File size: 33,342 Bytes
40ee6b4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
"""
Training script for the RNN Meta-Controller.

This module provides a complete training pipeline for the RNN-based meta-controller,
including data generation/loading, model training with early stopping, validation,
checkpointing, and comprehensive evaluation with per-class metrics.
"""

import argparse
import json
import logging
from pathlib import Path
from typing import Any

import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import DataLoader, TensorDataset

from src.agents.meta_controller.rnn_controller import (
    RNNMetaControllerModel,
)
from src.training.data_generator import MetaControllerDataGenerator

# Braintrust integration (optional)
try:
    from src.observability.braintrust_tracker import BraintrustTracker, create_training_tracker

    BRAINTRUST_AVAILABLE = True
except ImportError:
    BRAINTRUST_AVAILABLE = False
    BraintrustTracker = None  # type: ignore


class RNNTrainer:
    """
    Trainer class for the RNN Meta-Controller model.

    Handles the complete training pipeline including data loading, training loops,
    validation, early stopping, model checkpointing, and comprehensive evaluation.

    Attributes:
        hidden_dim: Dimension of the GRU hidden state.
        num_layers: Number of GRU layers.
        dropout: Dropout probability for regularization.
        lr: Learning rate for the optimizer.
        batch_size: Batch size for training and evaluation.
        epochs: Maximum number of training epochs.
        early_stopping_patience: Number of epochs to wait for improvement before stopping.
        seed: Random seed for reproducibility.
        device: PyTorch device for computation.
        model: The RNNMetaControllerModel instance.
        optimizer: Adam optimizer for training.
        criterion: CrossEntropyLoss for classification.
        logger: Logger instance for progress reporting.

    Example:
        >>> trainer = RNNTrainer(hidden_dim=64, epochs=10)
        >>> generator = MetaControllerDataGenerator(seed=42)
        >>> features, labels = generator.generate_balanced_dataset(100)
        >>> X, y = generator.to_tensor_dataset(features, labels)
        >>> splits = generator.split_dataset(X, y)
        >>> history = trainer.train(
        ...     train_data=(splits['X_train'], splits['y_train']),
        ...     val_data=(splits['X_val'], splits['y_val'])
        ... )
    """

    AGENT_NAMES = ["hrm", "trm", "mcts"]
    LABEL_TO_INDEX = {"hrm": 0, "trm": 1, "mcts": 2}
    INDEX_TO_LABEL = {0: "hrm", 1: "trm", 2: "mcts"}

    def __init__(
        self,
        hidden_dim: int = 64,
        num_layers: int = 1,
        dropout: float = 0.1,
        lr: float = 1e-3,
        batch_size: int = 32,
        epochs: int = 10,
        early_stopping_patience: int = 3,
        seed: int = 42,
        device: str | None = None,
        braintrust_tracker: Any | None = None,
    ) -> None:
        """
        Initialize the RNN trainer.

        Args:
            hidden_dim: Dimension of GRU hidden state. Defaults to 64.
            num_layers: Number of stacked GRU layers. Defaults to 1.
            dropout: Dropout probability for regularization. Defaults to 0.1.
            lr: Learning rate for Adam optimizer. Defaults to 1e-3.
            batch_size: Batch size for training and evaluation. Defaults to 32.
            epochs: Maximum number of training epochs. Defaults to 10.
            early_stopping_patience: Epochs to wait for improvement before early stopping.
                Defaults to 3.
            seed: Random seed for reproducibility. Defaults to 42.
            device: Device to run training on ('cpu', 'cuda', 'mps').
                If None, auto-detects best available device.
            braintrust_tracker: Optional BraintrustTracker for experiment tracking.
        """
        # Store hyperparameters
        self.hidden_dim = hidden_dim
        self.num_layers = num_layers
        self.dropout = dropout
        self.lr = lr
        self.batch_size = batch_size
        self.epochs = epochs
        self.early_stopping_patience = early_stopping_patience
        self.seed = seed

        # Set random seeds for reproducibility
        torch.manual_seed(seed)
        if torch.cuda.is_available():
            torch.cuda.manual_seed_all(seed)

        # Auto-detect device if not specified
        if device is None:
            if torch.cuda.is_available():
                self.device = torch.device("cuda")
            elif hasattr(torch.backends, "mps") and torch.backends.mps.is_available():
                self.device = torch.device("mps")
            else:
                self.device = torch.device("cpu")
        else:
            self.device = torch.device(device)

