class DQNG(DQN): def __init__(self, args, env, env_test, logger): super(DQNG, self).__init__(args, env, env_test, logger) def init(self, args, env): names = ['state0', 'action', 'state1', 'reward', 'terminal', 'goal'] self.buffer = ReplayBuffer(limit=int(1e6), names=names.copy()) if args['--imit'] != '0': names.append('expVal') self.bufferImit = ReplayBuffer(limit=int(1e6), names=names.copy()) self.critic = CriticDQNG(args, env) def train(self): if self.buffer.nb_entries > self.batch_size: exp = self.buffer.sample(self.batch_size) s0, a0, s1, r, t, g = [exp[name] for name in self.buffer.names] targets_dqn = self.critic.get_targets_dqn(r, t, s1, g) inputs = [s0, a0, g] loss = self.critic.qvalModel.train_on_batch(inputs, targets_dqn) for i, metric in enumerate(self.critic.qvalModel.metrics_names): self.metrics[metric] += loss[i] if self.args[ '--imit'] != '0' and self.bufferImit.nb_entries > self.batch_size: exp = self.bufferImit.sample(self.batch_size) s0, a0, s1, r, t, g, e = [ exp[name] for name in self.bufferImit.names ] targets_dqn = self.critic.get_targets_dqn(r, t, s1, g) targets = [ targets_dqn, np.zeros((self.batch_size, 1)), np.zeros((self.batch_size, 1)) ] inputs = [s0, a0, g, e] loss = self.critic.imitModel.train_on_batch(inputs, targets) for i, metric in enumerate( self.critic.imitModel.metrics_names): self.imitMetrics[metric] += loss[i] self.critic.target_train() def make_input(self, state, t): input = [np.expand_dims(i, axis=0) for i in [state, self.env.goal]] # temp = self.env.explor_temp(t) input.append(np.expand_dims([0.5], axis=0)) return input
class DDPGG(DDPG): def __init__(self, args, env, env_test, logger): super(DDPGG, self).__init__(args, env, env_test, logger) def init(self, args, env): names = ['s0', 'a', 's1', 'r', 't', 'g'] metrics = ['loss_dqn', 'loss_actor'] self.buffer = ReplayBuffer(limit=int(1e6), names=names.copy(), args=args) self.actorCritic = ActorCriticDDPGG(args, env) for metric in metrics: self.metrics[metric] = 0 def train(self): if self.buffer.nb_entries > self.batch_size: exp = self.buffer.sample(self.batch_size) targets_dqn = self.actorCritic.get_targets_dqn( exp['r'], exp['t'], exp['s1'], exp['g']) inputs = [exp['s0'], exp['a'], exp['g'], targets_dqn] loss_dqn = self.actorCritic.trainQval(inputs) action, criticActionGrads, invertedCriticActionGrads = self.actorCritic.trainActor( [exp['s0'], exp['g']]) self.metrics['loss_dqn'] += np.squeeze(loss_dqn) self.actorCritic.target_train() def make_input(self, state, mode): if mode == 'train': input = [np.expand_dims(i, axis=0) for i in [state, self.env.goal]] else: input = [ np.expand_dims(i, axis=0) for i in [state, self.env_test.goal] ] return input def reset(self): if self.trajectory: self.env.end_episode(self.trajectory) for expe in self.trajectory: self.buffer.append(expe.copy()) if self.args['--her'] != '0': augmented_ep = self.env.augment_episode(self.trajectory) for e in augmented_ep: self.buffer.append(e) self.trajectory.clear() state = self.env.reset() self.episode_step = 0 return state
class Qoff(Agent): def __init__(self, args, env, env_test, logger): super(Qoff, self).__init__(args, env, env_test, logger) self.args = args self.gamma = 0.99 self.lr = 0.1 self.names = ['state0', 'action', 'state1', 'reward', 'terminal'] self.init(args, env) def init(self, args, env): self.critic = np.zeros(shape=(5, 5, 4)) self.buffer = ReplayBuffer(limit=int(1e6), names=self.names) def train(self): if self.buffer.nb_entries > self.batch_size: exp = self.buffer.sample(self.batch_size) s0, a0, s1, r, t, g, m = [exp[name] for name in self.names] for k in range(self.batch_size): target = r[k] + (1 - t[k]) * self.gamma * np.max( self.critic[tuple(s1[k])]) self.critic[tuple(s0[k])][a0[k]] = self.lr * target + \ (1 - self.lr) * self.critic[tuple(s0[k])][a0[k]] def act(self, state): if np.random.rand() < 0.2: action = np.random.randint(self.env.action_space.n) else: action = np.argmax(self.critic[tuple(state)]) return action def reset(self): if self.trajectory: self.env.processEp(self.trajectory) for expe in reversed(self.trajectory): self.buffer.append(expe.copy()) self.trajectory.clear() state = self.env.reset() self.episode_step = 0 return state
