File size: 19,568 Bytes
aa654a4 | 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 | # rl_agent.py
"""
Implements a Deep Q-Network (DQN) Reinforcement Learning agent for Tensorus.
The agent interacts with an environment, stores experiences (S, A, R, S', D)
in TensorStorage, samples experiences, and trains its Q-network.
Note on Experience Storage:
- Large tensors (state, next_state) are stored individually in a 'rl_states' dataset.
- Experience tuples containing IDs of states/next_states and scalar action/reward/done
are stored as metadata in a placeholder tensor within the 'rl_experiences' dataset.
- This approach balances tensor-native storage with manageable metadata, but sampling
requires retrieving linked state tensors, which might be slow depending on storage backend.
"""
from typing import Any
import torch
import torch.nn as nn
import torch.optim as optim
import torch.nn.functional as F
import random
import math
import logging
from typing import Tuple, Optional, Dict, Any
# Import necessary Tensorus components
from tensor_storage import TensorStorage
from dummy_env import DummyEnv # Import our dummy environment
# Configure logging
logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(name)s - %(levelname)s - %(message)s')
logger = logging.getLogger(__name__)
# --- DQN Network Definition ---
class DQN(nn.Module):
"""Simple MLP Q-Network."""
def __init__(self, n_observations: int, n_actions: int, hidden_size: int = 128):
super(DQN, self).__init__()
self.layer1 = nn.Linear(n_observations, hidden_size)
self.layer2 = nn.Linear(hidden_size, hidden_size)
self.layer3 = nn.Linear(hidden_size, n_actions)
def forward(self, x: torch.Tensor) -> torch.Tensor:
"""Forward pass: returns Q-values for each action."""
# Ensure input is float
if x.dtype != torch.float32:
x = x.float()
x = F.relu(self.layer1(x))
x = F.relu(self.layer2(x))
return self.layer3(x)
# --- RL Agent Class ---
class RLAgent:
"""DQN Agent interacting with TensorStorage."""
def __init__(self,
tensor_storage: TensorStorage,
state_dim: int,
action_dim: int,
hidden_size: int = 128,
lr: float = 1e-4,
gamma: float = 0.99,
epsilon_start: float = 0.9,
epsilon_end: float = 0.05,
epsilon_decay: int = 10000,
target_update_freq: int = 500,
batch_size: int = 128,
experience_dataset: str = "rl_experiences",
state_dataset: str = "rl_states"):
"""
Initializes the RL Agent.
Args:
tensor_storage: Instance of TensorStorage.
state_dim: Dimensionality of the environment state.
action_dim: Number of discrete actions.
hidden_size: Hidden layer size for the DQN.
lr: Learning rate for the optimizer.
gamma: Discount factor for future rewards.
epsilon_*: Epsilon-greedy exploration parameters.
target_update_freq: How often (in steps) to update the target network.
batch_size: Number of experiences to sample for training.
experience_dataset: Name of the dataset to store experience metadata.
state_dataset: Name of the dataset to store state/next_state tensors.
"""
if not isinstance(tensor_storage, TensorStorage):
raise TypeError("tensor_storage must be an instance of TensorStorage")
self.tensor_storage = tensor_storage
self.state_dim = state_dim
self.action_dim = action_dim
self.gamma = gamma
self.batch_size = batch_size
self.epsilon_start = epsilon_start
self.epsilon_end = epsilon_end
self.epsilon_decay = epsilon_decay
self.target_update_freq = target_update_freq
self.experience_dataset = experience_dataset
self.state_dataset = state_dataset
# Ensure datasets exist
for ds_name in [self.experience_dataset, self.state_dataset]:
try:
self.tensor_storage.get_dataset(ds_name)
except ValueError:
logger.info(f"Dataset '{ds_name}' not found. Creating it.")
