DQN_PER.py

from IPython.core.debugger import set_trace
import numpy as np
import random
from collections import namedtuple, deque
from model import QNetwork

import torch
import torch.nn.functional as F
import torch.optim as optim

BUFFER_SIZE = int(1e5)     # replay buffer size
BATCH_SIZE = 64         # minibatch size
GAMMA = 0.995            # discount factor
TAU = 1e-2              # for soft update of target parameters
LR = 4.8e-4              # learning rate 
UPDATE_EVERY = 4        # how often to update the network

device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")

class Agent():
    """Interacts with and learns from the environment."""

    def __init__(self, state_size, action_size, seed, lr_decay=0.9999):
        """Initialize an Agent object.
        
        Params
        ======
            state_size (int): dimension of each state
            action_size (int): dimension of each action
            lr_decay (float): multiplicative factor of learning rate decay
            seed (int): random seed
        """
        self.state_size = state_size
        self.action_size = action_size
        self.seed = random.seed(seed)

        # Q-Network
        self.qnetwork_local = QNetwork(state_size, action_size, seed).to(device)
        self.qnetwork_target = QNetwork(state_size, action_size, seed).to(device)
        self.optimizer = optim.Adam(self.qnetwork_local.parameters(), lr=LR)   #only update the local network parameters
        #self.lr_scheduler = optim.lr_scheduler.ExponentialLR(self.optimizer, lr_decay)
         
        # prioritized Replay memory
        self.memory = PrioritizedReplayBuffer(action_size, BUFFER_SIZE, BATCH_SIZE, seed, device,
                                                  alpha=0.6, beta=0.4, beta_scheduler=1.0)
        # Initialize time step (for updating every UPDATE_EVERY steps)
        self.t_step = 0
    
    def step(self, state, action, reward, next_state, done):
        # Save experience in replay memory
        self.memory.add(state, action, reward, next_state, done)
        
        # Learn every UPDATE_EVERY time steps.
        self.t_step = (self.t_step + 1) % UPDATE_EVERY
        if self.t_step == 0:
        # If enough samples are available in memory, get random subset and learn
            if len(self.memory) > BATCH_SIZE:
                experiences = self.memory.sample()
                self.learn(experiences, GAMMA)

    def act(self, state, eps=0.):
        """Returns actions for given state as per current policy and Q value.
        
        Params
        ======
            state (array_like): current state
            eps (float): epsilon, for epsilon-greedy action selection
        """
        state = torch.from_numpy(state).float().unsqueeze(0).to(device)
        self.qnetwork_local.eval()
        with torch.no_grad():
            #set_trace()
            action_values = self.qnetwork_local(state)
        self.qnetwork_local.train()

        # Epsilon-greedy action selection
        if random.random() > eps:
            return np.argmax(action_values.cpu().data.numpy()),  np.max(action_values.cpu().data.numpy())
        else:
            #set_trace()
            random_action = random.choice(np.arange(self.action_size))
            action_values = action_values.cpu().data.numpy()
            return random_action, action_values[0,0,random_action]
        
    def learn(self, experiences, gamma):
        """Update value parameters using given batch of experience tuples.

        Params
        ======
            experiences (Tuple[torch.Tensor]): tuple of (s, a, r, s', done, w) tuples  
            gamma (float): discount factor
        """
        states, actions, rewards, next_states, dones, w = experiences
            
        # Get max predicted Q values (for next states) from target model
        Q_targets_next = self.qnetwork_target(next_states).detach().max(1)[0].unsqueeze(1)
        # Compute Q targets for current states
        Q_targets = rewards + (gamma * Q_targets_next * (1 - dones))
        #Get expected Q values from local model
        Q_expected = self.qnetwork_local(states).gather(1, actions)
        # Compute loss
        Q_targets.sub_(Q_expected)
        Q_targets.squeeze_()
        Q_targets.pow_(2)
        with torch.no_grad():
            TD_error = Q_targets.detach()
            TD_error.pow_(0.5)
            self.memory.update_priorities(TD_error)            
        Q_targets.mul_(w)
        loss = Q_targets.mean()        
        # Minimize the loss
        self.optimizer.zero_grad()
        loss.backward()
        self.optimizer.step()
        #self.lr_scheduler.step()
    
        # ------------------- update target network ------------------- #
        self.soft_update(self.qnetwork_local, self.qnetwork_target, TAU)    #updating the target network parameters                    

    def soft_update(self, local_model, target_model, tau):
        """Soft update model parameters.
        θ_target = τ*θ_local + (1 - τ)*θ_target

        Params
        ======
            local_model (PyTorch model): weights will be copied from
            target_model (PyTorch model): weights will be copied to
            tau (float): interpolation parameter 
        """
        #self.qnetwork_local.eval()
        #self.qnetwork_target.eval()
        for target_param, local_param in zip(target_model.parameters(), local_model.parameters()):
            target_param.data.copy_(tau*local_param.data + (1.0-tau)*target_param.data)

class PrioritizedReplayBuffer:
    """Fixed-size prioritized buffer to store experience tuples."""
    