        # Setup logging
        self._setup_logging()
        self.logger.info(f"Initializing RNNTrainer with device: {self.device}")

        # Initialize model
        self.model = RNNMetaControllerModel(
            input_dim=10,  # Fixed based on features_to_tensor output
            hidden_dim=hidden_dim,
            num_layers=num_layers,
            num_agents=len(self.AGENT_NAMES),
            dropout=dropout,
        )
        self.model = self.model.to(self.device)
        self.logger.info(f"Model initialized: hidden_dim={hidden_dim}, num_layers={num_layers}, dropout={dropout}")

        # Setup optimizer
        self.optimizer = optim.Adam(self.model.parameters(), lr=lr)
        self.logger.info(f"Optimizer: Adam with lr={lr}")

        # Setup loss function
        self.criterion = nn.CrossEntropyLoss()
        self.logger.info("Loss function: CrossEntropyLoss")

        # Braintrust experiment tracking (optional)
        self.braintrust_tracker = braintrust_tracker
        if self.braintrust_tracker and hasattr(self.braintrust_tracker, "is_available"):
            if self.braintrust_tracker.is_available:
                self.logger.info("Braintrust experiment tracking enabled")
                self.braintrust_tracker.log_hyperparameters(
                    {
                        "hidden_dim": hidden_dim,
                        "num_layers": num_layers,
                        "dropout": dropout,
                        "learning_rate": lr,
                        "batch_size": batch_size,
                        "max_epochs": epochs,
                        "early_stopping_patience": early_stopping_patience,
                        "seed": seed,
                        "device": str(self.device),
                    }
                )
            else:
                self.logger.info("Braintrust tracker provided but not available")

    def _setup_logging(self) -> None:
        """
        Setup logging configuration for the trainer.

        Creates a logger with console handler and appropriate formatting.
        """
        self.logger = logging.getLogger("RNNTrainer")
        self.logger.setLevel(logging.INFO)

        # Avoid duplicate handlers
        if not self.logger.handlers:
            console_handler = logging.StreamHandler()
            console_handler.setLevel(logging.INFO)
            formatter = logging.Formatter(
                "%(asctime)s - %(name)s - %(levelname)s - %(message)s",
                datefmt="%Y-%m-%d %H:%M:%S",
            )
            console_handler.setFormatter(formatter)
            self.logger.addHandler(console_handler)

    def create_dataloader(
        self,
        X: torch.Tensor,
        y: torch.Tensor,
        batch_size: int | None = None,
        shuffle: bool = True,
    ) -> DataLoader:
        """
        Create a PyTorch DataLoader from feature and label tensors.

        Args:
            X: Feature tensor of shape (N, 10).
            y: Label tensor of shape (N,).
            batch_size: Batch size for the DataLoader. If None, uses self.batch_size.
            shuffle: Whether to shuffle the data. Defaults to True.

        Returns:
            DataLoader instance for iterating over batches.

        Example:
            >>> trainer = RNNTrainer()
            >>> X = torch.randn(100, 10)
            >>> y = torch.randint(0, 3, (100,))
            >>> loader = trainer.create_dataloader(X, y, batch_size=16)
            >>> len(loader)
            7
        """
        if batch_size is None:
            batch_size = self.batch_size

        # Ensure tensors are on CPU for DataLoader
        if X.device != torch.device("cpu"):
            X = X.cpu()
        if y.device != torch.device("cpu"):
            y = y.cpu()

        dataset = TensorDataset(X, y)
        loader = DataLoader(
            dataset,
            batch_size=batch_size,
            shuffle=shuffle,
            num_workers=0,  # Use main process for data loading
            pin_memory=self.device.type == "cuda",
        )

        return loader

    def train_epoch(self, train_loader: DataLoader) -> float:
        """
        Train the model for one epoch.

        Args:
            train_loader: DataLoader providing training batches.

        Returns:
            Average training loss for the epoch.

