if step > UPDATE_START and step % UPDATE_INTERVAL == 0: # Randomly sample a batch of transitions B = {(s, a, r, s', d)} from D batch = random.sample(D, BATCH_SIZE) batch = { k: torch.cat([d[k] for d in batch], dim=0) for k in batch[0].keys() } # Compute targets y = batch['reward'] + DISCOUNT * (1 - batch['done']) * target_agent( batch['next_state']).max(dim=1)[0] # Update Q-function by one step of gradient descent value_loss = ( agent(batch['state']).gather(1, batch['action']).squeeze(dim=1) - y).pow(2).mean() optimiser.zero_grad() value_loss.backward() optimiser.step() if step > UPDATE_START and step % TARGET_UPDATE_INTERVAL == 0: # Update target network target_agent = create_target_network(agent) if step > UPDATE_START and step % TEST_INTERVAL == 0: agent.eval() total_reward = test(agent) pbar.set_description('Step: %i | Reward: %f' % (step, total_reward)) plot(step, total_reward, 'dqn') agent.train()
class QAgent: def __init__(self, epsilon_start, epsilon_end, epsilon_anneal, nb_actions, learning_rate, gamma, batch_size, replay_memory_size, hidden_size, model_input_size, use_PER, use_ICM): self.device = torch.device( "cuda" if torch.cuda.is_available() else "cpu") self.epsilon_start = epsilon_start self.epsilon_end = epsilon_end self.epsilon_anneal_over_steps = epsilon_anneal self.num_actions = nb_actions self.gamma = gamma self.batch_size = batch_size self.learning_rate = learning_rate self.step_no = 0 self.policy = DQN(hidden_size=hidden_size, inputs=model_input_size, outputs=nb_actions).to(self.device) self.target = DQN(hidden_size=hidden_size, inputs=model_input_size, outputs=nb_actions).to(self.device) self.target.load_state_dict(self.policy.state_dict()) self.target.eval() self.hidden_size = hidden_size self.optimizer = torch.optim.AdamW(self.policy.parameters(), lr=self.learning_rate) self.use_PER = use_PER if use_PER: self.replay = Prioritized_Replay_Memory(replay_memory_size) else: self.replay = Replay_Memory(replay_memory_size) self.loss_function = torch.nn.MSELoss() self.use_ICM = use_ICM if use_ICM: self.icm = ICM(model_input_size, nb_actions) # Get the current epsilon value according to the start/end and annealing values def get_epsilon(self): eps = self.epsilon_end if self.step_no < self.epsilon_anneal_over_steps: eps = self.epsilon_start - self.step_no * \ ((self.epsilon_start - self.epsilon_end) / self.epsilon_anneal_over_steps) return eps # select an action with epsilon greedy def select_action(self, state): self.step_no += 1 if np.random.uniform() > self.get_epsilon(): with torch.no_grad(): return torch.argmax(self.policy(state)).view(1) else: return torch.tensor([random.randrange(self.num_actions)], device=self.device, dtype=torch.long) # update the model according to one step td targets def update_model(self): if self.use_PER: batch_index, batch, ImportanceSamplingWeights = self.replay.sample( self.batch_size) else: batch = self.replay.sample(self.batch_size) batch_tuple = Transition(*zip(*batch)) state = torch.stack(batch_tuple.state) action = torch.stack(batch_tuple.action) reward = torch.stack(batch_tuple.reward) next_state = torch.stack(batch_tuple.next_state) done = torch.stack(batch_tuple.done) self.optimizer.zero_grad() if self.use_ICM: self.icm.optimizer.zero_grad() forward_loss = self.icm.get_forward_loss(state, action, next_state) inverse_loss = self.icm.get_inverse_loss(state, action, next_state) icm_loss = (1 - self.icm.beta) * inverse_loss.mean( ) + self.ICM.beta * forward_loss.mean() td_estimates = self.policy(state).gather(1, action).squeeze() td_targets = reward + (1 - done.float()) * self.gamma * \ self.target(next_state).max(1)[0].detach_() if self.use_PER: loss = (torch.tensor(ImportanceSamplingWeights, device=self.device) * self.loss_function(td_estimates, td_targets) ).sum() * self.loss_function(td_estimates, td_targets) errors = td_estimates - td_targets self.replay.batch_update(batch_index, errors.data.numpy()) else: loss = self.loss_function(td_estimates, td_targets) if self.use_ICM: loss = self.icm.lambda_weight * loss + icm_loss loss.backward() for param in self.policy.parameters(): param.grad.data.clamp_(-1, 1) if self.use_ICM: self.icm.optimizer.step() self.optimizer.step() return loss.item() # set target net parameters to policy net parameters def update_target(self): self.target.load_state_dict(self.policy.state_dict()) # save model def save(self, path, name): dirname = os.path.dirname(__file__) filename = os.path.join(dirname, os.path.join(path, name + ".pt")) torch.save(self.policy.state_dict(), filename) # load a model def load(self, path): dirname = os.path.dirname(__file__) filename = os.path.join(dirname, path) self.policy.load_state_dict(torch.load(filename)) # store experience in replay memory def cache(self, state, action, reward, next_state, done): self.replay.push(state, action, reward, next_state, done)
class DQNAgent(): """Interacts with and learns from the environment.""" def __init__(self, name, state_size, action_size, use_double_dqn=False, use_dueling=False, seed=0, lr_decay=0.9999, use_prioritized_replay=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.name = name self.state_size = state_size self.action_size = action_size self.use_double_dqn = use_double_dqn self.use_dueling = use_dueling self.seed = random.seed(seed) self.use_prioritized_replay = use_prioritized_replay # Q-Network if use_dueling: self.qnetwork_local = DuelingDQN(state_size, action_size, seed).to(device) self.qnetwork_target = DuelingDQN(state_size, action_size, seed).to(device) else: self.qnetwork_local = DQN(state_size, action_size, seed).to(device) self.qnetwork_target = DQN(state_size, action_size, seed).to(device) self.qnetwork_target.eval() self.optimizer = optim.Adam(self.qnetwork_local.parameters(), lr=LR) self.lr_scheduler = optim.lr_scheduler.ExponentialLR(self.optimizer, lr_decay) # Replay memory if self.use_prioritized_replay: self.memory = PrioritizedReplayBuffer(BUFFER_SIZE, seed, alpha=0.2, beta=0.8, beta_scheduler=1.0) else: self.memory = ReplayBuffer(BUFFER_SIZE, seed) # 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(BATCH_SIZE) self.learn(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) # Epsilon-greedy action selection if random.random() > eps: self.qnetwork_local.eval() with torch.no_grad(): action_values = self.qnetwork_local(state) self.qnetwork_local.train() 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.Tensor]): tuple of (s, a, r, s', done) tuples gamma (float): discount factor """ if self.use_prioritized_replay: states, actions, rewards, next_states, dones, indices, weights = experiences else: states, actions, rewards, next_states, dones = experiences with torch.no_grad(): # Get max predicted Q values (for next states) from target model if self.use_double_dqn: best_local_actions = self.qnetwork_local(states).max(1)[1].unsqueeze(1) Q_targets_next = self.qnetwork_target(next_states).gather(1, best_local_actions).max(1)[0].unsqueeze(1) else: 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) if self.use_prioritized_replay: 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) td_error.mul_(weights) self.memory.update_priorities(indices, td_error) Q_targets.mul_(weights) loss = Q_targets.mean() else: # Compute loss loss = F.mse_loss(Q_expected, Q_targets) # 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) 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)
import torch.optim as optim import copy import pickle from utils import * from models import DQN initial_Q = AER_initial_Q() # initial_Q = torch.zeros(n_actions, device=device) policy_net = DQN(recent_k, n_agents, n_actions, initial_Q).to(device) target_net = DQN(recent_k, n_agents, n_actions, initial_Q).to(device) target_net.load_state_dict(policy_net.state_dict()) target_net.eval() optimizer = optim.RMSprop(policy_net.parameters()) Q = torch.zeros(n_actions, n_actions, n_actions, device=device) for i in range(n_actions): for j in range(n_actions): Q[i, j, :] = initial_Q.view(-1) memory = ReplayMemory(MEM_SIZE) heat = torch.zeros(n_agents, n_actions, n_actions, device=device) heat_unique0 = [] heat_freq0 = [] heat_unique1 = [] heat_freq1 = []
