class Agent(): """Interacts with and learns from the environment.""" def __init__(self, state_size, action_size, seed): """Initialize an Agent object. Params ====== state_size (int): dimension of each state action_size (int): dimension of each action seed (int): random seed model (string): which network to use """ self.state_size = state_size self.action_size = action_size self.seed = random.seed(seed) # Q-Network self.qnetwork_local = DuelingQNetwork(state_size, action_size, seed).to(device) self.qnetwork_target = DuelingQNetwork(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) # 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=0.0): """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 # Double DQN # Local network picks action next_action = self.qnetwork_local(next_states).detach().argmax( 1).unsqueeze(1) # Target network estimates the value of said action Q_targets_next = self.qnetwork_target(next_states).gather( 1, next_action) # 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 loss = F.mse_loss(Q_expected, Q_targets) # Minimize the loss self.optimizer.zero_grad() loss.backward() self.optimizer.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)
class Agent(): """Interacts with and learns from the environment.""" def __init__(self, state_size, action_size, num_episodes, seed): """Initialize an Agent object. Params ====== state_size (int): dimension of each state action_size (int): dimension of each action seed (int): random seed num_episodes (int): number of training epochs """ self.state_size = state_size self.action_size = action_size self.seed = seed # Q-Network self.qnetwork_local = DuelingQNetwork(state_size, action_size, seed).to(device) self.qnetwork_target = DuelingQNetwork(state_size, action_size, seed).to(device) self.optimizer = optim.Adam(self.qnetwork_local.parameters(), lr=LR) # Replay memory self.anneal_beta = (1. - BETA) / num_episodes self.memory = ReplayBuffer(action_size, BUFFER_SIZE, BATCH_SIZE, seed, ALPHA, BETA) # Initialize time step (for updating every UPDATE_EVERY steps) self.t_step = 0 self.t_learning_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 update_weights(self): self.memory.anneal_beta(self.anneal_beta) 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.Tensor]): tuple of (s, a, r, s', done) tuples gamma (float): discount factor """ states, actions, rewards, next_states, dones, idxs, weights = 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) # update priorities updates = torch.abs(Q_expected - Q_targets).cpu().data.squeeze(1).numpy() self.memory.update_priorities(idxs, updates) # Compute loss loss = F.l1_loss(Q_expected, Q_targets) # Minimize the loss self.optimizer.zero_grad() (loss * weights).mean().backward() self.optimizer.step() # ------------------- update target network ------------------- # self.t_learning_step += 1 if self.t_learning_step % UPDATE_TARGET_STEPS == 0: 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()): # PyTorch copy: destination.data.copy(source.data) target_param.data.copy_(tau*local_param.data + (1.0-tau)*target_param.data)
class DDDQNPolicy(Policy): """Dueling Double DQN policy""" def __init__(self, state_size, action_size, parameters, evaluation_mode=False): self.evaluation_mode = evaluation_mode self.state_size = state_size self.action_size = action_size self.double_dqn = True self.hidsize = 1 if not evaluation_mode: self.hidsize = parameters.hidden_size self.buffer_size = parameters.buffer_size self.batch_size = parameters.batch_size self.update_every = parameters.update_every self.learning_rate = parameters.learning_rate self.tau = parameters.tau self.gamma = parameters.gamma self.buffer_min_size = parameters.buffer_min_size # Device if parameters.use_gpu and torch.cuda.is_available(): self.device = torch.device("cuda:0") print(" Using GPU") print(" GPU") else: self.device = torch.device("cpu") print(" Using CPU") # Q-Network self.qnetwork_local = DuelingQNetwork(state_size, action_size, hidsize1=self.hidsize, hidsize2=self.hidsize).to( self.device) if not evaluation_mode: self.qnetwork_target = copy.deepcopy(self.qnetwork_local) self.optimizer = optim.Adam(self.qnetwork_local.parameters(), lr=self.learning_rate) self.memory = ReplayBuffer(action_size, self.buffer_size, self.batch_size, self.device) self.t_step = 0 self.loss = 0.0 def act(self, state, eps=0.): state = torch.from_numpy(state).float().unsqueeze(0).to(self.