class Agent(): """Interacts with and learns from the environment.""" def __init__(self, state_size: int, action_size: int, seed: int, n_agent: int): """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.n_agent = n_agent self.seed = random.seed(seed) self.global_step = 0 self.update_step = 0 # Initialize actor and critic local and target networks self.actor = Actor(state_size, action_size, seed, ACTOR_NETWORK_LINEAR_SIZES, batch_normalization=ACTOR_BATCH_NORM).to(device) self.actor_target = Actor( state_size, action_size, seed, ACTOR_NETWORK_LINEAR_SIZES, batch_normalization=ACTOR_BATCH_NORM).to(device) self.critic = Critic(state_size, action_size, seed, CRITIC_NETWORK_LINEAR_SIZES, batch_normalization=CRITIC_BATCH_NORM).to(device) self.critic_second = Critic( state_size, action_size, seed, CRITIC_SECOND_NETWORK_LINEAR_SIZES, batch_normalization=CRITIC_BATCH_NORM).to(device) self.critic_second_target = Critic( state_size, action_size, seed, CRITIC_SECOND_NETWORK_LINEAR_SIZES, batch_normalization=CRITIC_BATCH_NORM).to(device) self.critic_target = Critic( state_size, action_size, seed, CRITIC_NETWORK_LINEAR_SIZES, batch_normalization=CRITIC_BATCH_NORM).to(device) self.actor_optimizer = optim.Adam(self.actor.parameters(), lr=ACTOR_LEARNING_RATE) self.critic_optimizer = optim.Adam(self.critic.parameters(), lr=CRITIC_LEARNING_RATE) self.critic_second_optimizer = optim.Adam( self.critic_second.parameters(), lr=CRITIC_LEARNING_RATE) # 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] * n_agent self.noise = OUNoise(action_size, seed, decay_period=50) # Copy parameters from local network to target network for target_param, param in zip(self.actor_target.parameters(), self.actor.parameters()): target_param.data.copy_(param.data) for target_param, param in zip(self.critic_target.parameters(), self.critic.parameters()): target_param.data.copy_(param.data) for target_param, param in zip(self.critic_second_target.parameters(), self.critic_second.parameters()): target_param.data.copy_(param.data) def step(self, state: np.array, action, reward, next_state, done, i_agent): # Save experience in replay memory self.memory.add(state, action, reward, next_state, done) # Learn every UPDATE_EVERY time steps. self.t_step[i_agent] = (self.t_step[i_agent] + 1) % UPDATE_EVERY # Learn, if enough samples are available in memory every UPDATE_EVERY if len(self.memory) > BATCH_SIZE and (not any(self.t_step)): for _ in range(LEARN_TIMES): experiences = self.memory.sample() self.learn(experiences, GAMMA) def noise_reset(self): self.noise.reset() def save_model(self, checkpoint_path: str = "./checkpoints/"): torch.save(self.actor.state_dict(), f"{checkpoint_path}/actor.pt") torch.save(self.critic.state_dict(), f"{checkpoint_path}/critic.pt") def load_model(self, checkpoint_path: str = "./checkpoints/checkpoint.pt"): self.actor.load_state_dict(torch.load(f"{checkpoint_path}/actor.pt")) self.critic.load_state_dict(torch.load(f"{checkpoint_path}/critic.pt")) def act(self, states: np.array, step: int): """Returns actions for given state as per current policy. Params ====== state (array_like): current state """ self.global_step += 1 if self.global_step < WARM_UP_STEPS: action_values = np.random.rand(self.n_agent, self.action_size) return action_values states = torch.from_numpy(states).float().to(device) self.actor.eval() with torch.no_grad(): action_values = self.actor(states).cpu().data.numpy() self.actor.train() action_values = [ self.noise.get_action(action, t=step) for action in action_values ] return action_values def learn(self, experiences: tuple, gamma=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 """ self.update_step += 1 states, actions, rewards, next_states, dones = experiences # Critic loss mask = torch.tensor(1 - dones).detach().to(device) Q_values = self.critic(states, actions) Q_values_second = self.critic_second(states, actions) next_actions = self.actor_target(next_states) next_Q = torch.min( self.critic_target(next_states, next_actions.detach()), self.critic_second_target(next_states, next_actions.detach())) Q_prime = rewards + gamma * next_Q * mask if self.update_step % CRITIC_UPDATE_EVERY == 0: critic_loss = F.mse_loss(Q_values, Q_prime.detach()) critic_second_loss = F.mse_loss(Q_values_second, Q_prime.detach()) # Update first critic network self.critic_optimizer.zero_grad() critic_loss.backward() if CRITIC_GRADIENT_CLIPPING_VALUE: torch.nn.utils.clip_grad_norm_(self.critic.parameters(), CRITIC_GRADIENT_CLIPPING_VALUE) self.critic_optimizer.step() # Update