File size: 17,991 Bytes
a888789
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""
Dueling Double DQN agent with Prioritized Experience Replay (PER).

Architecture upgrades over vanilla DDQN:
  - Dueling Network: Splits Q(s,a) = V(s) + A(s,a) - mean(A) for better
    state evaluation even when actions don't matter much.
  - Prioritized Experience Replay: Samples high-TD-error transitions more
    frequently, accelerating learning on surprising outcomes.
  - Double DQN: Decouples action selection (main net) from evaluation
    (target net) to reduce overestimation bias.

Backward compatible: `DQNAgent.load()` auto-detects old model format
and loads into the legacy QNetwork architecture seamlessly.
"""

from __future__ import annotations

from collections import deque
from dataclasses import dataclass
from typing import Deque, Dict, List, Optional, Tuple

import random
import numpy as np
import torch
import torch.nn as nn
import torch.optim as optim


# ---------------------------------------------------------------------------
# Q-networks
# ---------------------------------------------------------------------------

class QNetwork(nn.Module):
    """
    Standard MLP Q-network (legacy architecture).
    Kept for backward compatibility with old saved models.
    """
    def __init__(self, obs_size: int, num_actions: int):
        super().__init__()
        self.net = nn.Sequential(
            nn.Linear(obs_size, 128),
            nn.ReLU(),
            nn.Linear(128, 128),
            nn.ReLU(),
            nn.Linear(128, num_actions),
        )

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        return self.net(x)


class DuelingQNetwork(nn.Module):
    """
    Dueling DQN architecture (Wang et al., 2016).

    Splits the Q-value into two streams:
      Q(s, a) = V(s) + A(s, a) - mean(A(s, ·))

    The Value stream learns "how good is this state?"
    The Advantage stream learns "how much better is action a vs. average?"

    This decomposition improves learning efficiency because the agent
    can learn the value of a state independently of action effects,
    which is especially useful when many actions have similar outcomes.
    """
    def __init__(self, obs_size: int, num_actions: int):
        super().__init__()
        self.feature = nn.Sequential(
            nn.Linear(obs_size, 128),
            nn.ReLU(),
        )
        # Value stream: scalar state value V(s)
        self.value_stream = nn.Sequential(
            nn.Linear(128, 128),
            nn.ReLU(),
            nn.Linear(128, 1),
        )
        # Advantage stream: per-action advantage A(s, a)
        self.advantage_stream = nn.Sequential(
            nn.Linear(128, 128),
            nn.ReLU(),
            nn.Linear(128, num_actions),
        )

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        features = self.feature(x)
        value = self.value_stream(features)              # (batch, 1)
        advantage = self.advantage_stream(features)      # (batch, actions)
        # Combine: Q = V + (A - mean(A))
        q_values = value + advantage - advantage.mean(dim=1, keepdim=True)
        return q_values


# ---------------------------------------------------------------------------
# Configuration
# ---------------------------------------------------------------------------

@dataclass
class DQNConfig:
    """Hyperparameters for Dueling DDQN + PER training."""
    gamma: float = 0.99
    lr: float = 5e-4
    batch_size: int = 128
    replay_size: int = 100_000
    min_replay_size: int = 2_000
    target_update_every: int = 1_000
    epsilon_start: float = 1.0
    epsilon_end: float = 0.05
    epsilon_decay_steps: int = 50_000
    epsilon_decay_mult: float = 0.998
    epsilon_reset_every_episodes: int = 0
    epsilon_reset_value: float = 0.3
    max_grad_norm: float = 1.0
    # PER hyperparameters
    per_alpha: float = 0.6      # prioritization exponent (0 = uniform, 1 = full priority)
    per_beta_start: float = 0.4 # importance sampling correction (anneals to 1.0)
    per_beta_end: float = 1.0
    per_beta_anneal_steps: int = 100_000
    per_epsilon: float = 1e-6   # small constant to prevent zero priority


# ---------------------------------------------------------------------------
# Prioritized Experience Replay buffer
# ---------------------------------------------------------------------------

class SumTree:
    """Binary sum-tree for O(log N) prioritized sampling."""