class FixedDQNAgent(DQNAgent): """ DQN Agent with a target network to compute Q-targets. Extends DQNAgent. """ def __init__(self, input_dim, output_dim, lr, gamma, max_memory_size, batch_size, eps_start, eps_end, eps_decay, device, target_update=100, linear1_units=64, linear2_units=64, decay_type="linear"): super().__init__(input_dim, output_dim, lr, gamma, max_memory_size, batch_size, eps_start, eps_end, eps_decay, device, linear1_units, linear2_units, decay_type) self.model_name = "FixedDQN" self.target_update_freq = target_update # networks self.output_dim = output_dim self.target_net = DQN(input_dim, output_dim, linear1_units, linear2_units).to(device) self.target_net.load_state_dict(self.policy_net.state_dict()) self.target_net.eval() self.updated = 0 def learn(self): """ Update the weights of the network, using target_net to compute Q-targets. Every self.target_update_freq updates, clone the policy_net. :return: the loss """ states, next_states, actions, rewards, dones = self.memory.sample( self.batch_size) curr_q_vals = self.policy_net(states).gather(1, actions) next_q_vals = self.target_net(next_states).max( 1, keepdim=True)[0].detach() target = (rewards + self.gamma * next_q_vals * (1 - dones)).to( self.device) loss = F.smooth_l1_loss(curr_q_vals, target) self.optim.zero_grad() loss.backward() self.optim.step() self.updated += 1 if self.updated % self.target_update_freq == 0: self.target_hard_update() return loss.item() def target_hard_update(self): """ Clone the policy net weights into the target net """ self.target_net.load_state_dict(self.policy_net.state_dict())
class DQNAgent(BaseAgent): """ Agent with a DQN network. """ def __init__(self, input_dim, output_dim, lr, gamma, max_memory_size, batch_size, eps_start, eps_end, eps_decay, device, linear1_units=64, linear2_units=64, decay_type="linear"): super().__init__(max_memory_size, batch_size, eps_start, eps_end, eps_decay, device, decay_type) self.model_name = "DQN" self.output_dim = output_dim self.policy_net = DQN(input_dim, output_dim, linear1_units, linear2_units).to(device) # optimizer self.optim = optim.Adam(self.policy_net.parameters(), lr=lr) self.gamma = gamma def choose_action(self, state, testing=False): """ Choose an action to perform. Uses eps-greedy approach. :param state: current state of the environment :param testing: if True, always choose greedy action :return: the action chosen """ self.curr_step += 1 if not testing and np.random.random() < self.curr_eps: return np.random.randint(0, self.output_dim) else: # we're using the network for inference only, we don't want to track the gradients in this case with torch.no_grad(): return self.policy_net(state).argmax().item() def learn(self): """ Update the weights of the network. :return: the loss """ states, next_states, actions, rewards, dones = self.memory.sample( self.batch_size) curr_q_vals = self.policy_net(states).gather(1, actions) next_q_vals = self.policy_net(next_states).max( 1, keepdim=True)[0].detach() target = (rewards + self.gamma * next_q_vals * (1 - dones)).to( self.device) loss = F.smooth_l1_loss(curr_q_vals, target) self.optim.zero_grad() loss.backward() self.optim.step() return loss.item() def set_test(self): """ Sets the network in evaluation mode """ self.policy_net.eval() def set_train(self): """ Sets the network in training mode """ self.policy_net.train() def save(self, filename): """ Save the network weights. :param filename: path """ self.policy_net.save(filename) def load(self, filename): """ Load the network weights. :param filename: path of the weight file """ self.policy_net.load(filename, self.device)