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 step(self, state, action, reward, next_state, done): assert not self.evaluation_mode, "Policy has been initialized for evaluation only." # 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) % self.update_every if self.t_step == 0: # If enough samples are available in memory, get random subset and learn if len(self.memory) > self.buffer_min_size and len( self.memory) > self.batch_size: self._learn() def _learn(self): experiences = self.memory.sample() states, actions, rewards, next_states, dones = experiences # Get expected Q values from local model q_expected = self.qnetwork_local(states).gather(1, actions) if self.double_dqn: # Double DQN q_best_action = self.qnetwork_local(next_states).max(1)[1] q_targets_next = self.qnetwork_target(next_states).gather( 1, q_best_action.unsqueeze(-1)) else: # DQN q_targets_next = self.qnetwork_target(next_states).detach().max( 1)[0].unsqueeze(-1) # Compute Q targets for current states q_targets = rewards + (self.gamma * q_targets_next * (1 - dones)) # Compute loss self.loss = F.mse_loss(q_expected, q_targets) # Minimize the loss self.optimizer.zero_grad() self.loss.backward() self.optimizer.step() # Update target network self._soft_update(self.qnetwork_local, self.qnetwork_target, self.tau) def _soft_update(self, local_model, target_model, tau): # Soft update model parameters. # θ_target = τ*θ_local + (1 - τ)*θ_target 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 save(self, filename): torch.save(self.qnetwork_local.state_dict(), filename + ".local") torch.save(self.qnetwork_target.state_dict(), filename + ".target") def load(self, filename): if os.path.exists(filename + ".local"): self.qnetwork_local.load_state_dict( torch.load(filename + ".local", map_location=torch.device('cpu'))) print('local') if os.path.exists(filename + ".target"): self.qnetwork_target.load_state_dict( torch.load(filename + ".target", map_location=torch.device('cpu'))) print('target')
class Agent(): """Interacts with and learns from the environment.""" def __init__(self, state_size, action_size, seed): #, writer): """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) # Q-Network self.qnetwork_local = DuelingQNetwork(state_size, action_size, seed).to(device) self.qnetwork_target = DuelingQNetwork(state_size, action_size, seed).to(device) self.optimizer = optim.Adam(self.qnetwork_local.parameters(), lr=LR) # TODO: Swap ReplayBuffer for PER buffer # Replay memory # self.memory = ReplayBuffer(action_size, BUFFER_SIZE, BATCH_SIZE, seed) self.memory = PrioritisedReplayBuffer(action_size, BUFFER_SIZE, BATCH_SIZE, ALPHA, EPSILON) # Initialize time step (for updating every UPDATE_EVERY steps) self.t_step = 0 self.beta = BETA_START # self.writer = writer def step(self, state, action, reward, next_state, done): # calculate error, and store experience in replay buffer accordingly # next_actions = self.qnetwork_local(next_states).max(1).indices.unsqueeze(1) # Q_targets_next = self.qnetwork_target(next_states).detach().max(1).values.unsqueeze(1) # << [64,1] of max Q values # Q_targets_next = self.qnetwork_target(next_states).gather(1, next_actions) # Q_targets = rewards + (gamma * Q_targets_next * (1 - dones)) # Q_expected = self.qnetwork_local(states).gather(1, actions) # gather uses the actions as indices to select the Qs # get next action from qnetwork_local, using next_state # get next reward using next action, from qnetwork_target # calc. target: reward + (gamma * next reward * done mask) # get expected from qnetwork_local, for current state/action s = torch.tensor([state]).float().to(device) ns = torch.tensor([next_state]).float().to(device) next_action = self.qnetwork_local(ns).max(1).indices.unsqueeze(1) next_reward = self.qnetwork_target(ns).detach()[0, next_action] target = reward + (GAMMA * next_reward * (1 - done)) expected = self.qnetwork_local(s).detach()[0, action] error = torch.abs(target - expected).cpu().detach() self.memory.add(error, state, action, reward, next_state, done) # Learn every UPDATE_EVERY time steps. self.t_step += 1 # self.writer.add_scalar('Timestep Error', error, self.t_step) if (self.t_step % UPDATE_EVERY) == 0: # If enough samples are available in memory, get random subset and learn if len(self.memory) > BATCH_SIZE: self.beta += ( (1 - self.beta) / BETA_STEPS ) # anneal the beta, from a starting value towards 1.0 self.beta = np.min([1., self.beta]) experiences = self.memory.sample(self.beta) 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) 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 action_values.max(1).indices.unsqueeze( 1).cpu().detach().