second critic network self.critic_second_optimizer.zero_grad() critic_second_loss.backward() if CRITIC_GRADIENT_CLIPPING_VALUE: torch.nn.utils.clip_grad_norm_(self.critic_second.parameters(), CRITIC_GRADIENT_CLIPPING_VALUE) self.critic_second_optimizer.step() # Actor loss if self.update_step % POLICY_UPDATE_EVERY == 0: policy_loss = -self.critic(states, self.actor(states)).mean() # Update actor network self.actor_optimizer.zero_grad() policy_loss.backward() if ACTOR_GRADIENT_CLIPPING_VALUE: torch.nn.utils.clip_grad_norm_(self.actor.parameters(), ACTOR_GRADIENT_CLIPPING_VALUE) self.actor_optimizer.step() self.actor_soft_update() self.critic_soft_update() self.critic_second_soft_update() def actor_soft_update(self, tau: float = TAU): """Soft update for actor target network Args: tau (float, optional). Defaults to TAU. """ for target_param, param in zip(self.actor_target.parameters(), self.actor.parameters()): target_param.data.copy_(tau * param.data + (1.0 - tau) * target_param.data) def critic_soft_update(self, tau: float = TAU): """Soft update for critic target network Args: tau (float, optional). Defaults to TAU. """ for target_param, param in zip(self.critic_target.parameters(), self.critic.parameters()): target_param.detach_() target_param.data.copy_(tau * param.data + (1.0 - tau) * target_param.data) def critic_second_soft_update(self, tau: float = TAU): """Soft update for critic target network Args: tau (float, optional). Defaults to TAU. """ for target_param, param in zip(self.critic_second_target.parameters(), self.critic_second.parameters()): target_param.detach_() target_param.data.copy_(tau * param.data + (1.0 - tau) * target_param.data)
class DDPGAgent(object): """ General class for DDPG agents (policy, critic, target policy, target critic, exploration noise) """ def __init__(self, num_in_pol, num_out_pol, num_in_critic, hidden_dim_actor=120, hidden_dim_critic=64,lr_actor=0.01,lr_critic=0.01,batch_size=64, max_episode_len=100,tau=0.02,gamma = 0.99,agent_name='one', discrete_action=False): """ Inputs: num_in_pol (int): number of dimensions for policy input num_out_pol (int): number of dimensions for policy output num_in_critic (int): number of dimensions for critic input """ self.policy = Actor(num_in_pol, num_out_pol, hidden_dim=hidden_dim_actor, discrete_action=discrete_action) self.critic = Critic(num_in_pol, 1,num_out_pol, hidden_dim=hidden_dim_critic) self.target_policy = Actor(num_in_pol, num_out_pol, hidden_dim=hidden_dim_actor, discrete_action=discrete_action) self.target_critic = Critic(num_in_pol, 1,num_out_pol, hidden_dim=hidden_dim_critic) hard_update(self.target_policy, self.policy) hard_update(self.target_critic, self.critic) self.policy_optimizer = Adam(self.policy.parameters(), lr=lr_actor) self.critic_optimizer = Adam(self.critic.parameters(), lr=lr_critic,weight_decay=0) self.policy = self.policy.float() self.critic = self.critic.float() self.target_policy = self.target_policy.float() self.target_critic = self.target_critic.float() self.agent_name = agent_name self.gamma = gamma self.tau = tau self.batch_size = batch_size #self.replay_buffer = ReplayBuffer(1e7) self.replay_buffer = ReplayBufferOption(500000,self.batch_size,12) self.max_replay_buffer_len = batch_size * max_episode_len self.replay_sample_index = None self.niter = 0 self.eps = 5.0 self.eps_decay = 1/(250*5) self.exploration = OUNoise(num_out_pol) self.discrete_action = discrete_action self.num_history = 2 self.states = [] self.actions = [] self.rewards = [] self.next_states = [] self.dones = [] def reset_noise(self): if not self.discrete_action: self.exploration.reset() def scale_noise(self, scale): if self.discrete_action: self.exploration = scale else: self.exploration.scale = scale def act(self, obs, explore=False): """ Take a step forward in environment for a minibatch of observations Inputs: obs : Observations for this agent explore (boolean): Whether or not to add exploration noise Outputs: action (PyTorch Variable): Actions for this agent """ #obs = obs.reshape(1,48) state = Variable(torch.Tensor(obs),requires_grad=False) self.policy.eval() with torch.no_grad(): action = self.policy(state) self.policy.train() # continuous action if explore: action += Variable(Tensor(self.eps * self.exploration.sample()),requires_grad=False) action = torch.clamp(action, min=-1, max=1) return action def