def __init__(self, n_agents, state_size, action_size, random_seed): """Initialize an Agent object. Params ====== n_agents (int): number of agents state_size (int): dimension of each state action_size (int): dimension of each action random_seed (int): random seed """ self.n_agents = n_agents self.state_size = state_size self.action_size = action_size self.seed = np.random.seed(random_seed) random.seed(random_seed) # Actor Network (w/ Target Network) # self.actor_local = Actor(state_size, action_size, seed=random_seed, leak=LEAKINESS).to(device) # self.actor_target = Actor(state_size, action_size, seed=random_seed, leak=LEAKINESS).to(device) self.actor_local = Actor(state_size, action_size, seed=random_seed).to(device) self.actor_target = Actor(state_size, action_size, seed=random_seed).to(device) self.actor_optimizer = optim.Adam(self.actor_local.parameters(), lr=LR_ACTOR) # Critic Network (w/ Target Network) # self.critic_local = Critic(state_size, action_size, seed=random_seed, leak=LEAKINESS).to(device) # self.critic_target = Critic(state_size, action_size, seed=random_seed, leak=LEAKINESS).to(device) self.critic_local = Critic(state_size, action_size, seed=random_seed).to(device) self.critic_target = Critic(state_size, action_size, seed=random_seed).to(device) self.critic_optimizer = optim.Adam(self.critic_local.parameters(), lr=LR_CRITIC, weight_decay=WEIGHT_DECAY) # self.critic_optimizer = optim.Adam(self.critic_local.parameters(), lr=LR_CRITIC) # initialize targets same as original networks self.hard_update(self.actor_target, self.actor_local) self.hard_update(self.critic_target, self.critic_local) # Noise process # self.noise = OUNoise((n_agents, action_size), random_seed) self.noise = OUNoise(action_size, random_seed) # Replay memory self.memory = ReplayBuffer(action_size, BUFFER_SIZE, BATCH_SIZE, random_seed) # Initialize time step (for updating every UPDATE_EVERY steps) self.t_step = 0
class MADDPGAgent(): """Interacts with and learns from the environment.""" def __init__(self, state_size, action_size, n_agents, current_agent, random_seed=0): """Initialize an Agent object. Params ====== state_size (int): dimension of each state action_size (int): dimension of each action n_agents (int): number of agents current_agent (int): index of current agent random_seed (int): random seed """ self.n_agents = n_agents self.current_agent = current_agent self.state_size = state_size self.action_size = action_size self.seed = np.random.seed(random_seed) random.seed(random_seed) # Actor Network (w/ Target Network) # self.actor_local = Actor(state_size, action_size, seed=random_seed, leak=LEAKINESS, use_bn=False).to(device) # self.actor_target = Actor(state_size, action_size, seed=random_seed, leak=LEAKINESS, use_bn=False).to(device) self.actor_local = Actor(state_size, action_size, seed=random_seed, use_bn=False).to(device) self.actor_target = Actor(state_size, action_size, seed=random_seed, use_bn=False).to(device) self.actor_optimizer = optim.Adam(self.actor_local.parameters(), lr=LR_ACTOR) # Critic Network (w/ Target Network) # self.critic_local = Critic(state_size * n_agents, action_size * n_agents, # seed=random_seed, leak=LEAKINESS, use_bn=False).to(device) # self.critic_target = Critic(self.state_size * n_agents, self.action_size * n_agents, # seed=random_seed, leak=LEAKINESS, use_bn=False).to(device) self.critic_local = Critic(state_size * n_agents, action_size * n_agents, seed=random_seed, use_bn=False).to(device) self.critic_target = Critic(self.state_size * n_agents, self.action_size * n_agents, seed=random_seed, use_bn=False).to(device) self.critic_optimizer = optim.Adam(self.critic_local.parameters(), lr=LR_CRITIC, weight_decay=WEIGHT_DECAY) # self.critic_optimizer = optim.Adam(self.critic_local.parameters(), lr=LR_CRITIC) # Noise process self.noise = OUNoise(action_size, random_seed) # self.noise = OUNoise((n_agents, action_size), random_seed) # Replay memory self.memory = ReplayBuffer(action_size, BUFFER_SIZE, BATCH_SIZE, random_seed) # Initialize time step (for updating every UPDATE_EVERY steps) self.t_step = 0 def act(self, state, add_noise=True): """Returns actions for given state as per current policy.""" state = torch.from_numpy(state).float().to(device) # Run inference in eval mode self.actor_local.eval() with torch.no_grad(): action = self.actor_local(state).cpu().data.numpy() self.actor_local.train() # add noise if true if add_noise: action += self.noise.sample() return np.clip(action, -1, 1) def reset(self): self.noise.reset() def learn(self, agents, experiences, gamma): """Update policy and value parameters using given batch of experience tuples. Q_targets = r + γ * critic_target(next_state, actor_target(next_state)) where: actor_target(state) -> action critic_target(state, action) -> Q-value Params ====== agents (Agent class): all agents object experiences (Tuple[torch.Tensor]): tuple of (s, a, r, s', done) tuples gamma (float): discount factor """ states, actions, rewards, next_states, dones = experiences # ---------------------------- update critic ---------------------------- # # Get predicted next-state actions and Q values from target models actions_next = torch.zeros( (len(states), self.n_agents, self.action_size)).to(device) for i, agent in enumerate(agents): actions_next[:, i] = agent.actor_target(states[:, i, :]) # print('\nactions_next:', actions_next.size(), '\nnext_states:', next_states.size()) # # Flatten state and action # # e.g from state (100,2,24) --> (100, 48) critic_states = flatten(next_states) actions_next = flatten(actions_next) # print('after flatten:\nactions_next:', actions_next.size(), '\nnext_states:', critic_states.size()) # calculate target and expected Q_targets_next = self.critic_target(critic_states, actions_next) # Q_targets_next = self.critic_target(next_states, actions_next) # print('Q_targets_next:', Q_targets_next.size()) Q_targets = rewards[:, self.current_agent, :] + ( gamma * Q_targets_next * (1 - dones[:, self.current_agent, :])) # print('Q_targets', Q_targets.size()) # print('\nactions:', actions.size(), '\nstates:', states.size()) # print('after flatten:\nactions:', flatten(actions).size(), '\nstates:', flatten(states).size()) Q_expected = self.critic_local(flatten(states), flatten(actions)) # Q_expected = self.critic_local(states, actions) # # Compute critic loss, use mse loss critic_loss = F.mse_loss(Q_expected, Q_targets) # critic_loss_value = critic_loss.item() self.critic_optimizer.zero_grad() critic_loss.backward() torch.nn.utils.clip_grad_norm_( self.critic_local.parameters(), 1) # Added because I got converge problems # then I found in adaptationio's solution (https://github.com/adaptationio/DDPG-Continuous-Control), he added this. # Here to see the purpose of doing so: https://discuss.pytorch.org/t/about-torch-nn-utils-clip-grad-norm/13873 # for param in self.critic_local.parameters(): # param.grad.data.clamp_(-1, 1) self.critic_optimizer.step() # ---------------------------- update actor ---------------------------- # # Update the predicted action of current agent actions_pred = torch.zeros( (len(states), self.n_agents, self.action_size)).to(device) actions_pred.data.copy_(actions.data) actions_pred[:, self.current_agent] = self.actor_local( states[:, self.current_agent]) actor_loss = -self.critic_local(flatten(states), flatten(actions_pred)).mean() # actor_loss = -self.critic_local(states, actions_pred).mean() # Minimize the loss self.actor_optimizer.zero_grad() actor_loss.backward() torch.nn.utils.clip_grad_norm_( self.actor_local.parameters(), 1) # Added because I got converge problems # for param in self.critic_local.parameters(): # param.grad.data.clamp_(-1, 1) self.actor_optimizer.step() # ----------------------- update target networks ----------------------- # if self.t_step == 0: # One time only, start local and target with same parameters self._copy_weights(self.critic_local, self.critic_target) self._copy_weights(self.actor_local, self.actor_target) else: self.soft_update(self.critic_local, self.critic_target, TAU) self.soft_update(self.actor_local, self.actor_target, TAU) self.t_step += 1 def _copy_weights(self, source_network, target_network): """Copy source network weights to target""" for target_param, source_param in zip(target_network.parameters(), source_network.parameters()): target_param.data.copy_(source_param.data) 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)