self.tensor_storage.create_dataset(ds_name)
# Device configuration (use GPU if available)
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
logger.info(f"RL Agent using device: {self.device}")
# Networks
self.policy_net = DQN(state_dim, action_dim, hidden_size).to(self.device)
self.target_net = DQN(state_dim, action_dim, hidden_size).to(self.device)
self.target_net.load_state_dict(self.policy_net.state_dict())
self.target_net.eval() # Target network is only for inference
# Optimizer
self.optimizer = optim.AdamW(self.policy_net.parameters(), lr=lr, amsgrad=True)
self.steps_done = 0 # Counter for epsilon decay and target updates
def select_action(self, state: torch.Tensor) -> int:
"""Selects an action using epsilon-greedy strategy."""
sample = random.random()
# Calculate current epsilon
eps_threshold = self.epsilon_end + (self.epsilon_start - self.epsilon_end) * \
math.exp(-1. * self.steps_done / self.epsilon_decay)
self.steps_done += 1 # Increment step counter here
if sample > eps_threshold:
# Exploit: choose the best action from the policy network
with torch.no_grad():
# Ensure state is on the correct device and has batch dimension
state = state.unsqueeze(0).to(self.device) if state.ndim == 1 else state.to(self.device)
q_values = self.policy_net(state)
action = q_values.max(1)[1].view(1, 1).item() # Get index of max Q value
logger.debug(f"Exploiting: Q-Values={q_values.cpu().numpy()}, Chosen Action={action}")
return action
else:
# Explore: choose a random action
action = random.randrange(self.action_dim)
logger.debug(f"Exploring: Chosen Action={action}")
return action
def store_experience(self, state: torch.Tensor, action: int, reward: float, next_state: Optional[torch.Tensor], done: bool) -> None:
"""Stores an experience tuple in TensorStorage."""
if state is None:
logger.error("Cannot store experience with None state.")
return
# 1. Store state tensor
state_id = self.tensor_storage.insert(self.state_dataset, state.cpu(), metadata={"component": "state"})
# 2. Store next_state tensor (if not None)
next_state_id = None
if next_state is not None:
next_state_id = self.tensor_storage.insert(self.state_dataset, next_state.cpu(), metadata={"component": "next_state"})
# 3. Create experience metadata
experience_metadata = {
"state_id": state_id,
"action": action, # Store action directly (assuming discrete & scalar)
"reward": reward, # Store reward directly
"next_state_id": next_state_id, # Can be None if done
"done": int(done) # Store boolean as int
}
# 4. Store placeholder tensor with experience metadata
# Using a small tensor as the primary data for this record
placeholder_tensor = torch.tensor([1.0])
exp_record_id = self.tensor_storage.insert(self.experience_dataset, placeholder_tensor, experience_metadata)
logger.debug(f"Stored experience record {exp_record_id}: state_id={state_id}, action={action}, reward={reward:.2f}, next_state_id={next_state_id}, done={done}")
def optimize_model(self) -> None:
"""Performs one step of optimization on the policy network."""
# Check if enough samples are available in the experience dataset
try:
# A bit inefficient to get the full list just for the count,
# TensorStorage could be enhanced with a count method.
experience_count = len(self.tensor_storage.get_dataset(self.experience_dataset))
except ValueError:
experience_count = 0
if experience_count < self.batch_size:
logger.debug(f"Not enough experiences ({experience_count}/{self.batch_size}) to optimize yet.")
return
# 1. Sample experience metadata records
try:
sampled_metadata_records = self.tensor_storage.sample_dataset(self.experience_dataset, self.batch_size)
except ValueError:
logger.error(f"Could not sample from dataset {self.experience_dataset}")
return
except Exception as e:
logger.error(f"Error sampling experiences: {e}", exc_info=True)
return
# 2. Retrieve actual state/next_state tensors based on IDs in metadata
states = []
actions = []
rewards = []
next_states = []
dones = []
non_final_mask_list = [] # Keep track of which next_states are not None
for record in sampled_metadata_records:
meta = record['metadata']
state_id = meta.get('state_id')
next_state_id = meta.get('next_state_id')
action = meta.get('action')
reward = meta.get('reward')
done_flag = bool(meta.get('done', 1)) # Default to True if missing? Risky. Assume present.
