File size: 9,732 Bytes
518db7a | 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 | import torch
import torch.nn as nn
import torch.nn.functional as F
from typing import Optional
class MetaController(nn.Module):
"""
Actor-Critic RL Meta-Controller for hypergraph optimization.
Reinforcement learning system that dynamically optimizes hypergraph architecture
during runtime using policy gradient methods. The meta-controller uses an actor-critic
architecture to learn optimal architectural modifications (pruning/creating hyperedges)
that improve overall system performance.
Architecture:
- Input: 16384-dim SDR from Module A (SimHash Encoder)
- Shared network: 2 layers (state_dim → 256 → 256) with GELU activation
- Actor head: Linear(256 → 10) outputs policy logits π(a|s) over 10 meta-actions
- Critic head: Linear(256 → 1) outputs state value estimate V(s)
Meta-Actions (0-indexed):
0. INCREASE_SPARSITY_THRESHOLD - Increase sparsity threshold for hyperedge activation
1. DECREASE_SPARSITY_THRESHOLD - Decrease sparsity threshold for hyperedge activation
2. PRUNE_WEAKEST_EDGE - Remove hyperedge with lowest weight
3. CREATE_RANDOM_EDGE - Add new hyperedge connecting random nodes
4. MERGE_SIMILAR_EDGES - Combine hyperedges with overlapping node sets
5. SPLIT_DENSE_EDGE - Divide hyperedge with many nodes into two smaller edges
6. BOOST_ACH - Boost acetylcholine (attention neuromodulator)
7. BOOST_NE - Boost norepinephrine (arousal neuromodulator)
8. TRIGGER_SLEEP - Trigger sleep consolidation mechanism
9. NO_OP - No operation, continue with current configuration
Reference:
- Sutton & Barto (2018) - Reinforcement Learning: An Introduction
- Chapter 13.5: Actor-Critic Methods
"""
# Meta-Actions: Dictionary mapping action indices to (name, description) tuples
META_ACTIONS = {
0: (
"INCREASE_SPARSITY_THRESHOLD",
"Increase sparsity threshold for hyperedge activation",
),
1: (
"DECREASE_SPARSITY_THRESHOLD",
"Decrease sparsity threshold for hyperedge activation",
),
2: ("PRUNE_WEAKEST_EDGE", "Remove hyperedge with lowest weight"),
3: ("CREATE_RANDOM_EDGE", "Add new hyperedge connecting random nodes"),
4: ("MERGE_SIMILAR_EDGES", "Combine hyperedges with overlapping node sets"),
5: (
"SPLIT_DENSE_EDGE",
"Divide hyperedge with many nodes into two smaller edges",
),
6: ("BOOST_ACH", "Boost acetylcholine (attention neuromodulator)"),
7: ("BOOST_NE", "Boost norepinephrine (arousal neuromodulator)"),
8: ("TRIGGER_SLEEP", "Trigger sleep consolidation mechanism"),
9: ("NO_OP", "No operation, continue with current configuration"),
}
def __init__(
self,
state_dim: int = 16384,
hidden_dim: int = 256,
num_actions: int = 10,
device: Optional[str] = None,
) -> None:
"""
Initialize MetaController with actor-critic architecture.
Args:
state_dim: Dimensionality of input state (SDR from Module A)
hidden_dim: Size of hidden layers in shared network
num_actions: Number of meta-actions (fixed at 10)
device: Device to run on ('cuda' or 'cpu'). If None, auto-detects GPU.
"""
super().__init__()
assert num_actions == 10, f"num_actions must be 10, got {num_actions}"
self.state_dim = state_dim
self.hidden_dim = hidden_dim
self.num_actions = num_actions
# Auto-detect device if not specified
if device is None:
device = "cuda" if torch.cuda.is_available() else "cpu"
self.device = torch.device(device)
# Shared feature extraction (2 layers with GELU)
self.shared = nn.Sequential(
nn.Linear(state_dim, hidden_dim),
nn.GELU(),
nn.Linear(hidden_dim, hidden_dim),
nn.GELU(),
)
# Actor head: policy π(a|s) - outputs logits over actions
self.actor = nn.Linear(hidden_dim, num_actions)
# Critic head: value V(s) - outputs scalar state value
self.critic = nn.Linear(hidden_dim, 1)
# Move to device
self.to(self.device)
def forward(self, state: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
"""
Forward pass through actor-critic network.