class SAC: def __init__(self, env, gamma=0.99, tau=0.005, learning_rate=3e-4, buffer_size=50000, learning_starts=100, train_freq=1, batch_size=64, target_update_interval=1, gradient_steps=1, target_entropy='auto', ent_coef='auto', random_exploration=0.0, discrete=True, regularized=True, feature_extraction="cnn"): self.env = env self.learning_starts = learning_starts self.random_exploration = random_exploration self.train_freq = train_freq self.target_update_interval = target_update_interval self.batch_size = batch_size self.gradient_steps = gradient_steps self.learning_rate = learning_rate self.graph = tf.Graph() with self.graph.as_default(): self.sess = tf.Session(graph=self.graph) self.replay_buffer = ReplayBuffer(buffer_size) self.agent = SACAgent(self.sess, env, discrete=discrete, regularized=regularized, feature_extraction=feature_extraction) self.model = SACModel(self.sess, self.agent, target_entropy, ent_coef, gamma, tau) with self.sess.as_default(): self.sess.run(tf.global_variables_initializer()) self.sess.run(self.model.target_init_op) self.num_timesteps = 0 def train(self, learning_rate): batch_obs, batch_actions, batch_rewards, batch_next_obs, batch_dones = self.replay_buffer.sample( self.batch_size) # print("batch_actions:", batch_actions.shape) # print("self.agent.actions_ph:", self.agent.actions_ph) feed_dict = { self.agent.obs_ph: batch_obs, self.agent.next_obs_ph: batch_next_obs, self.model.rewards_ph: batch_rewards.reshape(self.batch_size, -1), self.model.terminals_ph: batch_dones.reshape(self.batch_size, -1), self.model.learning_rate_ph: learning_rate } if not self.agent.discrete: feed_dict[self.agent.actions_ph] = batch_actions else: batch_actions = batch_actions.reshape(-1) feed_dict[self.agent.actions_ph] = batch_actions policy_loss, qf1_loss, qf2_loss, value_loss, *values = self.sess.run( self.model.step_ops, feed_dict) return policy_loss, qf1_loss, qf2_loss def learn(self, total_timesteps): learning_rate = get_schedule_fn(self.learning_rate) episode_rewards = [0] mb_losses = [] obs = self.env.reset() for step in range(total_timesteps): if self.num_timesteps < self.learning_starts or np.random.rand( ) < self.random_exploration: unscaled_action = self.env.action_space.sample() action = scale_action(self.env.action_space, unscaled_action) else: action = self.agent.step(obs[None]).flatten() unscaled_action = unscale_action(self.env.action_space, action) # print("\nunscaled_action:", unscaled_action) new_obs, reward, done, _ = self.env.step(unscaled_action) self.num_timesteps += 1 self.replay_buffer.add(obs, action, reward, new_obs, done) obs = new_obs if self.num_timesteps % self.train_freq == 0: for grad_step in range(self.gradient_steps): if not self.replay_buffer.can_sample( self.batch_size ) or self.num_timesteps < self.learning_starts: break frac = 1.0 - step / total_timesteps current_lr = learning_rate(frac) mb_losses.append(self.train(current_lr)) if (step + grad_step) % self.target_update_interval == 0: self.sess.run(self.model.target_update_op) episode_rewards[-1] += reward if done: obs = self.env.reset() episode_rewards.append(0) mean_reward = round(float(np.mean(episode_rewards[-101:-1])), 1) loss_str = "/".join([f"{x:.3f}" for x in np.mean(mb_losses, 0) ]) if len(mb_losses) > 0 else "NaN" print(f"Step {step} - reward: {mean_reward} - loss: {loss_str}", end="\n" if step % 500 == 0 else "\r")
class DDPG(Agent): def __init__(self, args, env, env_test, logger): super(DDPG, self).