step(self, agent_id, state, action, reward, next_state, done,t_step): self.states.append(state) self.actions.append(action) self.rewards.append(reward) self.next_states.append(next_state) self.dones.append(done) #self.replay_buffer.add(state, action, reward, next_state, done) if t_step % self.num_history == 0: # Save experience / reward self.replay_buffer.add(self.states, self.actions, self.rewards, self.next_states, self.dones) self.states = [] self.actions = [] self.rewards = [] self.next_states = [] self.dones = [] # Learn, if enough samples are available in memory if len(self.replay_buffer) > self.batch_size: obs, acs, rews, next_obs, don = self.replay_buffer.sample() self.update(agent_id ,obs, acs, rews, next_obs, don,t_step) def update(self, agent_id, obs, acs, rews, next_obs, dones ,t_step, logger=None): obs = torch.from_numpy(obs).float() acs = torch.from_numpy(acs).float() rews = torch.from_numpy(rews[:,agent_id]).float() next_obs = torch.from_numpy(next_obs).float() dones = torch.from_numpy(dones[:,agent_id]).float() acs = acs.view(-1,2) # --------- update critic ------------ # self.critic_optimizer.zero_grad() all_trgt_acs = self.target_policy(next_obs) target_value = (rews + self.gamma * self.target_critic(next_obs,all_trgt_acs) * (1 - dones)) actual_value = self.critic(obs,acs) vf_loss = MSELoss(actual_value, target_value.detach()) # Minimize the loss vf_loss.backward() #torch.nn.utils.clip_grad_norm_(self.critic.parameters(), 1) self.critic_optimizer.step() # --------- update actor --------------- # self.policy_optimizer.zero_grad() if self.discrete_action: curr_pol_out = self.policy(obs) curr_pol_vf_in = gumbel_softmax(curr_pol_out, hard=True) else: curr_pol_out = self.policy(obs) curr_pol_vf_in = curr_pol_out pol_loss = -self.critic(obs,curr_pol_vf_in).mean() #pol_loss += (curr_pol_out**2).mean() * 1e-3 pol_loss.backward() torch.nn.utils.clip_grad_norm_(self.policy.parameters(), 1) self.policy_optimizer.step() self.update_all_targets() self.eps -= self.eps_decay self.eps = max(self.eps, 0) if logger is not None: logger.add_scalars('agent%i/losses' % self.agent_name, {'vf_loss': vf_loss, 'pol_loss': pol_loss}, self.niter) def update_all_targets(self): """ Update all target networks (called after normal updates have been performed for each agent) """ soft_update(self.critic, self.target_critic, self.tau) soft_update(self.policy, self.target_policy, self.tau) def get_params(self): return {'policy': self.policy.state_dict(), 'critic': self.critic.state_dict(), 'target_policy': self.target_policy.state_dict(), 'target_critic': self.target_critic.state_dict(), 'policy_optimizer': self.policy_optimizer.state_dict(), 'critic_optimizer': self.critic_optimizer.state_dict()} def load_params(self, params): self.policy.load_state_dict(params['policy']) self.critic.load_state_dict(params['critic']) self.target_policy.load_state_dict(params['target_policy']) self.target_critic.load_state_dict(params['target_critic']) self.policy_optimizer.load_state_dict(params['policy_optimizer']) self.critic_optimizer.load_state_dict(params['critic_optimizer'])
class MADDPG(object): device = 'cuda' if torch.cuda.is_available() else 'cpu' def __init__(self, state_dim, action_dim, max_action, agent_n, logger): # 存在于 GPU 的神经网络 self.actor = Actor(state_dim, action_dim, max_action).to(self.device) # origin_network self.actor_target = Actor(state_dim, action_dim, max_action).to(self.device) # target_network self.actor_target.load_state_dict(self.actor.state_dict( )) # initiate actor_target with actor's parameters # pytorch 中的 tensor 默认requires_grad 属性为false,即不参与梯度传播运算,特别地,opimizer中模型参数是会参与梯度优化的 self.actor_optimizer = optim.Adam( self.actor.parameters(), pdata.LEARNING_RATE) # 以pdata.LEARNING_RATE指定学习率优化actor中的参数 self.critic = CriticCentral(agent_n).to(self.device) self.critic_target = CriticCentral(agent_n).to(self.device) self.critic_target.load_state_dict(self.critic.state_dict()) self.critic_optimizer = optim.Adam(self.critic.parameters(), pdata.LEARNING_RATE) # self.replay_buffer 取消:每个Agent不再有独立的经验池 self.writer = SummaryWriter(pdata.DIRECTORY + 'runs') self.num_critic_update_iteration = 0 self.num_actor_update_iteration = 0 self.num_training = 0 self.logger = logger def select_action(self, state): state = torch.FloatTensor(state.reshape(1, -1)).to(self.device) # numpy.ndarray.flatten(): 返回一个 ndarray对象的copy,并且将该ndarray压缩成一维数组 action = self.actor(state).cpu().data.numpy().flatten() return action # 未归一化的 def select_target_actions(self, states): # state 的输入必须是归一化的[[],[]] np.ndarray united_state = torch.FloatTensor(states).to(self.device) action = self.actor_target(united_state).cpu().data.numpy() return action # 未归一化的 def select_current_actions(self, states): united_state = torch.FloatTensor(states).to(self.device) action = self.actor(united_state).cpu().data.numpy() return action # 未归一化的 # update the parameters in actor network and critic network def update(self, central_replay_buffer_ma, agent_list, agent_idx): critic_loss_list = [] actor_performance_list = [] # Sample replay buffer: [(united_normalized_states, united_normalized_next_states, united__normalized_action, [reward_1, ...], done), (...)] x, y, u, r, d = central_replay_buffer_ma.sample( pdata.BATCH_SIZE ) # 随机获取 batch_size 个五元组样本(sample random minibatch) # TODO: 从这里以下要改造成MADDPG的范式 next_actions = [ agent_list[i].select_target_actions( y[:, i * pdata.STATE_DIMENSION:i * pdata.STATE_DIMENSION + pdata.STATE_DIMENSION]) for i in range(len(agent_list)) ] next_actions = np.concatenate(next_actions, axis=1) united_next_action_batch = torch.FloatTensor(next_actions).to( self.device) united_state_batch = torch.FloatTensor(x).to(self.device) united_action_batch = torch.FloatTensor(u).to(self.device) united_next_state_batch = torch.FloatTensor(y).to(self.device) done = torch.FloatTensor(d).to(self.device) reward_batch = torch.FloatTensor(r[:, agent_idx]) # torch.Size([64]) reward_batch = reward_batch[:, None].to(self.device) # torch.Size([64,1]) target_Q = self.critic_target( united_next_state_batch, united_next_action_batch) # torch.Size([64,1]) target_Q = reward_batch + ( (1 - done) * pdata.GAMMA * target_Q).detach() # batch_size个维度 current_Q = self.critic(united_state_batch, united_action_batch) critic_loss = F.mse_loss(current_Q, target_Q) self.writer.add_scalar('critic_loss', critic_loss, global_step=self.num_critic_update_iteration) # self.logger.write_to_log('critic_loss:{loss}'.format(loss=critic_loss)) # self.logger.add_to_critic_buffer(critic_loss.item()) critic_loss_list.append(critic_loss.item()) self.critic_optimizer.zero_grad() # zeros the gradient buffer critic_loss.backward() # back propagation on a dynamic graph self.critic_optimizer.step() current_actions = [ agent_list[i].select_current_actions( x[:, i * pdata.STATE_DIMENSION:i * pdata.STATE_DIMENSION + pdata.STATE_DIMENSION]) for i in range(len(agent_list)) ] current_actions_batch = torch.FloatTensor( np.concatenate(current_actions, axis=1)).to(self.device) actor_loss = -self.critic(united_state_batch, current_actions_batch).mean() self.writer.add_scalar('actor_performance', actor_loss, global_step=self.num_actor_update_iteration) # self.logger.write_to_log('actor_loss:{loss}'.format(loss=actor_loss)) # self.logger.add_to_actor_buffer(actor_loss.item()) actor_performance_list.append(actor_loss.item()) # Optimize the actor self.actor_optimizer.zero_grad( ) # Clears the gradients of all optimized torch.Tensor actor_loss.backward() self.actor_optimizer.step() # perform a single optimization step # 这里是两个 target网络的 soft update for param, target_param in zip(self.critic.parameters(), self.critic_target.parameters()): target_param.data.copy_(pdata.TAU * param.data + (1 - pdata.TAU) * target_param.data) for param, target_param in zip(self.actor.parameters(), self.actor_target.parameters()): target_param.data.copy_(pdata.TAU * param.data + (1 - pdata.TAU) * target_param.data) self.num_actor_update_iteration += 1 self.num_critic_update_iteration += 1 actor_performance = np.mean(np.array(actor_performance_list)).item() self.logger.add_to_actor_buffer(actor_performance) critic_loss = np.mean(critic_loss_list).item() self.logger.add_to_critic_buffer(critic_loss) def save(self, mark_str): torch.save(self.actor.state_dict(), pdata.DIRECTORY + mark_str + '_actor_ma.pth') torch.save(self.critic.state_dict(), pdata.DIRECTORY + mark_str + '_critic_ma.pth') def load(self, mark_str): file_actor = pdata.DIRECTORY + mark_str + '_actor_ma.pth' file_critic = pdata.DIRECTORY + mark_str + '_critic_ma.pth' if os.path.exists(file_actor) and os.path.exists(file_critic): self.actor.load_state_dict(torch.load(file_actor)) self.critic.load_state_dict(torch.load(file_critic))