def __init__(self, n_agents, state_size, action_size, random_seed, prioritized_reply=False): """Initialize an Agent object. Params ====== n_agents (int): number of agents state_size (int): dimension of each state action_size (int): dimension of each action random_seed (int): random seed prioritized_reply (bool): True or False for using prioritized reply buffer, default is False """ self.n_agents = n_agents self.state_size = state_size self.action_size = action_size self.seed = np.random.seed(random_seed) random.seed(random_seed) self.prioritized_reply = prioritized_reply # Actor Network (w/ Target Network) self.actor_local = Actor(state_size, action_size, seed=random_seed, leak=LEAKINESS).to(device) self.actor_target = Actor(state_size, action_size, seed=random_seed, leak=LEAKINESS).to(device) self.actor_optimizer = optim.Adam(self.actor_local.parameters(), lr=LR_ACTOR) # Critic Network (w/ Target Network) self.critic_local = Critic(state_size, action_size, seed=random_seed, leak=LEAKINESS).to(device) self.critic_target = Critic(state_size, action_size, seed=random_seed, leak=LEAKINESS).to(device) self.critic_optimizer = optim.Adam(self.critic_local.parameters(), lr=LR_CRITIC) # Noise process self.noise = OUNoise((20, action_size), random_seed) # Replay memory if self.prioritized_reply: self.memory = PreReplayBuffer(action_size, BUFFER_SIZE, BATCH_SIZE, random_seed, ALPHA) # Initialize learning step for updating beta self.learn_step = 0 else: self.memory = ReplayBuffer(action_size, BUFFER_SIZE, BATCH_SIZE, random_seed) # Initialize time step (for updating every UPDATE_EVERY steps) self.t_step = 0
class Agent(): """Interacts with and learns from the environment.""" def __init__(self, n_agents, state_size, action_size, random_seed): """Initialize an Agent object. Params ====== n_agents (int): number of agents state_size (int): dimension of each state action_size (int): dimension of each action random_seed (int): random seed """ self.n_agents = n_agents self.state_size = state_size self.action_size = action_size self.seed = np.random.seed(random_seed) random.seed(random_seed) # Actor Network (w/ Target Network) # self.actor_local = Actor(state_size, action_size, seed=random_seed, leak=LEAKINESS).to(device) # self.actor_target = Actor(state_size, action_size, seed=random_seed, leak=LEAKINESS).to(device) self.actor_local = Actor(state_size, action_size, seed=random_seed).to(device) self.actor_target = Actor(state_size, action_size, seed=random_seed).to(device) self.actor_optimizer = optim.Adam(self.actor_local.parameters(), lr=LR_ACTOR) # Critic Network (w/ Target Network) # self.critic_local = Critic(state_size, action_size, seed=random_seed, leak=LEAKINESS).to(device) # self.critic_target = Critic(state_size, action_size, seed=random_seed, leak=LEAKINESS).to(device) self.critic_local = Critic(state_size, action_size, seed=random_seed).to(device) self.critic_target = Critic(state_size, action_size, seed=random_seed).to(device) self.critic_optimizer = optim.Adam(self.critic_local.parameters(), lr=LR_CRITIC, weight_decay=WEIGHT_DECAY) # self.critic_optimizer = optim.Adam(self.critic_local.parameters(), lr=LR_CRITIC) # initialize targets same as original networks self.hard_update(self.actor_target, self.actor_local) self.hard_update(self.critic_target, self.critic_local) # Noise process # self.noise = OUNoise((n_agents, action_size), random_seed) self.noise = OUNoise(action_size, random_seed) # Replay memory self.memory = ReplayBuffer(action_size, BUFFER_SIZE, BATCH_SIZE, random_seed) # Initialize time step (for updating every UPDATE_EVERY steps) self.t_step = 0 def step(self, states, actions, rewards, next_states, dones): """Save experience in replay memory, and use random sample from buffer to learn.""" # Save experience / reward for state, action, reward, next_state, done in