        Example:
            >>> trainer = RNNTrainer()
            >>> X = torch.randn(100, 10)
            >>> y = torch.randint(0, 3, (100,))
            >>> loader = trainer.create_dataloader(X, y)
            >>> loss = trainer.train_epoch(loader)
            >>> isinstance(loss, float)
            True
        """
        self.model.train()
        total_loss = 0.0
        num_batches = 0

        for batch_X, batch_y in train_loader:
            # Move data to device
            batch_X = batch_X.to(self.device)
            batch_y = batch_y.to(self.device)

            # Zero gradients
            self.optimizer.zero_grad()

            # Forward pass
            logits = self.model(batch_X)

            # Compute loss
            loss = self.criterion(logits, batch_y)

            # Backward pass
            loss.backward()

            # Update weights
            self.optimizer.step()

            # Accumulate loss
            total_loss += loss.item()
            num_batches += 1

        average_loss = total_loss / num_batches if num_batches > 0 else 0.0
        return average_loss

    def validate(self, val_loader: DataLoader) -> tuple[float, float]:
        """
        Evaluate the model on the validation set.

        Args:
            val_loader: DataLoader providing validation batches.

        Returns:
            Tuple of (average_loss, accuracy).
            - average_loss: Mean cross-entropy loss over validation set.
            - accuracy: Classification accuracy as a fraction [0, 1].

        Example:
            >>> trainer = RNNTrainer()
            >>> X = torch.randn(50, 10)
            >>> y = torch.randint(0, 3, (50,))
            >>> loader = trainer.create_dataloader(X, y, shuffle=False)
            >>> loss, acc = trainer.validate(loader)
            >>> 0.0 <= acc <= 1.0
            True
        """
        self.model.eval()
        total_loss = 0.0
        correct = 0
        total = 0

        with torch.no_grad():
            for batch_X, batch_y in val_loader:
                # Move data to device
                batch_X = batch_X.to(self.device)
                batch_y = batch_y.to(self.device)

                # Forward pass
                logits = self.model(batch_X)

                # Compute loss
                loss = self.criterion(logits, batch_y)
                total_loss += loss.item()

                # Compute accuracy
                predictions = torch.argmax(logits, dim=1)
                correct += (predictions == batch_y).sum().item()
                total += batch_y.size(0)

        num_batches = len(val_loader)
        average_loss = total_loss / num_batches if num_batches > 0 else 0.0
        accuracy = correct / total if total > 0 else 0.0

        return average_loss, accuracy

    def train(
        self,
        train_data: tuple[torch.Tensor, torch.Tensor],
        val_data: tuple[torch.Tensor, torch.Tensor],
        save_path: str | None = None,
    ) -> dict[str, Any]:
        """
        Main training loop with early stopping and model checkpointing.

        Trains the model for the specified number of epochs, monitoring validation
        loss for early stopping. If save_path is provided, saves the best model
        checkpoint based on validation loss.

        Args:
            train_data: Tuple of (X_train, y_train) tensors.
            val_data: Tuple of (X_val, y_val) tensors.
            save_path: Optional path to save the best model checkpoint.

        Returns:
            Dictionary containing training history:
            - 'train_losses': List of training losses per epoch.
            - 'val_losses': List of validation losses per epoch.
            - 'val_accuracies': List of validation accuracies per epoch.
            - 'best_epoch': Epoch with best validation loss.
            - 'best_val_loss': Best validation loss achieved.
            - 'best_val_accuracy': Validation accuracy at best epoch.
            - 'stopped_early': Whether training stopped early.
            - 'total_epochs': Total number of epochs trained.

        Example:
            >>> trainer = RNNTrainer(epochs=5)
            >>> X_train = torch.randn(100, 10)
            >>> y_train = torch.randint(0, 3, (100,))
            >>> X_val = torch.randn(20, 10)
            >>> y_val = torch.randint(0, 3, (20,))
            >>> history = trainer.train((X_train, y_train), (X_val, y_val))
            >>> 'train_losses' in history
            True
            >>> len(history['train_losses']) <= 5
            True
        """
        self.logger.info("Starting training...")
        self.logger.info(f"Training samples: {train_data[0].shape[0]}")
        self.logger.info(f"Validation samples: {val_data[0].shape[0]}")
        self.logger.info(f"Batch size: {self.batch_size}")
        self.logger.info(f"Max epochs: {self.epochs}")
        self.logger.info(f"Early stopping patience: {self.early_stopping_patience}")

        # Create data loaders
        train_loader = self.create_dataloader(train_data[0], train_data[1], shuffle=True)
        val_loader = self.create_dataloader(val_data[0], val_data[1], shuffle=False)

