
             last_loss = float(loss.detach().cpu())
         return last_loss
@@ -431,15 +472,21 @@ class DeepCFRTrainer:
     ) -> float:
         last_loss = 0.0
         network.train()
-        for _step in range(max(self.config.optimization.strategy_train_steps, 0)):
+        for _step in range(self.config.optimization.resolved_strategy_train_steps()):
             x, y, legal = self._batch_tensors(
-                self.strategy_memory.sample(self.config.optimization.batch_size, self.rng)
+                self.strategy_memory.sample(
+                    self.config.optimization.resolved_strategy_batch_size(), self.rng
+                )
             )
             logits = network(x).masked_fill(~legal, torch.finfo(torch.float32).min)
             log_probs = nn.functional.log_softmax(logits, dim=-1).masked_fill(~legal, 0.0)
             loss = -(y * log_probs).sum(dim=-1).mean()
             optimizer.zero_grad(set_to_none=True)
             loss.backward()
+            if self.config.optimization.grad_clip > 0.0:
+                torch.nn.utils.clip_grad_norm_(
+                    network.parameters(), self.config.optimization.grad_clip
+                )
             optimizer.step()
             last_loss = float(loss.detach().cpu())
         return last_loss

src/coolrl_lost_cities/games/classic/deep_cfr/workers.py

@@ -40,7 +40,7 @@ def run_traversal_worker_batch(batch: TraversalWorkerBatch) -> TraversalWorkerRe
     cfg = config_from_dict(batch.config)
     device = torch.device("cpu")
     networks = [
-        DeepCFRMLP(batch.input_dim, batch.action_size, cfg.network.hidden_size).to(device)
+        DeepCFRMLP.from_config(batch.input_dim, batch.action_size, cfg.network).to(device)
         for _ in range(2)
     ]
     for network, state_dict in zip(networks, batch.advantage_networks, strict=True):
@@ -49,7 +49,7 @@ def run_traversal_worker_batch(batch: TraversalWorkerBatch) -> TraversalWorkerRe
     league_networks: list[list[torch.nn.Module]] = []
     for snapshot in batch.league_advantage_networks:
         snapshot_networks = [
-            DeepCFRMLP(batch.input_dim, batch.action_size, cfg.network.hidden_size).to(device)
+            DeepCFRMLP.from_config(batch.input_dim, batch.action_size, cfg.network).to(device)
             for _ in range(2)
         ]
         for network, state_dict in zip(snapshot_networks, snapshot, strict=True):
@@ -69,7 +69,7 @@ def run_traversal_worker_batch(batch: TraversalWorkerBatch) -> TraversalWorkerRe
         store_strategy_on_traverser_nodes=cfg.traversal.store_strategy_on_traverser_nodes,
         store_strategy_on_opponent_nodes=cfg.traversal.store_strategy_on_opponent_nodes,
         max_depth=cfg.traversal.max_depth,
-        max_nodes=cfg.traversal.max_nodes,
+        max_nodes=cfg.traversal.resolved_max_nodes(),
         outcome_sampling_epsilon=cfg.traversal.outcome_sampling_epsilon,
         outcome_sampling_value_clip=cfg.traversal.outcome_sampling_value_clip,
         outcome_unsampled_regret=cfg.traversal.outcome_unsampled_regret,

tests/games/classic/test_deep_cfr_trainer.py

@@ -26,6 +26,33 @@ def test_deep_cfr_loads_smoke_yaml_config() -> None:
     assert config.checkpoint.directory == "runs/deep_cfr/smoke"
 
 
+def test_deep_cfr_loads_mapped_legacy_reproduction_config() -> None:
+    config = load_config(
+        "configs/deep_cfr/pure_self_play_zero_pit_poc_full_depth_slot_aware_playability.yaml"
+    )
+
+    assert config.run.experiment_name.endswith("slot_aware_playability")
+    assert config.run.seed == 79
+    assert config.run.max_iterations is None
+    assert config.run.max_hours == 4
+    assert config.encoding.derived_playability is True
+    assert config.encoding.slot_aware_playability is True
+    assert config.network.hidden_size == 256
+    assert config.network.num_layers == 3
+    assert config.traversal.resolved_traversals_per_player() == 70
+    assert config.traversal.max_depth is None
+    assert config.traversal.resolved_max_nodes() == 1000
+    assert config.traversal.resolved_worker_chunk_size() == 8
+    assert config.optimization.resolved_advantage_batch_size() == 1024
+    assert config.optimization.resolved_strategy_batch_size() == 1024
+    assert config.optimization.resolved_advantage_train_steps() == 256
+    assert config.optimization.resolved_strategy_train_steps() == 256
+    assert config.optimization.weight_decay == 0.0001
+    assert config.optimization.grad_clip == 1.0
+    assert config.evaluation.on_max_steps == "score_diff"
+    assert config.checkpoint.save_iteration_interval == 10
+
+
 def test_deep_cfr_trainer_smoke_run() -> None:
     trainer = DeepCFRTrainer(
         _deep_cfr_config(