class DQNAgent(): """Interacts with and learns from the environment.""" def __init__(self, state_size, action_size, seed, ddqn=False, dueling=False, priority=False): """Initialize an Agent object. Params ====== state_size (int): dimension of each state action_size (int): dimension of each action seed (int): random seed """ self.state_size = state_size self.action_size = action_size self.seed = random.seed(seed) self.ddqn = ddqn # 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) # Replay memory self.memory = ReplayBuffer(action_size, BUFFER_SIZE, BATCH_SIZE, seed, priority) # 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 experiences = self.memory.sample(get_n=UPDATE_EVERY) self.update_error(experiences, GAMMA) if len(self.memory) > BATCH_SIZE: experiences = self.memory.sample() self.learn(experiences, GAMMA) self.update_error(experiences, GAMMA) def act(self, state, eps=0.): """Returns actions for given state as per current policy. 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(): 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()) else: return random.choice(np.arange(self.action_size)) def learn(self, experiences, gamma): """Update value parameters using given batch of experience tuples. Params ====== experiences (Tuple[torch.Variable]): tuple of (s, a, r, s', done) tuples gamma (float): discount factor """ states, actions, rewards, next_states, dones, _ = experiences ## TODO: compute and minimize the loss "*** YOUR CODE HERE ***" if self.ddqn: old_val = self.qnetwork_local(states).gather(-1, actions) with torch.no_grad(): actions = self.qnetwork_local(next_states).argmax(-1, keepdim=True) maxQ = self.qnetwork_target(next_states).gather(-1, actions) target = rewards + gamma * maxQ * (1 - dones) else: # Normal DQN with torch.no_grad(): maxQ = self.qnetwork_target(next_states).max(-1, keepdim=True)[0] target = rewards + gamma * maxQ * (1 - dones) old_val = self.qnetwork_local(states).gather(-1, actions) self.optimizer.zero_grad() loss = self.qnetwork_local.criterion(old_val, target) loss.backward() self.optimizer.step() # ------------------- update target network ------------------- # self.soft_update(self.qnetwork_local, self.qnetwork_target, TAU) def update_error(self, experiences, gamma): states, actions, rewards, next_states, dones, idx = experiences ## TODO: compute and minimize the loss "*** YOUR CODE HERE ***" if self.ddqn: old_val = self.qnetwork_local(states).gather(-1, actions) with torch.no_grad(): actions = self.qnetwork_local(next_states).argmax(-1, keepdim=True) maxQ = self.qnetwork_target(next_states).gather(-1, actions) target = rewards + gamma * maxQ * (1 - dones) else: # Normal DQN with torch.no_grad(): maxQ = self.qnetwork_target(next_states).max(-1, keepdim=True)[0] target = rewards + gamma * maxQ * (1 - dones) old_val = self.qnetwork_local(states).gather(-1, actions) error = torch.abs(old_val - target).detach().numpy().squeeze() for i, err in zip(idx, error): self.memory.error_buffer[i] = err 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 """ 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 Agent(): def __init__(self, state_size, action_size, seed): ''' Initialization of the agent ''' # Initialize state / action space sizes, and the counter for updating self.state_size = state_size self.action_size = action_size self.seed = random.seed(seed) self.t_step = 0 # Initialize the two Q-networks (the local and target) and define the optimizer 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) # Initialize agent's replay buffer self.memory = ReplayBuffer(action_size, BUFFER_SIZE, BATCH_SIZE, seed) def step(self, state, action, reward, next_state, done): ''' Store experience and learn if it is time to do so ''' # 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 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 ''' state = torch.from_numpy(state).float().unsqueeze(0).to(device) # Set the Q-network to evaluation mode (turn off training layers such as dropouts) and do a forward pass for the state # without computing gradients. Afterwards, return to training mode self.qnetwork_local.eval() with torch.no_grad(): 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()) else: return random.choice(np.arange(self.action_size)) def soft_update(self, local_model, target_model, tau): ''' Soft update for the target Q-network ''' 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) def learn(self, experiences, gamma): ''' Update local network weights using sampled experience tuples (Tuple[torch.Variable]): tuple of (s, a, r, s', done) tuples ''' # Unpack experiences states, actions, rewards, next_states, dones = experiences # Get predictions from local network, i.e. Q-values of the (state, action) pairs in the sampled batch of experiences Q_expected = self.qnetwork_local(states).gather(1, actions) # Get training targets, i.e. current reward + max predicted Q-value for next state by target network, for each (state, action) Q_targets_next = self.qnetwork_target(next_states).detach().max( 1)[0].unsqueeze(1) Q_targets = rewards + (gamma * Q_targets_next * (1 - dones)) # Now compute the loss wrt these new targets loss = self.qnetwork_local.criterion(Q_expected, Q_targets) self.optimizer.zero_grad() # ?! loss.backward() self.optimizer.step() # ------------------- update target network ------------------- # self.soft_update(self.qnetwork_local, self.qnetwork_target, TAU)