        # Initialize tracking variables
        train_losses: list[float] = []
        val_losses: list[float] = []
        val_accuracies: list[float] = []

        best_val_loss = float("inf")
        best_val_accuracy = 0.0
        best_epoch = 0
        best_model_state = None
        patience_counter = 0
        stopped_early = False

        # Training loop
        for epoch in range(1, self.epochs + 1):
            # Train for one epoch
            train_loss = self.train_epoch(train_loader)
            train_losses.append(train_loss)

            # Validate
            val_loss, val_accuracy = self.validate(val_loader)
            val_losses.append(val_loss)
            val_accuracies.append(val_accuracy)

            # Log progress
            self.logger.info(
                f"Epoch {epoch}/{self.epochs} - "
                f"Train Loss: {train_loss:.4f}, "
                f"Val Loss: {val_loss:.4f}, "
                f"Val Accuracy: {val_accuracy:.4f}"
            )

            # Log to Braintrust if available
            if self.braintrust_tracker and hasattr(self.braintrust_tracker, "log_epoch_summary"):
                self.braintrust_tracker.log_epoch_summary(
                    epoch=epoch,
                    train_loss=train_loss,
                    val_loss=val_loss,
                    val_accuracy=val_accuracy,
                )

            # Check for improvement
            if val_loss < best_val_loss:
                best_val_loss = val_loss
                best_val_accuracy = val_accuracy
                best_epoch = epoch
                best_model_state = self.model.state_dict().copy()
                patience_counter = 0
                self.logger.info(f"  -> New best validation loss: {val_loss:.4f}")

                # Save checkpoint if path provided
                if save_path:
                    torch.save(best_model_state, save_path)
                    self.logger.info(f"  -> Model checkpoint saved to {save_path}")
            else:
                patience_counter += 1
                self.logger.info(f"  -> No improvement for {patience_counter} epoch(s)")

                # Check for early stopping
                if patience_counter >= self.early_stopping_patience:
                    self.logger.info(f"Early stopping triggered at epoch {epoch}. Best epoch was {best_epoch}.")
                    stopped_early = True
                    break

        # Restore best model state
        if best_model_state is not None:
            self.model.load_state_dict(best_model_state)
            self.logger.info(
                f"Restored best model from epoch {best_epoch} "
                f"with val_loss={best_val_loss:.4f}, val_accuracy={best_val_accuracy:.4f}"
            )

        # Final save if path provided and not already saved
        if save_path and best_model_state is not None:
            torch.save(best_model_state, save_path)
            self.logger.info(f"Final model saved to {save_path}")

        # Compile history
        history = {
            "train_losses": train_losses,
            "val_losses": val_losses,
            "val_accuracies": val_accuracies,
            "best_epoch": best_epoch,
            "best_val_loss": best_val_loss,
            "best_val_accuracy": best_val_accuracy,
            "stopped_early": stopped_early,
            "total_epochs": len(train_losses),
        }

        self.logger.info("Training completed!")
        self.logger.info(f"Best epoch: {best_epoch}")
        self.logger.info(f"Best validation loss: {best_val_loss:.4f}")
        self.logger.info(f"Best validation accuracy: {best_val_accuracy:.4f}")

        # Log final model artifact to Braintrust
        if self.braintrust_tracker and hasattr(self.braintrust_tracker, "log_model_artifact"):
            self.braintrust_tracker.log_model_artifact(
                model_path=str(save_path) if save_path else "in_memory",
                model_type="rnn",
                metrics={
                    "best_val_loss": best_val_loss,
                    "best_val_accuracy": best_val_accuracy,
                    "best_epoch": float(best_epoch),
                    "total_epochs": float(len(train_losses)),
                },
            )

        return history

    def evaluate(self, test_loader: DataLoader) -> dict[str, Any]:
        """
        Comprehensive evaluation on the test set.

        Computes overall metrics and per-class precision, recall, and F1-score.

        Args:
            test_loader: DataLoader providing test batches.

