File size: 10,650 Bytes
9f8cf99
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch
import torch.nn as nn
import torch.nn.functional as F
from typing import Any, Tuple, Optional
import numpy as np

from .titans import TitansMemory
from .replay_buffer import ReplayBuffer, Experience


class BaseAgent:
    """Base class for RL agents."""
    def __init__(self):
        pass

    def act(self, state: Any) -> Any:
        raise NotImplementedError

    def update(self, *args, **kwargs):
        raise NotImplementedError

class NestedLearningAgent(BaseAgent):
    """
    Nested Learning Agent for QEC with experience replay and proper RL.

    Implements nested optimization with:
    - Inner loop: Fast updates to decoder weights using policy gradients
    - Outer loop: Slow updates to memory/hyperparameters
    - Experience replay buffer for sample efficiency
    - Actor-critic architecture for stable learning
    """

    def __init__(
        self,
        state_dim: int,
        action_dim: int,
        hidden_dim: int = 64,
        replay_capacity: int = 10000,
        gamma: float = 0.99,
        gae_lambda: float = 0.95,
    ):
        super().__init__()
        self.state_dim = state_dim
        self.action_dim = action_dim
        self.gamma = gamma
        self.gae_lambda = gae_lambda

        # TITANS memory for long-term syndrome history
        self.memory_module = TitansMemory(input_dim=state_dim, hidden_dim=hidden_dim)

        # Actor (policy network)
        self.actor = nn.Sequential(
            nn.Linear(hidden_dim, hidden_dim),
            nn.ReLU(),
            nn.Linear(hidden_dim, action_dim)
        )

        # Critic (value network) for advantage estimation
        self.critic = nn.Sequential(
            nn.Linear(hidden_dim, hidden_dim),
            nn.ReLU(),
            nn.Linear(hidden_dim, 1)
        )

        # Inner loop optimizers (fast)
        self.actor_optimizer = torch.optim.Adam(self.actor.parameters(), lr=1e-3)
        self.critic_optimizer = torch.optim.Adam(self.critic.parameters(), lr=1e-3)

        # Outer loop optimizer (slow)
        self.outer_optimizer = torch.optim.Adam(self.memory_module.parameters(), lr=1e-4)

        # Experience replay buffer
        self.replay_buffer = ReplayBuffer(capacity=replay_capacity, prioritized=True)

        # Training statistics
        self.training_step = 0
        self.episode_rewards = []

    def act(self, state: torch.Tensor, deterministic: bool = False) -> Tuple[int, torch.Tensor]:
        """Select action using current policy.

        Returns:
            Tuple of (action_index, log_probability)
        """
        with torch.no_grad():
            if state.dim() == 1:
                state = state.unsqueeze(0)

            context = self.memory_module(state)
            logits = self.actor(context)
            probs = F.softmax(logits, dim=-1)

            if deterministic:
                action = torch.argmax(probs, dim=-1)
            else:
                action = torch.multinomial(probs, 1).squeeze(-1)

            # Compute log probability
            log_prob = F.log_softmax(logits, dim=-1)
            action_log_prob = log_prob.gather(-1, action.unsqueeze(-1)).squeeze(-1)

        return action.item(), action_log_prob

    def compute_advantages(
        self,
        rewards: torch.Tensor,
        values: torch.Tensor,
        dones: torch.Tensor,
    ) -> Tuple[torch.Tensor, torch.Tensor]:
        """Compute GAE advantages and returns.

        Args:
            rewards: Tensor of rewards (batch_size,)
            values: Tensor of value estimates (batch_size,)
            dones: Tensor of done flags (batch_size,)

        Returns:
            Tuple of (advantages, returns)
        """
        batch_size = rewards.shape[0]
        advantages = torch.zeros_like(rewards)
        returns = torch.zeros_like(rewards)

        # Compute GAE advantages
        next_advantage = 0
        for t in reversed(range(batch_size)):
            if t == batch_size - 1:
                next_value = 0
            else:
                next_value = values[t + 1]

            delta = rewards[t] + self.gamma * next_value * (1 - dones[t]) - values[t]
            advantages[t] = delta + self.gamma * self.gae_lambda * next_advantage * (1 - dones[t])
            next_advantage = advantages[t]

        returns = advantages + values

        return advantages, returns

    def inner_loop(
        self,
        batch_size: int = 32,
        clip_epsilon: float = 0.2,
        value_coef: float = 0.5,
        entropy_coef: float = 0.01,
    ) -> Tuple[float, float, float]:
        """Policy gradient update using experience replay.

        Implements PPO-style clipped surrogate objective with GAE advantages.