    def __init__(self, capacity: int):
        self.capacity = int(capacity)
        self.tree = np.zeros(2 * self.capacity - 1, dtype=np.float64)
        self.data = [None] * self.capacity
        self.write_idx = 0
        self.size = 0

    def _propagate(self, idx: int, change: float) -> None:
        parent = (idx - 1) // 2
        self.tree[parent] += change
        if parent > 0:
            self._propagate(parent, change)

    def _retrieve(self, idx: int, s: float) -> int:
        left = 2 * idx + 1
        right = left + 1
        if left >= len(self.tree):
            return idx
        if s <= self.tree[left]:
            return self._retrieve(left, s)
        return self._retrieve(right, s - self.tree[left])

    @property
    def total(self) -> float:
        return float(self.tree[0])

    @property
    def max_priority(self) -> float:
        leaf_start = self.capacity - 1
        return float(max(self.tree[leaf_start:leaf_start + self.size])) if self.size > 0 else 1.0

    def add(self, priority: float, data) -> None:
        idx = self.write_idx + self.capacity - 1
        self.data[self.write_idx] = data
        self.update(idx, priority)
        self.write_idx = (self.write_idx + 1) % self.capacity
        self.size = min(self.size + 1, self.capacity)

    def update(self, idx: int, priority: float) -> None:
        change = priority - self.tree[idx]
        self.tree[idx] = priority
        self._propagate(idx, change)

    def get(self, s: float):
        idx = self._retrieve(0, s)
        data_idx = idx - self.capacity + 1
        return idx, float(self.tree[idx]), self.data[data_idx]


class PrioritizedReplayBuffer:
    """
    Prioritized Experience Replay (Schaul et al., 2016).

    Samples transitions with probability proportional to their TD-error,
    so the agent focuses learning on "surprising" transitions.
    """

    def __init__(self, capacity: int, alpha: float = 0.6, seed: int = 0):
        self.tree = SumTree(capacity)
        self.alpha = alpha
        self.rng = np.random.default_rng(seed)
        self._max_priority = 1.0

    def __len__(self) -> int:
        return self.tree.size

    def add(self, s: np.ndarray, a: int, r: float, s2: np.ndarray, done: bool) -> None:
        data = (s.astype(np.float32), int(a), float(r), s2.astype(np.float32), bool(done))
        priority = self._max_priority ** self.alpha
        self.tree.add(priority, data)

    def sample(
        self, batch_size: int, beta: float = 0.4
    ) -> Tuple[np.ndarray, np.ndarray, np.ndarray, np.ndarray, np.ndarray, np.ndarray, List[int]]:
        """Sample a batch with importance-sampling weights."""
        indices = []
        priorities = []
        batch = []

        segment = self.tree.total / batch_size

        for i in range(batch_size):
            low = segment * i
            high = segment * (i + 1)
            s_val = float(self.rng.uniform(low, high))
            idx, priority, data = self.tree.get(s_val)
            if data is None:
                # Fallback: resample from valid range
                s_val = float(self.rng.uniform(0, self.tree.total))
                idx, priority, data = self.tree.get(s_val)
            if data is None:
                continue
            indices.append(idx)
            priorities.append(priority)
            batch.append(data)

        if len(batch) == 0:
            raise RuntimeError("PER buffer sampling failed — buffer may be empty")

        # Importance-sampling weights
        priorities_arr = np.array(priorities, dtype=np.float64)
        probs = priorities_arr / (self.tree.total + 1e-12)
        weights = (len(self) * probs + 1e-12) ** (-beta)
        weights = weights / (weights.max() + 1e-12)  # normalize

        s, a, r, s2, d = zip(*batch)
        return (
            np.stack(s),
            np.array(a, dtype=np.int64),
            np.array(r, dtype=np.float32),
            np.stack(s2),
            np.array(d, dtype=np.float32),
            weights.astype(np.float32),
            indices,
        )

    def update_priorities(self, indices: List[int], td_errors: np.ndarray, epsilon: float = 1e-6) -> None:
        for idx, td in zip(indices, td_errors):
            priority = (abs(float(td)) + epsilon) ** self.alpha
            self._max_priority = max(self._max_priority, priority)
            self.tree.update(idx, priority)