zip( states, actions, rewards, next_states, dones): self.memory.add(state, action, reward, next_state, done) # self.memory.add(states, actions, rewards, next_states, dones) # Learn, if enough samples are available in memory self.t_step = self.t_step + 1 if (len(self.memory) > BATCH_SIZE) and (self.t_step % UPDATE_EVERY == 0): # experiences = self.memory.sample() # self.learn(experiences, GAMMA) for _ in range(10): experiences = self.memory.sample() self.learn(experiences, GAMMA) def act(self, state, add_noise=True): """Returns actions for given state as per current policy.""" state = torch.from_numpy(state).float().to(device) self.actor_local.eval() with torch.no_grad(): action = self.actor_local(state).cpu().data.numpy() self.actor_local.train() if add_noise: action += self.noise.sample() return np.clip(action, -1, 1) def reset(self): self.noise.reset() def learn(self, experiences, gamma): """Update policy and value parameters using given batch of experience tuples. Q_targets = r + γ * critic_target(next_state, actor_target(next_state)) where: actor_target(state) -> action critic_target(state, action) -> Q-value Params ====== experiences (Tuple[torch.Tensor]): tuple of (s, a, r, s', done) tuples gamma (float): discount factor """ states, actions, rewards, next_states, dones = experiences # ---------------------------- update critic ---------------------------- # # Get predicted next-state actions and Q values from target models actions_next = self.actor_target(next_states) Q_targets_next = self.critic_target(next_states, actions_next) # Compute Q targets for current states (y_i) Q_targets = rewards + (gamma * Q_targets_next * (1 - dones)) # Compute critic loss Q_expected = self.critic_local(states, actions) # Compute critic loss critic_loss = F.mse_loss(Q_expected, Q_targets) # Minimize the loss self.critic_optimizer.zero_grad() critic_loss.backward() torch.nn.utils.clip_grad_norm_( self.critic_local.parameters(), 1) # Added because I got converge problems # then I found in adaptationio's solution (https://github.com/adaptationio/DDPG-Continuous-Control), he added this. # Here to see the purpose of doing so: https://discuss.pytorch.org/t/about-torch-nn-utils-clip-grad-norm/13873 self.critic_optimizer.step() # ---------------------------- update actor ---------------------------- # # Compute actor loss actions_pred = self.actor_local(states) actor_loss = -self.critic_local(states, actions_pred).mean() # Minimize the loss self.actor_optimizer.zero_grad() actor_loss.backward() torch.nn.utils.clip_grad_norm_( self.actor_local.parameters(), 1) # Added because I got converge problems self.actor_optimizer.step() # ----------------------- update target networks ----------------------- # self.soft_update(self.critic_local, self.critic_target, TAU) self.soft_update(self.actor_local, self.actor_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) def hard_update(self, target, source): """ Copy network parameters from source to target Inputs: target (torch.nn.Module): Net to copy parameters to source (torch.nn.Module): Net whose parameters to copy """ for target_param, param in zip(target.parameters(), source.parameters()): target_param.data.copy_(param.data)
class Agent(): """Interacts with and learns from the environment.""" def __init__(self, n_agents, state_size, action_size, random_seed, prioritized_reply=False): """Initialize an Agent object. Params ====== n_agents (int): number of agents state_size (int): dimension of each state action_size (int): dimension of each action random_seed (int): random seed prioritized_reply (bool): True or False for using prioritized reply buffer, default is False """ self.n_agents = n_agents self.state_size = state_size self.action_size = action_size self.seed = np.random.seed(random_seed) random.seed(random_seed) self.prioritized_reply = prioritized_reply # Actor Network (w/ Target Network) self.actor_local = Actor(state_size, action_size, seed=random_seed, leak=LEAKINESS).to(device) self.actor_target = Actor(state_size, action_size, seed=random_seed, leak=LEAKINESS).to(device) self.actor_optimizer = optim.Adam(self.actor_local.parameters(), lr=LR_ACTOR) # Critic Network (w/ Target Network) self.critic_local = Critic(state_size, action_size, seed=random_seed, leak=LEAKINESS).to(device) self.critic_target = Critic(state_size, action_size, seed=random_seed, leak=LEAKINESS).to(device) self.critic_optimizer = optim.Adam(self.critic_local.parameters(), lr=LR_CRITIC) # Noise process self.noise = OUNoise((20, action_size), random_seed) # Replay memory if self.prioritized_reply: self.memory = PreReplayBuffer(action_size, BUFFER_SIZE, BATCH_SIZE, random_seed, ALPHA) # Initialize learning step for updating beta self.learn_step = 0 else: self.memory = ReplayBuffer(action_size, BUFFER_SIZE, BATCH_SIZE, random_seed) # Initialize time step (for updating every UPDATE_EVERY steps) self.t_step = 0 def step(self, states, actions, rewards, next_states, dones): """Save experience in replay memory, and use random sample from buffer to learn.""" # Save experience / reward self.t_step = self.t_step + 1 for state, action, reward, next_state, done in zip( states, actions, rewards, next_states, dones): self.memory.add(state, action, reward, next_state, done) # Learn every UPDATE_EVERY time steps # If enough samples are available in memory, get random subset and learn if (len(self.memory) > BATCH_SIZE) and (self.t_step % UPDATE_EVERY == 0): if self.prioritized_reply: experiences = self.memory.sample() self.learn(experiences, GAMMA, BETA) else: for _ in range(10): experiences = self.memory.sample() self.learn(experiences, GAMMA) def act(self, state, add_noise=True): """Returns actions for given state as per current policy.""" state = torch.from_numpy(state).float().to(device) self.actor_local.eval() with torch.no_grad(): action = self.actor_local(state).cpu().data.numpy() self.actor_local.train() if add_noise: action += self.noise.sample() return np.clip(action, -1, 1) def reset(self): self.noise.reset() def learn(self, experiences, gamma, beta=None): """Update policy and value parameters using given batch of experience tuples. Q_targets = r + γ * critic_target(next_state, actor_target(next_state)) where: actor_target(state) -> action critic_target(state, action) -> Q-value Params ====== experiences (Tuple[torch.Tensor]): tuple of (s, a, r, s', done) tuples gamma (float): discount factor beta (float): reliance of importance sampling weight on prioritization """ if self.prioritized_reply: # 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 else: states, actions, rewards, next_states, dones = experiences # ---------------------------- update critic ---------------------------- # # Get predicted next-state actions and Q values from target models actions_next = self.actor_target(next_states) Q_targets_next = self.critic_target(next_states, actions_next) # Compute Q targets for current states (y_i) Q_targets = rewards + (gamma * Q_targets_next * (1 - dones)) # Compute critic loss Q_expected = self.critic_local(states, actions) if self.prioritized_reply: # 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 critic loss critic_loss = F.mse_loss(Q_expected, Q_targets) # Minimize the loss self.critic_optimizer.zero_grad() critic_loss.backward() torch.nn.utils.clip_grad_norm_( self.critic_local.parameters(), 1) # Added because I got converge problems # then I found in adaptationio's solution (https://github.com/adaptationio/DDPG-Continuous-Control), he added this. # Here to see the purpose of doing so: https://discuss.pytorch.org/t/about-torch-nn-utils-clip-grad-norm/13873 self.critic_optimizer.step() # ---------------------------- update actor ---------------------------- # # Compute actor loss actions_pred = self.actor_local(states) actor_loss = -self.critic_local(states, actions_pred).mean() # Minimize the loss self.actor_optimizer.zero_grad() actor_loss.backward() self.actor_optimizer.step() # ----------------------- update target networks ----------------------- # self.soft_update(self.critic_local, self.critic_target, TAU) self.soft_update(self.actor_local, self.actor_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)