class DDPG(object): device = 'cuda' if torch.cuda.is_available() else 'cpu' veer = 'straight' origin = 'west' # state_dim : 状态维度 # action_dim: 动作维度 # max_action:动作限制向量 def __init__(self, state_dim, action_dim, max_action, origin_str, veer_str, logger): # 存在于 GPU 的神经网络 self.actor = Actor(state_dim, action_dim, max_action).to(self.device) # origin_network self.actor_target = Actor(state_dim, action_dim, max_action).to(self.device) # target_network self.actor_target.load_state_dict(self.actor.state_dict( )) # initiate actor_target with actor's parameters # pytorch 中的 tensor 默认requires_grad 属性为false,即不参与梯度传播运算,特别地,opimizer中模型参数是会参与梯度优化的 self.actor_optimizer = optim.Adam( self.actor.parameters(), pdata.LEARNING_RATE) # 以pdata.LEARNING_RATE指定学习率优化actor中的参数 self.critic = Critic(state_dim, action_dim).to(self.device) self.critic_target = Critic(state_dim, action_dim).to(self.device) self.critic_target.load_state_dict(self.critic.state_dict()) self.critic_optimizer = optim.Adam(self.critic.parameters(), pdata.LEARNING_RATE) # self.replay_buffer = Replay_buffer() # initiate replay-buffer self.replay_buffer = FilterReplayBuffer() self.writer = SummaryWriter(pdata.DIRECTORY + 'runs') self.num_critic_update_iteration = 0 self.num_actor_update_iteration = 0 self.num_training = 0 self.veer = veer_str self.origin = origin_str self.logger = logger def select_action(self, state): state = torch.FloatTensor(state.reshape(1, -1)).to(self.device) # numpy.ndarray.flatten(): 返回一个 ndarray对象的copy,并且将该ndarray压缩成一维数组 action = self.actor(state).cpu().data.numpy().flatten() return action # update the parameters in actor network and critic network # 只有 replay_buffer 中的 storage 超过了样本数量才会调用 update函数 def update(self): critic_loss_list = [] actor_performance_list = [] for it in range(pdata.UPDATE_ITERATION): # Sample replay buffer x, y, u, r, d = self.replay_buffer.sample( pdata.BATCH_SIZE ) # 随机获取 batch_size 个五元组样本(sample random minibatch) state = torch.FloatTensor(x).to(self.device) action = torch.FloatTensor(u).to(self.device) next_state = torch.FloatTensor(y).to(self.device) done = torch.FloatTensor(d).to(self.device) reward = torch.FloatTensor(r).to(self.device) # Compute the target Q value —— Q(S', A') is an value evaluated with next_state and predicted action # 这里的 target_Q 是 sample 个 一维tensor target_Q = self.critic_target(next_state, self.actor_target(next_state)) # detach(): Return a new tensor, detached from the current graph # evaluate Q: targetQ = R + γQ'(s', a') target_Q = reward + ( (1 - done) * pdata.GAMMA * target_Q).detach() # batch_size个维度 # Get current Q estimate current_Q = self.critic(state, action) # 1 维 # Compute critic loss : a mean-square error # 由论文,critic_loss 其实计算的是每个样本估计值与每个critic网络输出的均值方差 # torch.nn.functional.mse_loss 为计算tensor中各个元素的的均值方差 critic_loss = F.mse_loss(current_Q, target_Q) self.writer.add_scalar( 'critic_loss', critic_loss, global_step=self.num_critic_update_iteration) # self.logger.write_to_log('critic_loss:{loss}'.format(loss=critic_loss)) # self.logger.add_to_critic_buffer(critic_loss.item()) critic_loss_list.append(critic_loss.item()) # Optimize the critic self.critic_optimizer.zero_grad() # zeros the gradient buffer critic_loss.backward() # back propagation on a dynamic graph self.critic_optimizer.step() # Compute actor loss # actor_loss:见论文中对公式 (6) 的理解 # mean():对tensor对象求所有element的均值 # backward() 以梯度下降的方式更新参数,则将 actor_loss 设置为反向梯度,这样参数便往梯度上升方向更新 actor_loss = -self.critic(state, self.actor(state)).mean() self.writer.add_scalar('actor_performance', actor_loss, global_step=self.num_actor_update_iteration) # self.logger.write_to_log('actor_loss:{loss}'.format(loss=actor_loss)) # self.logger.add_to_actor_buffer(actor_loss.item()) actor_performance_list.append(actor_loss.item()) # Optimize the actor self.actor_optimizer.zero_grad( ) # Clears the gradients of all optimized torch.Tensor actor_loss.backward() self.actor_optimizer.step() # perform a single optimization step # 这里是两个 target网络的 soft update for param, target_param in zip(self.critic.parameters(), self.critic_target.parameters()): target_param.data.copy_(pdata.TAU * param.data + (1 - pdata.TAU) * target_param.data) for param, target_param in zip(self.actor.parameters(), self.actor_target.parameters()): target_param.data.copy_(pdata.TAU * param.data + (1 - pdata.TAU) * target_param.data) self.num_actor_update_iteration += 1 self.num_critic_update_iteration += 1 actor_performance = np.mean(np.array(actor_performance_list)).item() self.logger.add_to_actor_buffer(actor_performance) critic_loss = np.mean(critic_loss_list).item() self.logger.add_to_critic_buffer(critic_loss) def save(self, mark_str): torch.save(self.actor.state_dict(), pdata.DIRECTORY + mark_str + '_actor.pth') torch.save(self.critic.state_dict(), pdata.DIRECTORY + mark_str + '_critic.pth') def load(self, mark_str): file_actor = pdata.DIRECTORY + mark_str + '_actor.pth' file_critic = pdata.DIRECTORY + mark_str + '_critic.pth' if os.path.exists(file_actor) and os.path.exists(file_critic): self.actor.load_state_dict(torch.load(file_actor)) self.critic.load_state_dict(torch.load(file_critic))