    def __init__(self, action_size, buffer_size, batch_size, seed, device, alpha=0., beta=1., beta_scheduler=1.):
        """Initialize a PrioritizedReplayBuffer object.
        Params
        ======
            action_size (int): dimension of each action
            buffer_size (int): maximum size of buffer
            batch_size (int): size of each training batch
            seed (int): random seed
            alpha (float): determines how much prioritization is used; α = 0 corresponding to the uniform case
            beta (float): amount of importance-sampling correction; β = 1 fully compensates for the non-uniform probabilities
            beta_scheduler (float): multiplicative factor (per sample) for increasing beta (should be >= 1.0)
        """
        
        self.action_size = action_size
        self.buffer_size = buffer_size
        self.batch_size = batch_size
        self.seed = np.random.seed(seed)
        self.device = device
        self.alpha = alpha
        self.beta = beta
        self.beta_scheduler = beta_scheduler
        
        # Create a Numpy Array to store namedtuples of experience
        self.memory = np.empty(buffer_size, dtype=[
            ("state", np.ndarray),
            ("action", np.int),
            ("reward", np.float),
            ("next_state", np.ndarray),
            ("done", np.bool),
            ('prob', np.double)]) # sel.memory = [(s,a,r,s',d,w),(s,a,r,s',d,w)]
        # Variable to control the memory buffer as being a circular list
        self.memory_idx_ctrl = 0
        
        # Variable to control the selected samples
        self.memory_samples_idx = np.empty(batch_size)
        # Numpy Array to store selected samples
        # Those samples could be controlled only by the index,
        # however keeping an allocated space in memory improves performance.
        # (Here we have a tradeoff between memory space and cumputacional processing)
        self.memory_samples = np.empty(batch_size, dtype=type(self.memory))

        # Each new experience is added to the memory with
        # the maximum probability of being choosen
        self.max_prob = 0.0001
        
        # Value to a non-zero probability
        self.nonzero_probability = 0.00001
        
        # Numpy Arrays to store probabilities and weights
        # (tradeoff between memory space and cumputacional processing)
        self.p = np.empty(buffer_size, dtype=np.double)
        self.w = np.empty(buffer_size, dtype=np.double)
    
    def add(self, state, action, reward, next_state, done):
        """Add a new experience to memory."""
        
        # Add the experienced parameters to the memory
        self.memory[self.memory_idx_ctrl]['state'] = state
        self.memory[self.memory_idx_ctrl]['action'] = action
        self.memory[self.memory_idx_ctrl]['reward'] = reward
        self.memory[self.memory_idx_ctrl]['next_state'] = next_state
        self.memory[self.memory_idx_ctrl]['done'] = done
        self.memory[self.memory_idx_ctrl]['prob'] = self.max_prob
        
        # Control memory as a circular list
        self.memory_idx_ctrl = (self.memory_idx_ctrl + 1) % self.buffer_size
    
    def sample(self):
        """Sample a batch of prioritized experiences from memory."""
        # Normalize the probability of being chosen for each one of the memory registers
        np.divide(self.memory['prob'], self.memory['prob'].sum(), out=self.p) # p = (p)/(sum of p)
        # Choose "batch_size" sample index following the defined probability
        self.memory_samples_idx = np.random.choice(self.buffer_size, self.batch_size, replace=False, p=self.p)
        # Get the samples from memory
        self.memory_samples = self.memory[self.memory_samples_idx]
        
        # Compute importance-sampling weights for each one of the memory registers
        # w = ((N * P) ^ -β) / max(w)
        np.multiply(self.memory['prob'], self.buffer_size, out=self.w)
        np.power(self.w, -self.beta, out=self.w, where=self.w!=0) # condition to avoid division by zero
        np.divide(self.w, self.w.max(), out=self.w) # normalize the weights
        
        self.beta = min(1, self.beta*self.beta_scheduler)
        
        # Split data into new variables
        states = torch.from_numpy(np.vstack(self.memory_samples['state'])).float().to(self.device)
        actions = torch.from_numpy(np.vstack(self.memory_samples['action'])).long().to(self.device)
        rewards = torch.from_numpy(np.vstack(self.memory_samples['reward'])).float().to(self.device)
        next_states = torch.from_numpy(np.vstack(self.memory_samples['next_state'])).float().to(self.device)
        dones = torch.from_numpy(np.vstack(self.memory_samples['done'])).float().to(self.device)
        weights = torch.from_numpy(self.w[self.memory_samples_idx]).float().to(self.device)
        
        return (states, actions, rewards, next_states, dones, weights)
    
    def update_priorities(self, td_error):
        # Balance the prioritization using the alpha value
        td_error.pow_(self.alpha)

        # Guarantee a non-zero probability
        td_error.add_(self.nonzero_probability)
        
        #convert cuda tensor to numpy array
        td_error = td_error.cpu().data.numpy()        
        
        # Update the probabilities in memory
        self.memory_samples['prob'] = td_error
        self.memory[self.memory_samples_idx] = self.memory_samples
        
        # Update the maximum probability value
        self.max_prob = self.memory['prob'].max()
        
       
    def __len__(self):
        """Return the current size of internal memory."""
        return self.buffer_size if self.memory_idx_ctrl // self.buffer_size > 0 else self.memory_idx_ctrl