        Returns:
            Tuple of (actor_loss, critic_loss, entropy)
        """
        if not self.replay_buffer.is_ready(batch_size):
            return 0.0, 0.0, 0.0

        # Sample from replay buffer
        states, actions, rewards, next_states, dones, weights, indices = self.replay_buffer.sample(batch_size)

        # Prepare states for TITANS (add sequence dimension if needed)
        if states.dim() == 2:
            states = states.unsqueeze(1)
            next_states = next_states.unsqueeze(1)

        # Get memory context
        context = self.memory_module(states.squeeze(1) if states.dim() == 3 else states)

        # Compute current policy logits and values
        logits = self.actor(context)
        values = self.critic(context).squeeze(-1)

        # Compute log probabilities of actions
        log_probs = F.log_softmax(logits, dim=-1)
        action_log_probs = log_probs.gather(-1, actions.unsqueeze(-1)).squeeze(-1)

        # Compute entropy for exploration
        probs = F.softmax(logits, dim=-1)
        entropy = -(probs * log_probs).sum(dim=-1).mean()

        # Compute advantages using GAE
        advantages, returns = self.compute_advantages(rewards, values, dones)

        # Normalize advantages
        advantages = (advantages - advantages.mean()) / (advantages.std() + 1e-8)

        # PPO-style clipped surrogate loss (without the ratio for simplicity, using vanilla PG)
        policy_loss = -(action_log_probs * advantages * weights).mean()

        # Value loss
        value_loss = F.mse_loss(values, returns, reduction='none')
        value_loss = (value_loss * weights).mean()

        # Combined loss
        actor_loss = policy_loss - entropy_coef * entropy
        critic_loss = value_coef * value_loss

        # Update actor
        self.actor_optimizer.zero_grad()
        actor_loss.backward(retain_graph=True)
        torch.nn.utils.clip_grad_norm_(self.actor.parameters(), 0.5)
        self.actor_optimizer.step()

        # Update critic
        self.critic_optimizer.zero_grad()
        critic_loss.backward()
        torch.nn.utils.clip_grad_norm_(self.critic.parameters(), 0.5)
        self.critic_optimizer.step()

        # Update replay buffer priorities (using TD errors as priorities)
        with torch.no_grad():
            td_errors = torch.abs(returns - values).cpu().numpy()
        self.replay_buffer.update_priorities(indices, td_errors)

        self.training_step += 1

        return policy_loss.item(), value_loss.item(), entropy.item()

    def outer_loop(self, states: torch.Tensor, performance_metric: float) -> float:
        """Slow update: Update memory representations and hyperparameters.

        Uses meta-learning signal (performance_metric) to adapt the memory module.
        """
        if states.dim() == 2:
            states = states.unsqueeze(1)

        self.outer_optimizer.zero_grad()

        # Get memory context
        context = self.memory_module(states.squeeze(1) if states.dim() == 3 else states)

        # Meta-learning loss: encourage representations that predict performance
        # Use auxiliary task: predict the performance metric from context
        performance_prediction = self.critic(context).squeeze(-1).mean()

        # Loss encourages accurate performance prediction
        meta_loss = F.mse_loss(performance_prediction, torch.tensor(performance_metric))

        # Add regularization to encourage diverse representations
        context_variance = context.var(dim=0).mean()
        diversity_bonus = -0.01 * context_variance  # Encourage variance

        total_loss = meta_loss + diversity_bonus
        total_loss.backward()

        # Gradient clipping for stability
        torch.nn.utils.clip_grad_norm_(self.memory_module.parameters(), 0.5)
        self.outer_optimizer.step()

        return total_loss.item()

    def store_experience(
        self,
        state: torch.Tensor,
        action: int,
        reward: float,
        next_state: torch.Tensor,
        done: bool,
        priority: Optional[float] = None,
    ) -> None:
        """Store experience in replay buffer with optional priority."""
        experience = Experience(
            state=state if state.dim() > 0 else state.unsqueeze(0),
            action=action,
            reward=reward,
            next_state=next_state if next_state.dim() > 0 else next_state.unsqueeze(0),
            done=done,
        )
        self.replay_buffer.push(experience, priority)

    def get_stats(self) -> dict:
        """Return training statistics."""
        return {
            "training_step": self.training_step,
            "buffer_size": len(self.replay_buffer),
            "episode_rewards": self.episode_rewards[-10:] if self.episode_rewards else [],
        }

    def save_checkpoint(self, path: str) -> None:
        """Save agent state to checkpoint."""
        torch.save({
            "memory_module": self.memory_module.state_dict(),
            "actor": self.actor.state_dict(),
            "critic": self.critic.state_dict(),
            "actor_optimizer": self.actor_optimizer.state_dict(),
            "critic_optimizer": self.critic_optimizer.state_dict(),
            "outer_optimizer": self.outer_optimizer.state_dict(),
            "training_step": self.training_step,
        }, path)

    def load_checkpoint(self, path: str) -> None:
        """Load agent state from checkpoint."""
        checkpoint = torch.load(path, map_location="cpu")
        self.memory_module.load_state_dict(checkpoint["memory_module"])
        self.actor.load_state_dict(checkpoint["actor"])
        self.critic.load_state_dict(checkpoint["critic"])
        self.actor_optimizer.load_state_dict(checkpoint["actor_optimizer"])
        self.critic_optimizer.load_state_dict(checkpoint["critic_optimizer"])
        self.outer_optimizer.load_state_dict(checkpoint["outer_optimizer"])
        self.training_step = checkpoint.get("training_step", 0)