# Basic validation
if state_id is None or action is None or reward is None:
logger.warning(f"Skipping invalid sampled record: {meta}")
continue
# Retrieve state
state_record = self.tensor_storage.get_tensor_by_id(self.state_dataset, state_id)
if state_record is None:
logger.warning(f"Could not find state tensor with ID {state_id} for experience {meta.get('record_id')}. Skipping sample.")
continue
states.append(state_record['tensor'])
# Retrieve next state if it exists (i.e., not a terminal state)
current_next_state = None
if not done_flag and next_state_id:
next_state_record = self.tensor_storage.get_tensor_by_id(self.state_dataset, next_state_id)
if next_state_record:
current_next_state = next_state_record['tensor']
else:
# This shouldn't happen if storage is consistent, but handle it
logger.warning(f"Could not find next_state tensor with ID {next_state_id} for non-terminal experience {meta.get('record_id')}. Treating as terminal.")
done_flag = True # Treat as done if next state is missing
next_states.append(current_next_state) # Will be None for terminal states
non_final_mask_list.append(not done_flag)
actions.append(torch.tensor([[action]], dtype=torch.long)) # Action needs to be [[action]] for gather()
rewards.append(torch.tensor([reward], dtype=torch.float32))
dones.append(done_flag) # Keep Python bool for now, convert later if needed
# If not enough valid samples were retrieved after lookup
if not states:
logger.warning("No valid samples retrieved after state lookup. Optimization step skipped.")
return
# 3. Batch the retrieved data
# Filter out None values in next_states for target Q calculation
non_final_next_states = torch.cat([ns for ns in next_states if ns is not None]).to(self.device) if any(non_final_mask_list) else None
state_batch = torch.cat(states).to(self.device)
action_batch = torch.cat(actions).to(self.device)
reward_batch = torch.cat(rewards).to(self.device)
non_final_mask = torch.tensor(non_final_mask_list, dtype=torch.bool, device=self.device)
# 4. Compute Q(s_t, a) - the model computes Q(s_t), then we select the columns of actions taken
# Ensure state_batch has correct dimensions if needed (e.g., B x C x H x W for images)
# Our dummy env state is simple (B x 1)
if state_batch.ndim == 1: # Ensure batch dimension exists
state_batch = state_batch.unsqueeze(-1)
state_action_values = self.policy_net(state_batch).gather(1, action_batch)
# 5. Compute V(s_{t+1}) for all next states.
# Expected values of actions for non_final_next_states are computed based
# on the "older" target_net; selecting their best reward with max(1)[0].
next_state_values = torch.zeros(len(states), device=self.device) # Start with zeros for all
if non_final_next_states is not None and non_final_next_states.numel() > 0:
with torch.no_grad():
if non_final_next_states.ndim == 1: # Ensure batch dimension
non_final_next_states = non_final_next_states.unsqueeze(-1)
next_state_values[non_final_mask] = self.target_net(non_final_next_states).max(1)[0]
# 6. Compute the expected Q values (Bellman equation)
expected_state_action_values = (next_state_values * self.gamma) + reward_batch
# 7. Compute loss
criterion = nn.SmoothL1Loss() # Huber loss
loss = criterion(state_action_values, expected_state_action_values.unsqueeze(1)) # Ensure target has same shape
# 8. Optimize the model
self.optimizer.zero_grad()
loss.backward()
# In-place gradient clipping (prevents exploding gradients)
torch.nn.utils.clip_grad_value_(self.policy_net.parameters(), 100)
self.optimizer.step()
logger.debug(f"Optimization step done. Loss: {loss.item():.4f}")
# 9. Periodically update the target network
if self.steps_done % self.target_update_freq == 0:
self._update_target_network()
def _update_target_network(self):
"""Copies weights from policy_net to target_net."""
logger.info(f"Updating target network at step {self.steps_done}")
self.target_net.load_state_dict(self.policy_net.state_dict())
def train(self, env: DummyEnv, num_episodes: int):
"""Runs the training loop for a number of episodes."""