        Returns:
            Dictionary containing:
            - 'loss': Average cross-entropy loss.
            - 'accuracy': Overall classification accuracy.
            - 'per_class_metrics': Dictionary with per-class metrics:
                - For each agent ('hrm', 'trm', 'mcts'):
                    - 'precision': Precision score.
                    - 'recall': Recall score.
                    - 'f1_score': F1 score.
                    - 'support': Number of samples in this class.
            - 'confusion_matrix': 3x3 confusion matrix as nested list.
            - 'total_samples': Total number of test samples.

        Example:
            >>> trainer = RNNTrainer()
            >>> X = torch.randn(50, 10)
            >>> y = torch.randint(0, 3, (50,))
            >>> loader = trainer.create_dataloader(X, y, shuffle=False)
            >>> results = trainer.evaluate(loader)
            >>> 'accuracy' in results
            True
            >>> 'per_class_metrics' in results
            True
        """
        self.model.eval()
        total_loss = 0.0
        all_predictions: list[int] = []
        all_labels: list[int] = []

        with torch.no_grad():
            for batch_X, batch_y in test_loader:
                # Move data to device
                batch_X = batch_X.to(self.device)
                batch_y = batch_y.to(self.device)

                # Forward pass
                logits = self.model(batch_X)

                # Compute loss
                loss = self.criterion(logits, batch_y)
                total_loss += loss.item()

                # Get predictions
                predictions = torch.argmax(logits, dim=1)
                all_predictions.extend(predictions.cpu().tolist())
                all_labels.extend(batch_y.cpu().tolist())

        # Calculate overall metrics
        num_batches = len(test_loader)
        average_loss = total_loss / num_batches if num_batches > 0 else 0.0

        correct = sum(p == label for p, label in zip(all_predictions, all_labels, strict=False))
        total = len(all_labels)
        accuracy = correct / total if total > 0 else 0.0

        # Calculate confusion matrix
        num_classes = len(self.AGENT_NAMES)
        confusion_matrix = [[0] * num_classes for _ in range(num_classes)]
        for pred, label in zip(all_predictions, all_labels, strict=False):
            confusion_matrix[label][pred] += 1

        # Calculate per-class metrics
        per_class_metrics: dict[str, dict[str, float]] = {}

        for class_idx, agent_name in enumerate(self.AGENT_NAMES):
            # True positives: predicted as this class and actually this class
            tp = confusion_matrix[class_idx][class_idx]

            # False positives: predicted as this class but actually other class
            fp = sum(confusion_matrix[i][class_idx] for i in range(num_classes) if i != class_idx)

            # False negatives: actually this class but predicted as other class
            fn = sum(confusion_matrix[class_idx][j] for j in range(num_classes) if j != class_idx)

            # Support: total number of samples in this class
            support = sum(confusion_matrix[class_idx])

            # Precision: TP / (TP + FP)
            precision = tp / (tp + fp) if (tp + fp) > 0 else 0.0

            # Recall: TP / (TP + FN)
            recall = tp / (tp + fn) if (tp + fn) > 0 else 0.0

            # F1 Score: 2 * (Precision * Recall) / (Precision + Recall)
            f1_score = 2 * (precision * recall) / (precision + recall) if (precision + recall) > 0 else 0.0

            per_class_metrics[agent_name] = {
                "precision": precision,
                "recall": recall,
                "f1_score": f1_score,
                "support": support,
            }

        results = {
            "loss": average_loss,
            "accuracy": accuracy,
            "per_class_metrics": per_class_metrics,
            "confusion_matrix": confusion_matrix,
            "total_samples": total,
        }

        self.logger.info("Evaluation Results:")
        self.logger.info(f"  Test Loss: {average_loss:.4f}")
        self.logger.info(f"  Test Accuracy: {accuracy:.4f}")
        self.logger.info(f"  Total Samples: {total}")
        self.logger.info("  Per-Class Metrics:")
        for agent_name, metrics in per_class_metrics.items():
            self.logger.info(
                f"    {agent_name}: "
                f"Precision={metrics['precision']:.4f}, "
                f"Recall={metrics['recall']:.4f}, "
                f"F1={metrics['f1_score']:.4f}, "
                f"Support={metrics['support']}"
            )

        return results


def main() -> None:
    """
    Main entry point for training the RNN Meta-Controller.