__init__(args, env, env_test, logger) self.args = args self.init(args, env) for metric in self.critic.model.metrics_names: self.metrics[self.critic.model.name + '_' + metric] = 0 def init(self, args, env): names = ['state0', 'action', 'state1', 'reward', 'terminal'] self.buffer = ReplayBuffer(limit=int(1e6), names=names.copy()) self.actorCritic = ActorCriticDDPG(args, env) # self.critic = CriticDDPG(args, env) # self.actor = ActorDDPG(args, env) def train(self): if self.buffer.nb_entries > self.batch_size: exp = self.buffer.sample(self.batch_size) s0, a0, s1, r, t = [exp[name] for name in self.buffer.names] a1 = self.actor.target_model.predict_on_batch(s1) a1 = np.clip(a1, self.env.action_space.low, self.env.action_space.high) q = self.critic.Tmodel.predict_on_batch([s1, a1]) targets = r + (1 - t) * self.critic.gamma * np.squeeze(q) targets = np.clip(targets, self.env.minR / (1 - self.critic.gamma), self.env.maxR) inputs = [s0, a0] loss = self.critic.model.train_on_batch(inputs, targets) for i, metric in enumerate(self.critic.model.metrics_names): self.metrics[metric] += loss[i] # a2 = self.actor.model.predict_on_batch(s0) # grads = self.critic.gradsModel.predict_on_batch([s0, a2]) # low = self.env.action_space.low # high = self.env.action_space.high # for d in range(grads[0].shape[0]): # width = high[d] - low[d] # for k in range(self.batch_size): # if grads[k][d] >= 0: # grads[k][d] *= (high[d] - a2[k][d]) / width # else: # grads[k][d] *= (a2[k][d] - low[d]) / width # self.actor.train(s0, grads) self.actor.target_train() self.critic.target_train() def reset(self): if self.trajectory: T = int(self.trajectory[-1]['terminal']) R = np.sum([ self.env.unshape(exp['reward'], exp['terminal']) for exp in self.trajectory ]) S = len(self.trajectory) self.env.processEp(R, S, T) for expe in reversed(self.trajectory): self.buffer.append(expe.copy()) self.trajectory.clear() state = self.env.reset() self.episode_step = 0 return state def make_input(self, state): input = [np.reshape(state, (1, self.actor.s_dim[0]))] return input def act(self, state): input = self.make_input(state) action = self.actor.model.predict(input, batch_size=1) noise = np.random.normal(0., 0.1, size=action.shape) action = noise + action action = np.clip(action, self.env.action_space.low, self.env.action_space.high) action = action.squeeze() return action
class MADDPG: def __init__(self, env, state_dim: int, action_dim: int, config: Dict, device=None, writer=None): self.logger = logging.getLogger("MADDPG") self.device = device if device is not None else DEVICE self.writer = writer self.env = env self.state_dim = state_dim self.action_dim = action_dim self.agents_number = config['agents_number'] hidden_layers = config.get('hidden_layers', (400, 300)) noise_scale = config.get('noise_scale', 0.2) noise_sigma = config.get('noise_sigma', 0.1) actor_lr = config.get('actor_lr', 1e-3) actor_lr_decay = config.get('actor_lr_decay', 0) critic_lr = config.get('critic_lr', 1e-3) critic_lr_decay = config.get('critic_lr_decay', 0) self.actor_tau = config.get('actor_tau', 0.002) self.critic_tau = config.get('critic_tau', 0.002) create_agent = lambda: DDPGAgent(state_dim, action_dim, agents=self.agents_number, hidden_layers=hidden_layers, actor_lr=actor_lr, actor_lr_decay=actor_lr_decay, critic_lr=critic_lr, critic_lr_decay=critic_lr_decay, noise_scale=noise_scale, noise_sigma=noise_sigma, device=self.device) self.agents = [create_agent() for _ in range(self.agents_number)] self.discount = 0.99 if 'discount' not in config else config['discount'] self.gradient_clip = 1.0 if 'gradient_clip' not in config else config[ 'gradient_clip'] self.warm_up = 1e3 if 'warm_up' not in config else config['warm_up'] self.buffer_size = int( 1e6) if 'buffer_size' not in config else config['buffer_size'] self.batch_size = config.get('batch_size', 128) self.p_batch_size = config.get('p_batch_size', int(self.batch_size // 2)) self.n_batch_size = config.get('n_batch_size', int(self.batch_size // 4)) self.buffer = ReplayBuffer(self.batch_size, self.buffer_size) self.update_every_iterations = config.get('update_every_iterations', 2) self.number_updates = config.get('number_updates', 2) self.reset() def reset(self): self.iteration = 0 self.reset_agents() def reset_agents(self): for agent in self.agents: agent.reset_agent() def step(self, state, action, reward, next_state, done) -> None: if np.isnan(state).any() or np.isnan(next_state).any(): print("State contains NaN. Skipping.") return self.iteration += 1 self.buffer.add(state, action, reward, next_state, done) if self.iteration < self.warm_up: return if len(self.buffer) > self.batch_size and ( self.iteration % self.update_every_iterations) == 0: self.evok_learning() def filter_batch(self, batch, agent_number): states, actions, rewards, next_states, dones = batch agent_states = states[:, agent_number * self.state_dim:(agent_number + 1) * self.state_dim].clone() agent_next_states = next_states[:, agent_number * self.state_dim:(agent_number + 1) * self.state_dim].clone() agent_rewards = rewards.select(1, agent_number).view(-1, 1).clone() agent_dones = dones.select(1, agent_number).view(-1, 1).clone() return (agent_states, states, actions, agent_rewards, agent_next_states, next_states, agent_dones) def evok_learning(self): for _ in range(self.number_updates): for agent_number in range(self.agents_number): batch = self.filter_batch(self.buffer.sample(), agent_number) self.learn(batch, agent_number) # self.update_targets() def act(self, states, noise: Union[None, List] = None): """get actions from all agents in the MADDPG object""" noise = [0] * self.agents_number if noise is None else noise tensor_states = torch.tensor(states).view(-1, self.agents_number, self.state_dim) with torch.no_grad(): actions = [] for agent_number, agent in enumerate(self.agents): agent.actor.eval() actions += agent.act(tensor_states.select(1, agent_number), noise[agent_number]) agent.actor.train() return torch.stack(actions) def learn(self, samples, agent_number: int) -> None: """update the critics and actors of all the agents """ action_offset = agent_number * self.action_dim flatten_actions = lambda a: a.view( -1, self.agents_number * self.action_dim) # No need to flip since there are no paralle agents agent_states, states, actions, rewards, agent_next_states, next_states, dones = samples agent = self.agents[agent_number] next_actions = actions.clone() next_actions[:, action_offset:action_offset + self.action_dim] = agent.target_actor(agent_next_states) # critic loss Q_target_next = agent.target_critic(next_states, flatten_actions(next_actions)) Q_target = rewards + (self.discount * Q_target_next * (1 - dones)) Q_expected = agent.critic(states, actions) critic_loss = F.mse_loss(Q_expected, Q_target) # Minimize the loss agent.critic_optimizer.zero_grad() critic_loss.backward() torch.nn.utils.clip_grad_norm_(agent.critic.parameters(), self.gradient_clip) agent.critic_optimizer.step() # Compute actor loss pred_actions = actions.clone() pred_actions[:, action_offset:action_offset + self.action_dim] = agent.actor(agent_states) actor_loss = -agent.critic(states, flatten_actions(pred_actions)).mean() agent.actor_optimizer.zero_grad() actor_loss.backward() agent.actor_optimizer.step() if self.writer: self.writer.add_scalar(f'agent{agent_number}/critic_loss', critic_loss.item(), self.iteration) self.writer.add_scalar(f'agent{agent_number}/actor_loss', abs(actor_loss.item()), self.iteration) self._soft_update(agent.target_actor, agent.actor, self.actor_tau) self._soft_update(agent.target_critic, agent.critic, self.critic_tau) def _soft_update(self, target: nn.Module, source: nn.Module, tau) -> None: for target_param, param in zip(target.parameters(), source.parameters()): target_param.data.copy_(target_param.data * (1.0 - tau) + param.data * tau)
class DQNAgent(): """Deep Q-learning agent.""" # def __init__(self, # env, device=DEVICE, summary_writer=writer, # noqa # hyperparameters=DQN_HYPERPARAMS): # noqa rewards = [] total_reward = 0 birth_time = 0 n_iter = 0 n_games = 0 ts_frame = 0 ts = time.time() # Memory = namedtuple( # 'Memory', ['obs', 'action', 'new_obs', 'reward', 'done'], # verbose=False, rename=False) Memory = namedtuple('Memory', ['obs', 'action', 'new_obs', 'reward', 'done'], rename=False) def __init__(self, env, hyperparameters, device, summary_writer=None): """Set parameters, initialize network.""" state_space_shape = env.observation_space.shape action_space_size = env.action_space.n self.env = env self.online_network = DQN(state_space_shape, action_space_size).to(device) self.target_network = DQN(state_space_shape, action_space_size).to(device) # XXX maybe not really necesary? self.update_target_network() self.experience_replay = None self.accumulated_loss = [] self.device = device self.optimizer = optim.Adam(self.online_network.parameters(), lr=hyperparameters['learning_rate']) self.double_DQN = hyperparameters['double_DQN'] # Discount factor self.gamma = hyperparameters['gamma'] # XXX ??? self.n_multi_step = hyperparameters['n_multi_step'] self.replay_buffer = ReplayBuffer(hyperparameters['buffer_capacity'], hyperparameters['n_multi_step'], hyperparameters['gamma']) self.birth_time = time.time() self.iter_update_target = hyperparameters['n_iter_update_target'] self.buffer_start_size = hyperparameters['buffer_start_size'] self.summary_writer = summary_writer # Greedy search hyperparameters self.epsilon_start = hyperparameters['epsilon_start'] self.epsilon = hyperparameters['epsilon_start'] self.epsilon_decay = hyperparameters['epsilon_decay'] self.epsilon_final = hyperparameters['epsilon_final'] def get_max_action(self, obs): ''' Forward pass of the NN to obtain the action of the given observations ''' # convert the observation in tensor state_t = torch.tensor(np.array([obs])).to(self.device) # forward pass q_values_t = self.online_network(state_t) # get the maximum value of the output (i.e. the best action to take) _, act_t = torch.max(q_values_t, dim=1) return int(act_t.item()) def act(self, obs): ''' Greedy action outputted by the NN in the CentralControl ''' return self.get_max_action(obs) def act_eps_greedy(self, obs): ''' E-greedy action ''' # In case of a noisy net, it takes a greedy action # if self.noisy_net: # return self.act(obs) if np.random.random() < self.epsilon: return self.env.action_space.sample() else: return self.act(obs) def update_target_network(self): """Update target network weights with current online network values.""" self.target_network.load_state_dict(self.online_network.state_dict()) def set_optimizer(self, learning_rate): self.optimizer = optim.Adam(self.online_network.parameters(), lr=learning_rate) def sample_and_optimize(self, batch_size): ''' Sample batch_size memories from the buffer and optimize them ''' # This should be the part where it waits until it has enough # experience if len(self.replay_buffer) > self.buffer_start_size: # sample mini_batch = self.replay_buffer.sample(batch_size) # optimize # l_loss = self.cc.optimize(mini_batch) l_loss = self.optimize(mini_batch) self.accumulated_loss.append(l_loss) # update target NN if self.n_iter % self.iter_update_target == 0: self.update_target_network() def optimize(self, mini_batch): ''' Optimize the NN ''' # reset the grads self.optimizer.zero_grad() # caluclate the loss of the mini batch loss = self._calulate_loss(mini_batch) loss_v = loss.item() # do backpropagation loss.backward() # one step of optimization self.optimizer.step() return loss_v def _calulate_loss(self, mini_batch): ''' Calculate mini batch's MSE loss. It support also the double DQN version ''' states, actions, next_states, rewards, dones = mini_batch # convert the data in tensors states_t = torch.as_tensor(states, device=self.device) next_states_t = torch.as_tensor(next_states, device=self.device) actions_t = torch.as_tensor(actions, device=self.device) rewards_t = torch.as_tensor(rewards, dtype=torch.float32, device=self.device) done_t = torch.as_tensor(dones, dtype=torch.uint8, device=self.device) # noqa # Value of the action taken previously (recorded in actions_v) # in state_t state_action_values = self.online_network(states_t).gather( 1, actions_t[:, None]).squeeze(-1) # NB gather is a differentiable function # Next state value with Double DQN. (i.e. get the value predicted # by the target nn, of the best action predicted by the online nn) if self.double_DQN: double_max_action = self.online_network(next_states_t).max(1)[1] double_max_action = double_max_action.detach() target_output = self.target_network(next_states_t) # NB: [:,None] add an extra dimension next_state_values = torch.gather( target_output, 1, double_max_action[:, None]).squeeze(-1) # Next state value in the normal configuration else: next_state_values = self.target_network(next_states_t).max(1)[0] next_state_values = next_state_values.detach() # No backprop # Use the Bellman equation expected_state_action_values = rewards_t + \ (self.gamma**self.n_multi_step) * next_state_values # compute the loss return nn.MSELoss()(state_action_values, expected_state_action_values) def reset_stats(self): ''' Reset the agent's statistics ''' self.rewards.append(self.total_reward) self.total_reward = 0 self.accumulated_loss = [] self.n_games += 1 def add_env_feedback(self, obs, action, new_obs, reward, done): ''' Acquire a new feedback from the environment. The feedback is constituted by the new observation, the reward and the done boolean. ''' # Create the new memory and update the buffer new_memory = self.Memory(obs=obs, action=action, new_obs=new_obs, reward=reward, done=done) # Append it to the replay buffer self.replay_buffer.append(new_memory) # update the variables self.n_iter += 1 # TODO check this... # decrease epsilon self.epsilon = max( self.epsilon_final, self.epsilon_start - self.n_iter / self.epsilon_decay) self.total_reward += reward def print_info(self): ''' Print information about the agent ''' fps = (self.n_iter - self.ts_frame) / (time.time() - self.ts) # TODO replace with proper logger print('%d %d rew:%d mean_rew:%.2f eps:%.2f, fps:%d, loss:%.4f' % (self.n_iter, self.n_games, self.total_reward, np.mean(self.rewards[-40:]), self.epsilon, fps, np.mean(self.accumulated_loss))) self.ts_frame = self.n_iter self.ts = time.time() if self.summary_writer is not None: self.summary_writer.add_scalar('reward', self.total_reward, self.n_games) self.summary_writer.add_scalar('mean_reward', np.mean(self.rewards[-40:]), self.n_games) self.summary_writer.add_scalar('10_mean_reward', np.mean(self.rewards[-10:]), self.n_games) self.summary_writer.add_scalar('epsilon', self.epsilon, self.n_games) self.summary_writer.add_scalar('loss', np.mean(self.accumulated_loss), self.n_games)
class TD3: def __init__(self, env, gamma=0.99, tau=1e-3, pol_lr=1e-4, q_lr=5e-3, batch_size=64, buffer_size=10000, target_noise=0.2, action_noise=0.1, clip_range=0.5, update_delay=2): # environment stuff self.env = env self.num_act = env.action_space.shape[0] self.num_obs = env.observation_space.shape[0] self.eval_env = copy.deepcopy(env) # hyper parameters self.gamma = gamma self.tau = tau self.pol_lr = pol_lr self.q_lr = q_lr self.batch_size = batch_size self.buffer_size = buffer_size self.target_noise = target_noise self.action_noise = action_noise self.clip_range = clip_range self.update_delay = 2 # networks self.pol = Actor(self.num_obs, self.num_act, [400, 300]).double() self.q1 = Critic(self.num_obs, self.num_act, [400, 300]).double() self.q2 = Critic(self.num_obs, self.num_act, [400, 300]).double() self.pol.init_weights() self.q1.init_weights() self.q2.init_weights() self.target_pol = copy.deepcopy(self.pol).double() self.target_q1 = copy.deepcopy(self.q1).double() self.target_q2 = copy.deepcopy(self.q2).double() # optimizers, buffer self.pol_opt = torch.optim.Adam(self.pol.parameters(), lr=self.pol_lr) self.q1_opt = torch.optim.Adam( self.q1.parameters(), lr=self.q_lr, ) self.q2_opt = torch.optim.Adam( self.q2.parameters(), lr=self.q_lr, ) self.buffer = ReplayBuffer(self.buffer_size, 1000) self.mse_loss = torch.nn.MSELoss() self.cum_q1_loss = 0 self.cum_q2_loss = 0 self.cum_obj = 0 def noise(self, noise, length): return torch.tensor(np.random.multivariate_normal( mean=np.array([0.0 for i in range(length)]), cov=np.diag([noise for i in range(length)])), dtype=torch.double) # fill up buffer first def prep_buffer(self): obs = self.env.reset() while not self.buffer.ready: pre_obs = obs action = self.env.action_space.sample() obs, reward, done, _ = self.env.step(action) self.buffer.insert((pre_obs, action, reward, obs, done)) if done: obs = self.env.reset() # for clipping off values def clip(self, x, l, u): if isinstance(l, (list, np.ndarray)): lower = torch.tensor(l, dtype=torch.double) upper = torch.tensor(u, dtype=torch.double) elif isinstance(l, (int, float)): lower = torch.tensor([l for i in range(len(x))], dtype=torch.double) upper = torch.tensor([u for i in range(len(x))], dtype=torch.double) else: assert (False, "Clipped wrong") return torch.max(torch.min(x, upper), lower) # update neural net def