# Legacy uniform replay buffer (kept for backward compat)
class ReplayBuffer:
    def __init__(self, capacity: int, seed: int = 0):
        self.capacity = int(capacity)
        self.rng = random.Random(seed)
        self.buf: Deque[Tuple[np.ndarray, int, float, np.ndarray, bool]] = deque(
            maxlen=self.capacity
        )

    def __len__(self) -> int:
        return len(self.buf)

    def add(self, s: np.ndarray, a: int, r: float, s2: np.ndarray, done: bool) -> None:
        self.buf.append(
            (s.astype(np.float32), int(a), float(r), s2.astype(np.float32), bool(done))
        )

    def sample(
        self, batch_size: int
    ) -> Tuple[np.ndarray, np.ndarray, np.ndarray, np.ndarray, np.ndarray]:
        batch = self.rng.sample(self.buf, k=int(batch_size))
        s, a, r, s2, d = zip(*batch)
        return (
            np.stack(s),
            np.array(a, dtype=np.int64),
            np.array(r, dtype=np.float32),
            np.stack(s2),
            np.array(d, dtype=np.float32),
        )


# ---------------------------------------------------------------------------
# Dueling Double DQN Agent with PER
# ---------------------------------------------------------------------------

class DQNAgent:
    """
    Production-grade Dueling Double DQN Agent with Prioritized Experience Replay.

    Key upgrades:
      1. Dueling Architecture: Q(s,a) = V(s) + A(s,a) - mean(A)
      2. Prioritized Replay: Focus learning on high-error transitions
      3. Double DQN: Decouple selection from evaluation
      4. Input Normalization: Min-Max scaling for stable gradients

    Backward compatible: loads old QNetwork models seamlessly.
    """

    NORM_DENOMS = np.array([12.0, 100.0, 30.0, 50.0, 50.0, 50.0, 200.0], dtype=np.float32)

    def __init__(
        self,
        obs_size: int,
        num_actions: int,
        config: Optional[DQNConfig] = None,
        seed: int = 0,
        device: Optional[str] = None,
        use_dueling: bool = True,
        use_per: bool = True,
    ):
        self.obs_size = int(obs_size)
        self.num_actions = int(num_actions)
        self.cfg = config or DQNConfig()
        self.rng = np.random.default_rng(seed)
        self.use_dueling = use_dueling
        self.use_per = use_per

        if device is None:
            device = "cuda" if torch.cuda.is_available() else "cpu"
        self.device = torch.device(device)

        # Networks — choose architecture
        NetClass = DuelingQNetwork if use_dueling else QNetwork
        self.q = NetClass(self.obs_size, self.num_actions).to(self.device)
        self.target = NetClass(self.obs_size, self.num_actions).to(self.device)
        self.target.load_state_dict(self.q.state_dict())
        self.target.eval()

        self.optim = optim.Adam(self.q.parameters(), lr=self.cfg.lr)

        # Replay buffer — choose type
        if use_per:
            self.replay = PrioritizedReplayBuffer(
                self.cfg.replay_size, alpha=self.cfg.per_alpha, seed=seed
            )
        else:
            self.replay = ReplayBuffer(self.cfg.replay_size, seed=seed)

        self.train_steps: int = 0
        self._epsilon_value: float = float(self.cfg.epsilon_start)
        self.episodes_seen: int = 0
        self._beta: float = float(self.cfg.per_beta_start)

    # --- Pipeline Steps ---

    def preprocess_state(self, obs: np.ndarray) -> torch.Tensor:
        """Normalize raw observation to [0, 1] range."""
        norm_obs = obs.astype(np.float32) / self.NORM_DENOMS
        return torch.tensor(norm_obs, dtype=torch.float32, device=self.device)

    def select_action(self, obs: np.ndarray, greedy: bool = False) -> int:
        """Epsilon-greedy action selection on the main network."""
        if (not greedy) and (self.rng.random() < self.epsilon()):
            return int(self.rng.integers(0, self.num_actions))
        with torch.no_grad():
            q_values = self.predict_q_values(obs)
            return int(np.argmax(q_values))

    def predict_q_values(self, obs: np.ndarray) -> np.ndarray:
        """Return raw Q-values for XAI transparency."""
        with torch.no_grad():
            x = self.preprocess_state(obs).unsqueeze(0)
            q_values = self.q(x).squeeze(0)
            return q_values.cpu().numpy()

    # --- Training Logic ---

    def train_step(self) -> Dict[str, float]:
        """
        Single training update with Dueling DDQN + PER.
        """
        if not self.can_train():
            return {"loss": float("nan")}

        if self.use_per:
            # Anneal beta toward 1.0
            self._beta = min(
                self.cfg.per_beta_end,
                self.cfg.per_beta_start + (self.cfg.per_beta_end - self.cfg.per_beta_start)
                * self.train_steps / max(1, self.cfg.per_beta_anneal_steps)
            )
            s, a, r, s2, d, weights, indices = self.replay.sample(
                self.cfg.batch_size, beta=self._beta
            )
            w_t = torch.tensor(weights, dtype=torch.float32, device=self.device).unsqueeze(-1)
        else:
            s, a, r, s2, d = self.replay.sample(self.cfg.batch_size)
            w_t = torch.ones(self.cfg.batch_size, 1, device=self.device)
            indices = None