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 """ # states, actions, rewards, next_states, dones = experiences states, actions, rewards, next_states, dones, weights, idxs = experiences next_actions = self.qnetwork_local(next_states).max( 1).indices.unsqueeze(1) # Q_targets_next = self.qnetwork_target(next_states).detach().max(1).values.unsqueeze(1) # << [64,1] of max Q values Q_targets_next = self.qnetwork_target(next_states).detach().gather( 1, next_actions) Q_targets = rewards + (gamma * Q_targets_next * (1 - dones)) Q_expected = self.qnetwork_local(states).gather( 1, actions) # gather uses the actions as indices to select the Qs # refresh errors in replay buffer errors = torch.abs(Q_expected - Q_targets).cpu().detach() for (idx, error) in zip(idxs, errors): self.memory.update(idx, error) loss = (weights.detach() * F.mse_loss(Q_expected, Q_targets)).mean() # weighted loss self.optimizer.zero_grad() loss.backward() self.optimizer.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)
class Agent(): """ Interacts with and learns from the environment. This agent implements a few improvements over the vanilla DQN, making it a Double Dueling Deep Q-Learning Network with Prioritized Experience Replay. * Deep Q-Learning Network: RL where a deep learning network is used for the Q-network estimate. * Double DQN: The local network from DQN is used to select the optimal action during learning, but the policy estimate for that action is computed using the target network. * Dueling DQN: The deep learning network explicitly estimates the value function and the advantage functions separately. * DQN-PER: Experiences are associated with a probability weight based upon the absolute error between the estimated Q-value and the target Q-value at time of estimation -- prioritizing experiences that help learn more. """ def __init__(self, state_size, action_size, buffer_size=int(1e5), batch_size=64, gamma=0.99, tau=1e-3, learn_rate=5e-4, update_every=4, per_epsilon=1e-5, per_alpha=0.6, per_beta=0.9, device=DEFAULT_DEVICE, seed=0): """ Initialize an object. :param state_size: (int) Dimension of each state :param action_size: (int) Dimension of each action :param buffer_size: (int) Replay buffer size :param batch_size: (int) Minibatch size used during learning :param gamma: (float) Discount factor :param tau: (float) Scaling parameter for soft update :param learn_rate: (float) Learning rate used by optimizer :param update_every: (int) Steps between updates of target network :param per_epsilon: (float) PER hyperparameter, constant added to each error :param per_alpha: (float) PER hyperparameter, exponent applied to each probability :param per_beta: (float) PER hyperparameter, bias correction exponent for probability weight :param device: (torch.device) Object representing the device where to allocate tensors :param seed: (int) Seed used for PRNG """ # Save copy of model parameters self.state_size = state_size self.action_size = action_size self.seed = random.seed(seed) self.device = device # Save copy of hyperparameters self.buffer_size = buffer_size self.batch_size = batch_size self.gamma = gamma self.tau = tau self.learn_rate = learn_rate self.update_every = update_every self.per_epsilon = per_epsilon self.per_alpha = per_alpha self.per_beta = per_beta # Q networks self.qnetwork_local = DuelingQNetwork(state_size, action_size, seed).to(device) self.qnetwork_target = DuelingQNetwork(state_size, action_size, seed).to(device) self.optimizer = optim.Adam(self.qnetwork_local.parameters(), lr=learn_rate) # Replay memory self.memory = PrioritizedReplayBuffer(memory_size=buffer_size, device=device, update_every=update_every, seed=seed) # Initialize time step (for updating every self.update_every steps) self.t_step = 0 self.episode = 0 def step(self, state, action, reward, next_state, done): """ Store a single agent step, learning every N steps :param state: (array-like) Initial state on the visit :param action: (int) Action on the visit :param reward: (float) Reward received on the visit :param next_state: (array-like) State reached after the visit :param done: (bool) Flag whether the next state is a terminal state """ # Save experience in replay memory self.memory.add(state, action, reward, next_state, done) # Learn every self.update_every time steps self.t_step = (self.t_step + 1) % self.update_every if self.t_step == 0: # If enough samples are available in memory, get random subset and learn if len(self.memory) > self.batch_size: experiences = self.memory.sample(batch_size=self.batch_size, alpha=self.per_alpha, beta=self.per_beta) self.learn(experiences) # Keep track of episode number if done: self.episode += 1 def act(self, state, eps=0.): """ Returns the selected action for the given state according to the current