update_networks(self): # (pre_obs, action, reward, obs, done) pre_obs = torch.tensor(self.batch[0], dtype=torch.double) actions = torch.tensor(self.batch[1], dtype=torch.double) rewards = torch.tensor(self.batch[2], dtype=torch.double) obs = torch.tensor(self.batch[3], dtype=torch.double) done = torch.tensor(self.batch[4], dtype=torch.double).unsqueeze(1) self.q1_opt.zero_grad() self.q2_opt.zero_grad() noise = self.clip( torch.tensor(self.noise(self.target_noise, self.num_act)), -self.clip_range, self.clip_range) target_action = self.clip( self.target_pol(obs) + noise, self.env.action_space.low, self.env.action_space.high) target_q1_val = self.target_q1(obs, target_action) target_q2_val = self.target_q2(obs, target_action) y = rewards + (self.gamma * (1.0 - done) * torch.min(target_q1_val, target_q2_val)) # loss = torch.sum((y - self.q(pre_obs, actions)) ** 2) / self.batch_size q1_loss = self.mse_loss(self.q1(pre_obs, actions), y) q2_loss = self.mse_loss(self.q2(pre_obs, actions), y) q1_loss.backward(retain_graph=True) q2_loss.backward() self.q1_opt.step() self.q2_opt.step() self.cum_q1_loss += q1_loss self.cum_q2_loss += q2_loss self.pol_opt.zero_grad() objective = -self.q1(pre_obs, self.pol(pre_obs)).mean() objective.backward() self.pol_opt.step() self.cum_obj += objective # update target networks with tau def update_target_networks(self): for target, actual in zip(self.target_q1.named_parameters(), self.q1.named_parameters()): target[1].data.copy_(self.tau * actual[1].data + (1 - self.tau) * target[1].data) for target, actual in zip(self.target_q2.named_parameters(), self.q2.named_parameters()): target[1].data.copy_(self.tau * actual[1].data + (1 - self.tau) * target[1].data) for target, actual in zip(self.target_pol.named_parameters(), self.pol.named_parameters()): target[1].data.copy_(self.tau * actual[1].data + (1 - self.tau) * target[1].data) def policy_eval(self): state = self.eval_env.reset() done = False rewards = [] while not done: inp = torch.tensor(state, dtype=torch.double) action = self.pol(inp) action = action.detach().numpy() next_state, r, done, _ = self.eval_env.step(action) rewards.append(r) # self.eval_env.render() # time.sleep(0.1) state = next_state total = sum(rewards) return total def train(self, num_iters=200000, eval_len=1000, render=False): print("Start") if render: self.env.render('human') self.prep_buffer() obs = self.env.reset() iter_info = [] # train for num_iters for i in range(int(num_iters / eval_len)): for j in trange(eval_len): # one step and put into buffer pre_obs = obs inp = torch.tensor(obs, dtype=torch.double) action = self.pol(inp) action = action + self.noise(self.action_noise, self.num_act) action = action.detach().numpy() obs, reward, done, _ = self.env.step(action) self.buffer.insert((pre_obs, action, reward, obs, done)) if render: self.env.render('human') time.sleep(0.000001) if done: obs = self.env.reset() # TD3 updates less often if j % self.update_delay == 0: # sample from buffer, train one step, update target networks self.batch = self.buffer.sample(self.batch_size) self.update_networks() self.update_target_networks() iter_reward = self.policy_eval() avg_q1_loss = self.cum_q1_loss / ((i + 1) * eval_len) avg_q2_loss = self.cum_q2_loss / ((i + 1) * eval_len) avg_obj = self.cum_obj / ((i + 1) * eval_len) print("Iteration {}/{}".format((i + 1) * eval_len, num_iters)) print("Rewards: {} | Q Loss: {}, {} | Policy Objective: {}".format( iter_reward, avg_q1_loss, avg_q2_loss, avg_obj)) iter_info.append((iter_reward, avg_q1_loss, avg_q2_loss, avg_obj)) return iter_info