    Parses command-line arguments, generates or loads dataset, trains the model,
    evaluates on test set, and saves results.
    """
    parser = argparse.ArgumentParser(
        description="Train the RNN Meta-Controller for agent selection.",
        formatter_class=argparse.ArgumentDefaultsHelpFormatter,
    )

    # Model hyperparameters
    parser.add_argument(
        "--hidden_dim",
        type=int,
        default=64,
        help="Dimension of GRU hidden state",
    )
    parser.add_argument(
        "--num_layers",
        type=int,
        default=1,
        help="Number of GRU layers",
    )
    parser.add_argument(
        "--dropout",
        type=float,
        default=0.1,
        help="Dropout probability",
    )

    # Training hyperparameters
    parser.add_argument(
        "--lr",
        type=float,
        default=1e-3,
        help="Learning rate for Adam optimizer",
    )
    parser.add_argument(
        "--batch_size",
        type=int,
        default=32,
        help="Batch size for training and evaluation",
    )
    parser.add_argument(
        "--epochs",
        type=int,
        default=10,
        help="Maximum number of training epochs",
    )
    parser.add_argument(
        "--patience",
        type=int,
        default=3,
        help="Early stopping patience (epochs without improvement)",
    )
    parser.add_argument(
        "--seed",
        type=int,
        default=42,
        help="Random seed for reproducibility",
    )

    # Data parameters
    parser.add_argument(
        "--num_samples",
        type=int,
        default=3000,
        help="Number of samples to generate (per class for balanced dataset)",
    )
    parser.add_argument(
        "--data_path",
        type=str,
        default=None,
        help="Path to load existing dataset (JSON format). If not provided, generates new data.",
    )

    # Output parameters
    parser.add_argument(
        "--save_path",
        type=str,
        default="rnn_meta_controller.pt",
        help="Path to save the trained model",
    )

    # Experiment tracking
    parser.add_argument(
        "--use_braintrust",
        action="store_true",
        help="Enable Braintrust experiment tracking",
    )
    parser.add_argument(
        "--experiment_name",
        type=str,
        default=None,
        help="Custom experiment name for Braintrust (auto-generated if not provided)",
    )

    args = parser.parse_args()

    # Setup logging for main
    logging.basicConfig(
        level=logging.INFO,
        format="%(asctime)s - %(name)s - %(levelname)s - %(message)s",
        datefmt="%Y-%m-%d %H:%M:%S",
    )
    logger = logging.getLogger("train_rnn")

    logger.info("=" * 60)
    logger.info("RNN Meta-Controller Training")
    logger.info("=" * 60)

    # Print configuration
    logger.info("Configuration:")
    for arg_name, arg_value in vars(args).items():
        logger.info(f"  {arg_name}: {arg_value}")
    logger.info("")

    try:
        # Initialize data generator
        generator = MetaControllerDataGenerator(seed=args.seed)

        # Load or generate dataset
        if args.data_path and Path(args.data_path).exists():
            logger.info(f"Loading dataset from {args.data_path}...")
            features_list, labels_list = generator.load_dataset(args.data_path)
            logger.info(f"Loaded {len(features_list)} samples")
        else:
            logger.info(f"Generating balanced dataset with {args.num_samples} samples per class...")
            features_list, labels_list = generator.generate_balanced_dataset(num_samples_per_class=args.num_samples)
            total_samples = len(features_list)
            logger.info(f"Generated {total_samples} total samples")

            # Optionally save generated dataset
            if args.data_path:
                logger.info(f"Saving generated dataset to {args.data_path}...")
                generator.save_dataset(features_list, labels_list, args.data_path)

        # Convert to tensors
        logger.info("Converting dataset to tensors...")
        X, y = generator.to_tensor_dataset(features_list, labels_list)
        logger.info(f"Feature tensor shape: {X.shape}")
        logger.info(f"Label tensor shape: {y.shape}")

        # Split dataset
        logger.info("Splitting dataset into train/val/test (70%/15%/15%)...")
        splits = generator.split_dataset(X, y, train_ratio=0.7, val_ratio=0.15)

        logger.info(f"Training set size: {splits['X_train'].shape[0]}")
        logger.info(f"Validation set size: {splits['X_val'].shape[0]}")
        logger.info(f"Test set size: {splits['X_test'].shape[0]}")
        logger.info("")