logger.info(f"--- Starting Training for {num_episodes} episodes ---")
episode_rewards = []
for i_episode in range(num_episodes):
state = env.reset()
# state should be a tensor from env.reset()
# state = torch.tensor(state, dtype=torch.float32, device=self.device).unsqueeze(0) # Add batch dim
# Note: env.reset now returns a tensor, no need to convert here
done = False
current_episode_reward = 0
steps_in_episode = 0
while not done:
# Select action
action = self.select_action(state) # state is already a tensor
# Perform action in environment
next_state, reward, done, _ = env.step(action) # next_state is a tensor
current_episode_reward += reward
steps_in_episode += 1
# Store experience in TensorStorage
# Note: store_experience handles moving tensors to CPU for storage if needed
self.store_experience(state, action, reward, next_state if not done else None, done)
# Move to the next state
state = next_state
# Perform one step of the optimization (on the policy network)
self.optimize_model()
# Check if env forces done (e.g. max steps reached in DummyEnv)
if done:
break # Exit loop if env says done
episode_rewards.append(current_episode_reward)
logger.info(f"Episode {i_episode+1}/{num_episodes} finished after {steps_in_episode} steps. Reward: {current_episode_reward:.2f}. Epsilon: {self.epsilon_end + (self.epsilon_start - self.epsilon_end) * math.exp(-1. * self.steps_done / self.epsilon_decay):.3f}")
# Optional: Add plotting or saving logic here
if (i_episode + 1) % 50 == 0: # Log average reward periodically
avg_reward = sum(episode_rewards[-50:]) / len(episode_rewards[-50:])
logger.info(f" Average reward over last 50 episodes: {avg_reward:.2f}")
logger.info("--- Training Finished ---")
return episode_rewards
# --- Example Usage ---
if __name__ == "__main__":
logger.info("--- Starting RL Agent Example ---")
# 1. Setup TensorStorage
storage = TensorStorage()
# 2. Setup Dummy Environment
env = DummyEnv(max_steps=100) # Episodes are max 100 steps
# 3. Create the RL Agent
agent = RLAgent(
tensor_storage=storage,
state_dim=env.state_dim,
action_dim=env.action_dim,
hidden_size=64, # Smaller network for simple env
lr=5e-4,
gamma=0.95,
epsilon_start=0.95,
epsilon_end=0.05,
epsilon_decay=20000, # Slower decay for demonstration
target_update_freq=200, # Update target net less frequently
batch_size=64, # Smaller batch size
experience_dataset="dummy_env_experiences", # Use specific names
state_dataset="dummy_env_states"
)
# 4. Run the training loop
num_episodes_to_run = 200 # Adjust as needed
rewards = agent.train(env, num_episodes=num_episodes_to_run)
# 5. Optional: Check TensorStorage contents
print("\n--- Checking TensorStorage contents (Sample) ---")
try:
exp_count = len(storage.get_dataset(agent.experience_dataset))
state_count = len(storage.get_dataset(agent.state_dataset))
print(f"Found {exp_count} experience records in '{agent.experience_dataset}'.")
print(f"Found {state_count} state records in '{agent.state_dataset}'.")
if exp_count > 0:
print("\nExample experience record (metadata):")
sample_exp = storage.sample_dataset(agent.experience_dataset, 1)
if sample_exp:
print(sample_exp[0]['metadata'])
state_id = sample_exp[0]['metadata'].get('state_id')
if state_id:
state_rec = storage.get_tensor_by_id(agent.state_dataset, state_id)
if state_rec:
print(f" -> Corresponding state tensor (retrieved): {state_rec['tensor']}")
except ValueError as e:
print(f"Could not retrieve datasets: {e}")
except Exception as e:
print(f"An error occurred checking storage: {e}")
logger.info("--- RL Agent Example Finished ---")
# Optional: Plot rewards
try:
import matplotlib.pyplot as plt
plt.figure(figsize=(10, 5))
plt.plot(rewards)
# Calculate a simple moving average
moving_avg = [sum(rewards[max(0, i-20):i+1])/len(rewards[max(0, i-20):i+1]) for i in range(len(rewards))]
plt.plot(moving_avg, linestyle='--', label='Moving Avg (20 episodes)')
plt.title("Episode Rewards over Time")
plt.xlabel("Episode")
plt.ylabel("Total Reward")
plt.grid(True)
plt.legend()
print("\nPlotting rewards... Close the plot window to exit.")
plt.show()
except ImportError:
print("\nMatplotlib not found. Skipping reward plot. Install with: pip install matplotlib") |