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 = 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) # 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, training_mode=True): """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) if training_mode is True: self.qnetwork_local.train() # Epsilon-greedy action selection if random.random() > eps: action = np.argmax(action_values.cpu().data.numpy()) else: action = random.choice(np.arange(self.action_size)) action = np.int32(action) return action 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 # 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 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 DQN(object): def __init__(self,state_space,action_space,seed,update_every,batch_size,buffer_size,learning_rate): self.action_space = action_space self.state_space = state_space self.seed = random.seed(seed) self.batch_size = batch_size self.buffer_size = buffer_size self.learning_rate = learning_rate self.update_every = update_every self.qnetwork_local = QNetwork(state_space,action_space) self.qnetwork_target = QNetwork(state_space,action_space) self.optimizer = optim.Adam(self.qnetwork_local.parameters(),lr=learning_rate) # Initialize replaybuffer self.memory = ReplayBuffer(action_space,buffer_size,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,GAMMA): # Save the experience self.memory.add_experience(state,action,reward,next_state,done) # learn from the experience self.t_step = (self.t_step + 1) % self.update_every if self.t_step == 0: if len(self.memory) > self.buffer_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.qnetwork_local.eval() with torch.no_grad(): action_values = self.qnetwork_local(state) self.qnetwork_local.train() if random.random() > eps: return np.argmax(action_values.cpu().data.numpy()) else: return random.sample(np.arange(self.action_space)) def learn(self,experiences,GAMMA): states,actions,rewards,next_states,dones = experiences target_values = self.qnetwork_target(next_states).detach().max(1)[0].unsqueeze(1) targets = reward + (GAMMA * target_values * (1-done)) action_values = self.qnetwork_local(states).gather(1,actions) loss = F.mse_loss(action_values,targets) loss.backward() self.optimizer.step() soft_update(TAU) def soft_update(self,tau): """Soft update model parameters. θ_target = τ*θ_local + (1 - τ)*θ_target """ for local_param,target_param in zip(self.qnetwork_local.parameters(),self.qnetwork_target.parameters()): local_param.data.copy_(tau*local_param.data + (1-tau)*target_param.data) # self.qnetwork_local.parameters() = TAU*self.qnetwork_local.parameters() + (1-TAU)*self.qnetwork_target.parameters()
class DQN_Agent(): """ Interacts an learns from the environment. """ def __init__(self, state_size, action_size, seed, GAMMA=GAMMA, TAU=TAU, LR=LR, UPDATE_EVERY=UPDATE_EVERY, BUFFER_SIZE=BUFFER_SIZE, BATCH_SIZE=BATCH_SIZE): """ Initialize the agent. ========== PARAMETERS ========== state_size (int) = observation dimension of the environment 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.gamma = GAMMA self.tau = TAU self.lr = LR self.update_every = UPDATE_EVERY self.buffer_size = BUFFER_SIZE self.batch_size = BATCH_SIZE self.device = torch.device( "cuda:0" if torch.cuda.is_available() else "cpu") # instantiate online local and target network for weight updates self.qnetwork_local = QNetwork(state_size, action_size, seed).to(self.device) self.qnetwork_target = QNetwork(state_size, action_size, seed).to(self.device) self.optimizer = optim.Adam(self.qnetwork_local.parameters(), lr=self.lr) # create a replay buffer self.memory = ReplayBuffer(action_size, self.buffer_size, self.batch_size, seed, self.device) # time steps for updating target network every time t_step % 4 == 0 self.t_step = 0 def step(self, state, action, reward, next_state, done): ''' Append a SARS sequence to memory, then every update_every steps learn from experiences''' self.memory.add(state, action, reward, next_state, done) self.t_step = (self.t_step + 1) % self.update_every if self.t_step == 0: # in case enough samples are available in internal memory, sample and learn if len(self.memory) > self.batch_size: experiences = self.memory.sample() self.learn(experiences, self.gamma) def act(self, state, eps=0.): """ Choose action from an epsilon-greedy policy ========== PARAMETERS ========== state (array) = current state space eps (float) = epsilon, for epsilon-greedy action choice """ state = torch.from_numpy(state).float().unsqueeze(0).to(self.device) self.qnetwork_local.eval() with torch.no_grad(): action_values = self.qnetwork_local.forward(state) self.qnetwork_local.train() 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 the value parameters using experience tuples sampled from ReplayBuffer ========== PARAMETERS ========== experiences = Tuple of torch.Variable: SARS', done gamma (float) = discount factor to weight rewards """ states, actions, rewards, next_states, dones = experiences # calculate max predicted Q values for the next states using target model Q_targets_next = self.qnetwork_target(next_states).detach().max( 1)[0].unsqueeze(1) Q_targets = rewards + (gamma * Q_targets_next * (1 - dones)) # calculate expected Q vaues from the local model Q_expected = self.qnetwork_local(states).gather(1, actions) # compute MSE Loss loss = F.mse_loss(Q_expected, Q_targets) self.optimizer.zero_grad() loss.backward() self.optimizer.step() self.soft_update(self.qnetwork_local, self.qnetwork_target, self.tau) def soft_update(self, local_model, target_model, tau): """ Soft update for model parameters, every update steps as defined above theta_target = tau * theta_local + (1-tau)*theta_target ========== PARAMETERS ========== local_model, target_model = PyTorch Models, weights will be copied from-to tau = interpolation parameter, type=float """ for target_param, local_param in zip(target_model.parameters(), local_model.parameters()): target_param.data.copy_(self.tau * local_param.data + (1.0 - self.tau) * target_param.data)