class DDPGAgent: def __init__(self, state_size=24, action_size=2, seed=1, num_agents=2): self.state_size = state_size self.action_size = action_size self.seed = random.seed(seed) self.num_agents = num_agents # DDPG specific configuration hidden_size = 512 self.CHECKPOINT_FOLDER = './' # Defining networks self.actor = Actor(state_size, hidden_size, action_size).to(device) self.actor_target = Actor(state_size, hidden_size, action_size).to(device) self.critic = Critic(state_size, self.action_size, hidden_size, 1).to(device) self.critic_target = Critic(state_size, self.action_size, hidden_size, 1).to(device) self.optimizer_actor = optim.Adam(self.actor.parameters(), lr=ACTOR_LR) self.optimizer_critic = optim.Adam(self.critic.parameters(), lr=CRITIC_LR) # Noise self.noises = OUNoise((num_agents, action_size), seed) # Initialize replay buffer self.memory = ReplayBuffer(action_size, BUFFER_SIZE, BATCH_SIZE, seed) def act(self, state, add_noise=True): ''' Returns action to be taken based on state provided as the input ''' state = torch.from_numpy(state).float().to(device) actions = np.zeros((self.num_agents, self.action_size)) self.actor.eval() with torch.no_grad(): for agent_num, state in enumerate(state): action = self.actor(state).cpu().data.numpy() actions[agent_num, :] = action self.actor.train() if add_noise: actions += self.noises.sample() return np.clip(actions, -1, 1) def reset(self): self.noises.reset() def learn(self, experiences): ''' Trains the actor critic network using experiences ''' states, actions, rewards, next_states, dones = experiences # Update Critic actions_next = self.actor_target(next_states) # print(next_states.shape, actions_next.shape) Q_targets_next = self.critic_target(next_states, actions_next) Q_targets = rewards + GAMMA*Q_targets_next*(1-dones) Q_expected = self.critic(states, actions) critic_loss = F.mse_loss(Q_expected, Q_targets) self.optimizer_critic.zero_grad() critic_loss.backward() self.optimizer_critic.step() # Update Actor actions_pred = self.actor(states) actor_loss = -self.critic(states, actions_pred).mean() self.optimizer_actor.zero_grad() actor_loss.backward() self.optimizer_actor.step() # Updating the local networks self.soft_update(self.critic, self.critic_target) self.soft_update(self.actor, self.actor_target) def soft_update(self, model, model_target): tau = TAU for target_param, local_param in zip(model_target.parameters(), model.parameters()): target_param.data.copy_(tau * local_param.data + (1.0 - tau) * target_param.data) def step(self, state, action, reward, next_state, done): ''' Adds experience to memory and learns if the memory contains sufficient samples ''' for i in range(self.num_agents): self.memory.add(state[i, :], action[i, :], reward[i], next_state[i, :], done[i]) if len(self.memory) > BATCH_SIZE: # print("Now Learning") experiences = self.memory.sample() self.learn(experiences) def checkpoint(self): ''' Saves the actor critic network on disk ''' torch.save(self.actor.state_dict(), self.CHECKPOINT_FOLDER + 'checkpoint_actor.pth') torch.save(self.critic.state_dict(), self.CHECKPOINT_FOLDER + 'checkpoint_critic.pth') def load(self): ''' Loads the actor critic network from disk ''' self.actor.load_state_dict(torch.load(self.CHECKPOINT_FOLDER + 'checkpoint_actor.pth')) self.actor_target.load_state_dict(torch.load(self.CHECKPOINT_FOLDER + 'checkpoint_actor.pth')) self.critic.load_state_dict(torch.load(self.CHECKPOINT_FOLDER + 'checkpoint_critic.pth')) self.critic_target.load_state_dict(torch.load(self.CHECKPOINT_FOLDER + 'checkpoint_critic.pth'))
class DDPG: def __init__(self, gamma, tau, hidden_size, num_inputs, action_space, train_mode, alpha, replay_size, normalize_obs=True, normalize_returns=False, critic_l2_reg=1e-2): if torch.cuda.is_available(): self.device = torch.device('cuda') torch.backends.cudnn.enabled = False self.Tensor = torch.cuda.FloatTensor else: self.device = torch.device('cpu') self.Tensor = torch.FloatTensor self.alpha = alpha self.train_mode = train_mode self.num_inputs = num_inputs self.action_space = action_space self.critic_l2_reg = critic_l2_reg self.actor = Actor(hidden_size, self.num_inputs, self.action_space).to(self.device) self.adversary = Actor(hidden_size, self.num_inputs, self.action_space).to(self.device) if