policy :param state: (array_like) Current state :param eps: (float) Epsilon, for epsilon-greedy action selection :return: action (int) """ state = torch.from_numpy(state).float().unsqueeze(0).to(self.device) self.qnetwork_local.eval() with torch.no_grad(): action_values = self.qnetwork_local(state) self.qnetwork_local.train() # Epsilon-greedy action selection # Convert types to np.int32 for compatibility with environment if random.random() > eps: return np.argmax(action_values.cpu().data.numpy()).astype(np.int32) else: return random.choice(np.arange(self.action_size)).astype(np.int32) def learn(self, experiences): """ Update value parameters using given batch of indexed experience tuples :param experiences: (Tuple[torch.Tensor, np.array]) (s, a, r, s', done, index) tuples """ states, actions, rewards, next_states, dones, indexes = experiences # Get max predicted Q values (for next states) from target model # Double DQN: use local network to select action with maximum value, # then use target network to get Q value for that action Q_next_indices = self.qnetwork_local(next_states).detach().argmax( 1).unsqueeze(1) Q_next_values = self.qnetwork_target(next_states).detach() Q_targets_next = Q_next_values.gather(1, Q_next_indices) # Compute Q target for current states Q_targets = rewards + self.gamma * Q_targets_next * (1 - dones) # Get expected Q values from local model Q_expected = self.qnetwork_local(states).gather(1, actions) # Compute estimation error (for Prioritized Experience Replay) and update weights Q_error = (torch.abs(Q_expected.detach() - Q_targets.detach()) + self.per_epsilon).squeeze() self.memory.update(indexes, Q_error) # Compute loss loss = F.mse_loss(Q_expected, Q_targets) # Minimize the loss self.optimizer.zero_grad() loss.backward() self.optimizer.step() # Update target network for target_param, local_param in zip(self.qnetwork_target.parameters(), self.qnetwork_local.parameters()): target_param.data.copy_(self.tau * local_param.data + (1.0 - self.tau) * target_param.data)
class Agent(): """Interacts with and learns from the environment.""" def __init__(self, state_size, action_size, seed, hidden_sizes=[64, 64], flavor='plain'): """Initialize an Agent object. Params ====== state_size (int): dimension of each state action_size (int): dimension of each action seed (int): random seed hidden_sizes (list): list of neurons in each layer flavor (str): flavor of the network - plain, double, dueling, double-dueling """ self.state_size = state_size self.action_size = action_size self.seed = random.seed(seed) self.hidden_sizes = hidden_sizes self.flavor = flavor # Q-Network if self.flavor == 'plain' or self.flavor == 'double': self.qnetwork_local = QNetwork(state_size, action_size, seed, hidden_sizes).to(device) self.qnetwork_target = QNetwork(state_size, action_size, seed, hidden_sizes).to(device) # Dueling Q-Network if self.flavor == 'dueling' or self.flavor == 'double-dueling': self.qnetwork_local = DuelingQNetwork(state_size, action_size, seed, hidden_sizes).to(device) self.qnetwork_target = DuelingQNetwork(state_size, action_size, seed, hidden_sizes).to(device) self.optimizer = optim.Adam(self.qnetwork_local.parameters(), lr=LR) # Replay memory self.memory = ReplayBuffer(action_size, BUFFER_SIZE, BATCH_SIZE, seed) # Initialize time step (for updating every UPDATE_EVERY steps) self.t_step = 0 def show_network(self): x = Variable(torch.randn(1, self.state_size)) y = self.qnetwork_local(x) return make_dot(y, params=dict( list(self.qnetwork_local.named_parameters()))) 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. 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 # Best actions ... if self.flavor == 'plain' or self.flavor == 'dueling': # ... according to target model for Double DQN best_actions = self.qnetwork_target(next_states).detach().argmax( dim=1).unsqueeze(1) if self.flavor == 'double' or self.flavor == 'double-dueling': # ... according to local model for Double DQN best_actions = self.qnetwork_local(next_states).detach().argmax( dim=1).unsqueeze(1) # Maximal predicted Q value for next state from target model Q_t_max = self.qnetwork_target(next_states).gather( 1, best_actions).detach() # Q targets for current state Q_t = rewards + (1 - dones) * gamma * Q_t_max # Expected Q values of local model Q_exp = self.qnetwork_local(states).gather(1, actions) # Loss function loss = F.mse_loss(Q_exp, Q_t) self.optimizer.zero_grad() loss.backward() self.optimizer.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)