        # Initialize Braintrust tracker if enabled
        braintrust_tracker = None
        if args.use_braintrust and BRAINTRUST_AVAILABLE:
            logger.info("Initializing Braintrust experiment tracker...")
            braintrust_tracker = create_training_tracker(
                model_type="rnn",
                config={
                    "hidden_dim": args.hidden_dim,
                    "num_layers": args.num_layers,
                    "dropout": args.dropout,
                    "lr": args.lr,
                    "batch_size": args.batch_size,
                    "epochs": args.epochs,
                    "patience": args.patience,
                    "seed": args.seed,
                    "num_samples": args.num_samples,
                },
            )
            if braintrust_tracker.is_available:
                logger.info("Braintrust experiment tracking enabled")
            else:
                logger.info("Braintrust not available (check API key)")
        elif args.use_braintrust and not BRAINTRUST_AVAILABLE:
            logger.warning("Braintrust requested but not installed. Install with: pip install braintrust")

        # Initialize trainer
        logger.info("Initializing trainer...")
        trainer = RNNTrainer(
            hidden_dim=args.hidden_dim,
            num_layers=args.num_layers,
            dropout=args.dropout,
            lr=args.lr,
            batch_size=args.batch_size,
            epochs=args.epochs,
            early_stopping_patience=args.patience,
            seed=args.seed,
            braintrust_tracker=braintrust_tracker,
        )
        logger.info("")

        # Train model
        logger.info("Starting training...")
        logger.info("-" * 60)
        history = trainer.train(
            train_data=(splits["X_train"], splits["y_train"]),
            val_data=(splits["X_val"], splits["y_val"]),
            save_path=args.save_path,
        )
        logger.info("-" * 60)
        logger.info("")

        # Evaluate on test set
        logger.info("Evaluating on test set...")
        logger.info("-" * 60)
        test_loader = trainer.create_dataloader(splits["X_test"], splits["y_test"], shuffle=False)
        test_results = trainer.evaluate(test_loader)
        logger.info("-" * 60)
        logger.info("")

        # Save training history
        history_path = Path(args.save_path).with_suffix(".history.json")
        logger.info(f"Saving training history to {history_path}...")

        # Combine history and test results
        full_results = {
            "config": {
                "hidden_dim": args.hidden_dim,
                "num_layers": args.num_layers,
                "dropout": args.dropout,
                "lr": args.lr,
                "batch_size": args.batch_size,
                "epochs": args.epochs,
                "patience": args.patience,
                "seed": args.seed,
                "num_samples": args.num_samples,
            },
            "training_history": history,
            "test_results": test_results,
        }

        with open(history_path, "w", encoding="utf-8") as f:
            json.dump(full_results, f, indent=2)

        logger.info(f"Training history saved to {history_path}")
        logger.info("")

        # Print final summary
        logger.info("=" * 60)
        logger.info("Training Summary")
        logger.info("=" * 60)
        logger.info(f"Model saved to: {args.save_path}")
        logger.info(f"History saved to: {history_path}")
        logger.info(f"Best validation accuracy: {history['best_val_accuracy']:.4f}")
        logger.info(f"Test accuracy: {test_results['accuracy']:.4f}")
        logger.info(f"Test loss: {test_results['loss']:.4f}")

        if history["stopped_early"]:
            logger.info(f"Training stopped early at epoch {history['total_epochs']}")
        else:
            logger.info(f"Training completed all {history['total_epochs']} epochs")

        logger.info("")
        logger.info("Per-class test performance:")
        for agent_name, metrics in test_results["per_class_metrics"].items():
            logger.info(
                f"  {agent_name}: F1={metrics['f1_score']:.4f}, "
                f"Precision={metrics['precision']:.4f}, "
                f"Recall={metrics['recall']:.4f}"
            )

        # End Braintrust experiment
        if braintrust_tracker and hasattr(braintrust_tracker, "end_experiment"):
            experiment_url = braintrust_tracker.end_experiment()
            if experiment_url:
                logger.info(f"Braintrust experiment URL: {experiment_url}")

        logger.info("=" * 60)
        logger.info("Training completed successfully!")
        logger.info("=" * 60)

    except FileNotFoundError as e:
        logger.error(f"File not found: {e}")
        raise
    except ValueError as e:
        logger.error(f"Invalid value: {e}")
        raise
    except RuntimeError as e:
        logger.error(f"Runtime error: {e}")
        raise
    except Exception as e:
        logger.error(f"Unexpected error: {e}")
        raise


if __name__ == "__main__":
    main()