self.train_mode: self.actor_target = Actor(hidden_size, self.num_inputs, self.action_space).to(self.device) self.actor_perturbed = Actor(hidden_size, self.num_inputs, self.action_space).to(self.device) self.actor_optim = Adam(self.actor.parameters(), lr=1e-4) self.critic = Critic(hidden_size, self.num_inputs, self.action_space).to(self.device) self.critic_target = Critic(hidden_size, self.num_inputs, self.action_space).to(self.device) self.critic_optim = Adam(self.critic.parameters(), lr=1e-3, weight_decay=critic_l2_reg) self.adversary_target = Actor(hidden_size, self.num_inputs, self.action_space).to(self.device) self.adversary_perturbed = Actor(hidden_size, self.num_inputs, self.action_space).to(self.device) self.adversary_optim = Adam(self.adversary.parameters(), lr=1e-4) hard_update( self.adversary_target, self.adversary) # Make sure target is with the same weight hard_update(self.actor_target, self.actor) # Make sure target is with the same weight hard_update(self.critic_target, self.critic) self.gamma = gamma self.tau = tau self.normalize_observations = normalize_obs self.normalize_returns = normalize_returns if self.normalize_observations: self.obs_rms = RunningMeanStd(shape=num_inputs) else: self.obs_rms = None if self.normalize_returns: self.ret_rms = RunningMeanStd(shape=1) self.ret = 0 self.cliprew = 10.0 else: self.ret_rms = None self.memory = ReplayMemory(replay_size) def eval(self): self.actor.eval() self.adversary.eval() if self.train_mode: self.critic.eval() def train(self): self.actor.train() self.adversary.train() if self.train_mode: self.critic.train() def select_action(self, state, action_noise=None, param_noise=None, mdp_type='mdp'): state = normalize( Variable(state).to(self.device), self.obs_rms, self.device) if mdp_type != 'mdp': if mdp_type == 'nr_mdp': if param_noise is not None: mu = self.actor_perturbed(state) else: mu = self.actor(state) mu = mu.data if action_noise is not None: mu += self.Tensor(action_noise()).to(self.device) mu = mu.clamp(-1, 1) * (1 - self.alpha) if param_noise is not None: adv_mu = self.adversary_perturbed(state) else: adv_mu = self.adversary(state) adv_mu = adv_mu.data.clamp(-1, 1) * self.alpha mu += adv_mu else: # mdp_type == 'pr_mdp': if np.random.rand() < (1 - self.alpha): if param_noise is not None: mu = self.actor_perturbed(state) else: mu = self.actor(state) mu = mu.data if action_noise is not None: mu += self.Tensor(action_noise()).to(self.device) mu = mu.clamp(-1, 1) else: if param_noise is not None: mu = self.adversary_perturbed(state) else: mu = self.adversary(state) mu = mu.data.clamp(-1, 1) else: if param_noise is not None: mu = self.actor_perturbed(state) else: mu = self.actor(state) mu = mu.data if action_noise is not None: mu += self.Tensor(action_noise()).to(self.device) mu = mu.clamp(-1, 1) return mu def update_robust(self, state_batch, action_batch, reward_batch, mask_batch, next_state_batch, adversary_update, mdp_type, robust_update_type): # TRAIN CRITIC if robust_update_type == 'full': if mdp_type == 'nr_mdp': next_action_batch = (1 - self.alpha) * self.actor_target(next_state_batch) \ + self.alpha * self.adversary_target(next_state_batch) next_state_action_values = self.critic_target( next_state_batch, next_action_batch) else: # mdp_type == 'pr_mdp': next_action_actor_batch = self.actor_target(next_state_batch) next_action_adversary_batch = self.adversary_target( next_state_batch) next_state_action_values = self.critic_target(next_state_batch, next_action_actor_batch) * ( 1 - self.alpha) \ + self.critic_target(next_state_batch, next_action_adversary_batch) * self.alpha expected_state_action_batch = reward_batch + self.gamma * mask_batch * next_state_action_values self.critic_optim.zero_grad() state_action_batch = self.critic(state_batch, action_batch) value_loss = F.mse_loss(state_action_batch, expected_state_action_batch) value_loss.backward() self.critic_optim.step() value_loss = value_loss.item() else: value_loss = 0 if adversary_update: # TRAIN ADVERSARY self.adversary_optim.zero_grad() if mdp_type == 'nr_mdp': with torch.no_grad(): real_action = self.actor_target(next_state_batch) action = ( 1 - self.alpha ) * real_action + self.alpha * self.adversary(next_state_batch) adversary_loss = self.critic(state_batch, action) else: # mdp_type == 'pr_mdp' action = self.adversary(next_state_batch) adversary_loss = self.critic(state_batch, action) * self.alpha adversary_loss = adversary_loss.mean() adversary_loss.backward() self.adversary_optim.step() adversary_loss = adversary_loss.item() policy_loss = 0 else: if robust_update_type == 'full': # TRAIN ACTOR self.actor_optim.zero_grad() if mdp_type == 'nr_mdp': with torch.no_grad(): adversary_action = self.adversary_target( next_state_batch) action = (1 - self.alpha) * self.actor( next_state_batch) + self.alpha * adversary_action policy_loss = -self.critic(state_batch, action) else: # mdp_type == 'pr_mdp': action = self.actor(next_state_batch) policy_loss = -self.critic(state_batch, action) * ( 1 - self.alpha) policy_loss = policy_loss.mean() policy_loss.backward() self.actor_optim.step() policy_loss = policy_loss.item() adversary_loss = 0 else: policy_loss = 0 adversary_loss = 0 return value_loss, policy_loss, adversary_loss def update_non_robust(self, state_batch, action_batch, reward_batch, mask_batch, next_state_batch): # TRAIN CRITIC next_action_batch = self.actor_target(next_state_batch) next_state_action_values = self.critic_target(next_state_batch, next_action_batch) expected_state_action_batch = reward_batch + self.gamma * mask_batch * next_state_action_values self.critic_optim.zero_grad() state_action_batch = self.critic(state_batch, action_batch) value_loss = F.mse_loss(state_action_batch, expected_state_action_batch) value_loss.backward() self.critic_optim.step() # TRAIN ACTOR self.actor_optim.zero_grad() action = self.actor(next_state_batch) policy_loss = -self.critic(state_batch, action) policy_loss = policy_loss.mean() policy_loss.backward() self.actor_optim.step() policy_loss = policy_loss.item() adversary_loss = 0 return value_loss.item(), policy_loss, adversary_loss def store_transition(self, state, action, mask, next_state, reward): B = state.shape[0] for b in range(B): self.memory.push(state[b], action[b], mask[b], next_state[b], reward[b]) if self.normalize_observations: self.obs_rms.update(state[b].cpu().numpy()) if self.normalize_returns: self.ret = self.ret * self.gamma + reward[b] self.ret_rms.update(np.array([self.ret])) if mask[b] == 0: # if terminal is True self.ret = 0 def update_parameters(self, batch_size, mdp_type='mdp', adversary_update=False, exploration_method='mdp'): transitions = self.memory.sample(batch_size) batch = Transition(*zip(*transitions)) if mdp_type != 'mdp': robust_update_type = 'full' elif exploration_method != 'mdp': robust_update_type = 'adversary' else: robust_update_type = None state_batch = normalize( Variable(torch.stack(batch.state)).to(self.device), self.obs_rms, self.device) action_batch = Variable(torch.stack(batch.action)).to(self.device) reward_batch = normalize( Variable(torch.stack(batch.reward)).to(self.device).unsqueeze(1), self.ret_rms, self.device) mask_batch = Variable(torch.stack(batch.mask)).to( self.device).unsqueeze(1) next_state_batch = normalize( Variable(torch.stack(batch.next_state)).to(self.device), self.obs_rms, self.device) if self.normalize_returns: reward_batch = torch.clamp(reward_batch, -self.cliprew, self.cliprew) value_loss = 0 policy_loss = 0 adversary_loss = 0 if robust_update_type is not None: _value_loss, _policy_loss, _adversary_loss = self.update_robust( state_batch, action_batch, reward_batch, mask_batch, next_state_batch, adversary_update, mdp_type, robust_update_type) value_loss += _value_loss policy_loss += _policy_loss adversary_loss += _adversary_loss if robust_update_type != 'full': _value_loss, _policy_loss, _adversary_loss = self.update_non_robust( state_batch, action_batch, reward_batch, mask_batch, next_state_batch) value_loss += _value_loss policy_loss += _policy_loss adversary_loss += _adversary_loss self.soft_update() return value_loss, policy_loss, adversary_loss def soft_update(self): soft_update(self.actor_target, self.actor, self.tau) soft_update(self.adversary_target, self.adversary, self.tau) soft_update(self.critic_target, self.critic, self.tau) def perturb_actor_parameters(self, param_noise): """Apply parameter noise to actor model, for exploration""" hard_update(self.actor_perturbed, self.actor) params = self.actor_perturbed.state_dict() for name in params: if 'ln' in name: pass param = params[name] param += torch.randn(param.shape).to( self.device) * param_noise.current_stddev """Apply parameter noise to adversary model, for exploration""" hard_update(self.adversary_perturbed, self.adversary) params = self.adversary_perturbed.state_dict() for name in params: if 'ln' in name: pass param = params[name] param += torch.randn(param.shape).to( self.device) * param_noise.current_stddev