class Agent(): """Interacts with and learns from the environment.""" def __init__(self, state_size, action_size, seed): """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) # Q-Network self.qnetwork_local = DuelingQNetwork(state_size, action_size, seed).to(device) self.qnetwork_target = DuelingQNetwork(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, ALPHA) # Initialize time step (for updating every UPDATE_EVERY steps) self.t_step = 0 # Initialize learning step for updating beta self.learn_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 prioritized subset and learn if len(self.memory) > BATCH_SIZE: experiences = self.memory.sample() self.learn(experiences, GAMMA, BETA) 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) # Choose action values according to local model 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, beta): """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 beta (float): reliance of importance sampling weight on priortization """ # Beta will reach 1 after 25,000 training steps (~325 episodes) b = min(1.0, beta + self.learn_step * (1.0 - beta) / 25000) self.learn_step += 1 states, actions, rewards, next_states, dones, probabilities, indices = experiences # # Get max predicted actions (for next states) from local model # next_local_actions = self.qnetwork_local(next_states).max(1)[1].unsqueeze(1) # # Evaluate the max predicted actions from the local model on the target model # # based on Double DQN # Q_targets_next_values = self.qnetwork_target(next_states).detach().gather(1, next_local_actions) # # Compute Q targets for current states # Q_targets = rewards + (gamma * Q_targets_next_values * (1 - dones)) # # Get expected Q values from local # Q_expected = self.qnetwork_local(states).gather(1, actions) ## Double DQN Q_expected = self.qnetwork_local(states).gather(1, actions) next_actions = self.qnetwork_local(next_states).argmax(-1, keepdim=True) Q_targets_next = self.qnetwork_target(next_states).gather(-1, next_actions) Q_targets = rewards + GAMMA * Q_targets_next * (1-dones) # Compute and update new priorities new_priorities = (abs(Q_expected - Q_targets) + 0.2).detach() self.memory.update_priority(new_priorities, indices) # Compute and apply importance sampling weights to TD Errors ISweights = (((1 / len(self.memory)) * (1 / probabilities)) ** b) max_ISweight = torch.max(ISweights) ISweights /= max_ISweight Q_targets *= ISweights Q_expected *= ISweights # Compute loss loss = F.mse_loss(Q_expected, Q_targets) self.last_loss = loss # Minimize the loss self.optimizer.zero_grad() loss.backward() self.optimizer.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)
class Agent(): def __init__(self, state_size, action_size, seed): self.state_size = state_size self.action_size = action_size self.seed = random.seed(seed) # Q-Network self.model = DuelingQNetwork(state_size, action_size, seed).to(device) # self.qnetwork_target = DuelingQNetwork(state_size, action_size, seed).to(device) # for target_param, param in zip(self.qnetwork_local.parameters(),self.qnetwork_target.parameters()): # target_param.data.copy_(param) self.optimizer = optim.Adam(self.model.parameters(), lr=LR) # Replay memory self.memory = ReplayBuffer(action_size, BUFFER_SIZE, BATCH_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() self.learn(experiences, GAMMA) def act(self, state, eps=0.): state = torch.from_numpy(state).float().unsqueeze(0).to(device) self.model.eval() with torch.no_grad(): action_values = self.model(state) self.model.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): states, actions, rewards, next_states, dones = experiences actions = actions.view(actions.size(0), 1) dones = dones.view(dones.size(0), 1) # curr_Q = self.qnetwork_local.forward(states).gather(1, actions) # next_Q = self.qnetwork_target.forward(next_states) # max_next_Q = torch.max(next_Q, 1)[0] # max_next_Q = max_next_Q.view(max_next_Q.size(0), 1) # expected_Q = rewards + (1 - dones) * gamma * max_next_Q # loss = F.mse_loss(curr_Q, expected_Q.detach()) curr_Q = self.model.forward(states).gather(1, actions.unsqueeze(1)) curr_Q = curr_Q.squeeze(1) next_Q = self.model.forward(next_states) max_next_Q = torch.max(next_Q, 1)[0] expected_Q = rewards.squeeze(1) + self.gamma * max_next_Q loss = self.MSE_loss(curr_Q, expected_Q) self.optimizer.zero_grad() loss.backward() self.optimizer.step() # self.soft_update(self.qnetwork_local, self.qnetwork_target, TAU) # for target_param, param in zip(self.qnetwork_target.parameters(), self.qnetwork_local.parameters()): # target_param.data.copy_(TAU * param + (1 - TAU) * target_param) def soft_update(self, local_model, target_model, tau): 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)