        # Preprocess
        s_t = self.preprocess_state(s)
        s2_t = self.preprocess_state(s2)
        a_t = torch.tensor(a, dtype=torch.int64, device=self.device).unsqueeze(-1)
        r_t = torch.tensor(r, dtype=torch.float32, device=self.device).unsqueeze(-1)
        d_t = torch.tensor(d, dtype=torch.float32, device=self.device).unsqueeze(-1)

        # Current Q-values
        q_sa = self.q(s_t).gather(1, a_t)

        # Double DQN target
        with torch.no_grad():
            next_actions = self.q(s2_t).argmax(dim=1, keepdim=True)
            q_target_next = self.target(s2_t).gather(1, next_actions)
            target_val = r_t + (1.0 - d_t) * self.cfg.gamma * q_target_next

        # TD errors for PER priority update
        td_errors = (q_sa - target_val).detach()

        # Weighted loss
        elementwise_loss = nn.functional.smooth_l1_loss(q_sa, target_val, reduction='none')
        loss = (w_t * elementwise_loss).mean()

        self.optim.zero_grad(set_to_none=True)
        loss.backward()
        nn.utils.clip_grad_norm_(self.q.parameters(), self.cfg.max_grad_norm)
        self.optim.step()

        # Update PER priorities
        if self.use_per and indices is not None:
            self.replay.update_priorities(
                indices,
                td_errors.squeeze(-1).cpu().numpy(),
                epsilon=self.cfg.per_epsilon,
            )

        # Housekeeping
        self.train_steps += 1
        self._epsilon_value = max(
            float(self.cfg.epsilon_end),
            float(self._epsilon_value) * float(self.cfg.epsilon_decay_mult),
        )
        if self.train_steps % self.cfg.target_update_every == 0:
            self.target.load_state_dict(self.q.state_dict())

        return {
            "loss": float(loss.item()),
            "epsilon": float(self.epsilon()),
            "avg_q": float(q_sa.mean().item()),
        }

    # --- Helpers ---

    def act(self, obs: np.ndarray, greedy: bool = False) -> int:
        """Legacy helper wrapping select_action."""
        return self.select_action(obs, greedy=greedy)

    def observe(self, s: np.ndarray, a: int, r: float, s2: np.ndarray, done: bool) -> None:
        self.replay.add(s, a, r, s2, done)

    def can_train(self) -> bool:
        return len(self.replay) >= self.cfg.min_replay_size

    def epsilon(self) -> float:
        return float(self._epsilon_value)

    def on_episode_end(self) -> None:
        self.episodes_seen += 1

    def save(self, path: str) -> None:
        payload = {
            "obs_size": self.obs_size,
            "num_actions": self.num_actions,
            "config": self.cfg.__dict__,
            "state_dict": self.q.state_dict(),
            "norm_denoms": self.NORM_DENOMS.tolist(),
            "architecture": "dueling" if self.use_dueling else "standard",
        }
        torch.save(payload, path)

    @classmethod
    def load(cls, path: str, device: Optional[str] = None) -> "DQNAgent":
        payload = torch.load(path, map_location="cpu", weights_only=False)

        # Detect architecture from saved model
        arch = payload.get("architecture", "standard")  # old models = "standard"
        use_dueling = (arch == "dueling")

        # Filter out PER-specific keys that old configs won't have
        config_dict = {}
        valid_fields = {f.name for f in DQNConfig.__dataclass_fields__.values()}
        for k, v in payload.get("config", {}).items():
            if k in valid_fields:
                config_dict[k] = v

        cfg = DQNConfig(**config_dict)
        agent = cls(
            payload["obs_size"],
            payload["num_actions"],
            cfg,
            seed=0,
            device=device,
            use_dueling=use_dueling,
            use_per=False,  # Don't need PER for inference
        )
        agent.q.load_state_dict(payload["state_dict"])
        agent.target.load_state_dict(payload["state_dict"])
        agent.target.eval()
        return agent