Args:
state: Input state tensor of shape (state_dim,) or (batch, state_dim)
Returns:
logits: Action logits of shape (num_actions,) or (batch, num_actions)
value: State value of shape () or (batch,) - note: scalar/squeezed output
Example:
>>> controller = MetaController()
>>> state = torch.randn(16384)
>>> logits, value = controller(state)
>>> logits.shape
torch.Size([10])
>>> value.shape
torch.Size([])
"""
# Pass through shared network
features = self.shared(state)
# Actor: policy logits
logits = self.actor(features)
# Critic: value estimate (squeeze to remove last dimension)
value = self.critic(features).squeeze(-1)
return logits, value
def select_action(
self, state: torch.Tensor, training: bool = True
) -> tuple[int, float]:
"""
Sample action from policy during training, greedy during evaluation.
Args:
state: torch.Tensor [state_dim] - Single state (not batched)
training: bool - If True, sample from policy; if False, use argmax
Returns:
action: int - Selected action index (0-9)
value: float - Predicted value of the state
Example:
>>> controller = MetaController()
>>> state = torch.randn(16384)
>>> action, value = controller.select_action(state, training=True)
>>> assert 0 <= action < 10
>>> assert isinstance(value, float)
"""
logits, value = self.forward(state)
safe_logits = torch.clamp(logits, -50.0, 50.0)
probs = F.softmax(safe_logits, dim=-1)
probs = torch.where(torch.isfinite(probs), probs, torch.zeros_like(probs))
probs = probs + 1e-8
total = probs.sum()
if not torch.isfinite(total) or total <= 0:
probs = torch.ones_like(probs) / probs.numel()
else:
probs = probs / total
if training:
# Sample from policy (exploration)
action = int(torch.multinomial(probs, 1).item())
else:
# Greedy (exploitation)
action = int(probs.argmax().item())
return action, float(value.item())
def compute_loss(
self,
state: torch.Tensor,
action: int,
reward: float,
old_value: float,
gamma: float = 0.99,
) -> torch.Tensor:
"""
Compute actor-critic loss with one-step TD error.
Loss = actor_loss + 0.5 * critic_loss
Args:
state: Current state (state_dim,) - SDR from Module A
action: Action taken (int, 0-9) - Selected meta-action index
reward: Reward received (float) - Performance improvement signal
old_value: Value estimated before taking action (float) - Baseline for advantage
gamma: Discount factor (not used in one-step version)
Returns:
loss: Combined actor + critic loss (scalar tensor)
Example:
>>> controller = MetaController()
>>> state = torch.randn(16384)
>>> action = 2
>>> reward = 1.0
>>> old_value = 0.5
>>> loss = controller.compute_loss(state, action, reward, old_value)
>>> assert loss.ndim == 0 # Scalar loss
>>> assert not torch.isnan(loss) # No NaN
"""
logits, value = self.forward(state)
# TD error (advantage) - one-step temporal difference
td_error = reward - old_value
# Critic loss: minimize squared TD error
critic_loss = td_error**2
# Actor loss: policy gradient with advantage
log_probs = F.log_softmax(logits, dim=-1)
actor_loss = -log_probs[action] * td_error
# Combined loss (weight critic by 0.5, standard practice)
loss = actor_loss + 0.5 * critic_loss
return loss
def execute_action(self, action: int, hypergraph: Optional[object]) -> None:
"""
Execute a meta-action on the hypergraph (Module B).
Stub implementation - will be completed in later subtasks.
Applies structural modifications to the hypergraph based on the selected
meta-action (e.g., pruning edges, adjusting sparsity, boosting neuromodulators).
Args:
action: Action index (0-9) - Must be valid meta-action from META_ACTIONS
hypergraph: HypergraphManifold instance from Module B (None allowed for testing)
Returns:
None
Raises:
ValueError: If action index is out of valid range [0-9]
Example:
>>> controller = MetaController()
>>> result = controller.execute_action(2, None) # PRUNE_WEAKEST_EDGE
>>> assert result is None # Stub returns None
"""
if not 0 <= action < self.num_actions:
raise ValueError(
f"Invalid action {action}. Must be in range [0, {self.num_actions - 1}]"
)
# Stub: actual implementation will modify hypergraph based on action
# Will be implemented in subtask-3-3 through subtask-3-12
return None
|