class DQNAgent(): ''' Agent class. It control all the agent functionalities ''' rewards = [] total_reward = 0 birth_time = 0 n_iter = 0 n_games = 0 ts_frame = 0 ts = time.time() Memory = namedtuple('Memory', ['obs', 'action', 'new_obs', 'reward', 'done'], rename=False) def __init__(self, env, device, hyperparameters, summary_writer=None): ''' Agent initialization. It create the CentralControl that control all the low ''' # The CentralControl is the 'brain' of the agent self.cc = CentralControl(env.observation_space.shape, env.action_space.n, hyperparameters['gamma'], hyperparameters['n_multi_step'], hyperparameters['double_DQN'], hyperparameters['noisy_net'], hyperparameters['dueling'], device) self.cc.set_optimizer(hyperparameters['learning_rate']) self.birth_time = time.time() self.iter_update_target = hyperparameters['n_iter_update_target'] self.buffer_start_size = hyperparameters['buffer_start_size'] self.epsilon_start = hyperparameters['epsilon_start'] self.epsilon = hyperparameters['epsilon_start'] self.epsilon_decay = hyperparameters['epsilon_decay'] self.epsilon_final = hyperparameters['epsilon_final'] self.accumulated_loss = [] self.device = device # initialize the replay buffer (i.e. the memory) of the agent self.replay_buffer = ReplayBuffer(hyperparameters['buffer_capacity'], hyperparameters['n_multi_step'], hyperparameters['gamma']) self.summary_writer = summary_writer self.noisy_net = hyperparameters['noisy_net'] self.env = env def act(self, obs): ''' Greedy action outputted by the NN in the CentralControl ''' return self.cc.get_max_action(obs) def act_eps_greedy(self, obs): ''' E-greedy action ''' # In case of a noisy net, it takes a greedy action if self.noisy_net: return self.act(obs) if np.random.random() < self.epsilon: return self.env.action_space.sample() else: return self.act(obs) def add_env_feedback(self, obs, action, new_obs, reward, done): ''' Acquire a new feedback from the environment. The feedback is constituted by the new observation, the reward and the done boolean. ''' # Create the new memory and update the buffer new_memory = self.Memory(obs=obs, action=action, new_obs=new_obs, reward=reward, done=done) self.replay_buffer.append(new_memory) # update the variables self.n_iter += 1 # decrease epsilon self.epsilon = max( self.epsilon_final, self.epsilon_start - self.n_iter / self.epsilon_decay) self.total_reward += reward def sample_and_optimize(self, batch_size): ''' Sample batch_size memories from the buffer and optimize them ''' if len(self.replay_buffer) > self.buffer_start_size: # sample mini_batch = self.replay_buffer.sample(batch_size) # optimize l_loss = self.cc.optimize(mini_batch) self.accumulated_loss.append(l_loss) # update target NN if self.n_iter % self.iter_update_target == 0: self.cc.update_target() def reset_stats(self): ''' Reset the agent's statistics ''' self.rewards.append(self.total_reward) self.total_reward = 0 self.accumulated_loss = [] self.n_games += 1 def print_info(self): ''' Print information about the agent ''' fps = (self.n_iter - self.ts_frame) / (time.time() - self.ts) print('%d %d rew:%d mean_rew:%.2f eps:%.2f, fps:%d, loss:%.4f' % (self.n_iter, self.n_games, self.total_reward, np.mean(self.rewards[-40:]), self.epsilon, fps, np.mean(self.accumulated_loss))) self.ts_frame = self.n_iter self.ts = time.time() if self.summary_writer != None: self.summary_writer.add_scalar('reward', self.total_reward, self.n_games) self.summary_writer.add_scalar('mean_reward', np.mean(self.rewards[-40:]), self.n_games) self.summary_writer.add_scalar('10_mean_reward', np.mean(self.rewards[-10:]), self.n_games) self.summary_writer.add_scalar('esilon', self.epsilon, self.n_games) self.summary_writer.add_scalar('loss', np.mean(self.accumulated_loss), self.n_games)
class D4PG_Agent: """ PyTorch Implementation of D4PG: "Distributed Distributional Deterministic Policy Gradients" (Barth-Maron, Hoffman, et al., 2018) As described in the paper at: https://arxiv.org/pdf/1804.08617.pdf Much thanks also to the original DDPG paper: "Continuous Control with Deep Reinforcement Learning" (Lillicrap, Hunt, et al., 2016) https://arxiv.org/pdf/1509.02971.pdf And to: "A Distributional Perspective on Reinforcement Learning" (Bellemare, Dabney, et al., 2017) https://arxiv.org/pdf/1707.06887.pdf D4PG utilizes distributional value estimation, n-step returns, prioritized experience replay (PER), distributed K-actor exploration, and off-policy actor-critic learning to achieve very fast and stable learning for continuous control tasks. This version of the Agent is written to interact with Udacity's Continuous Control robotic arm manipulation environment which provides 20 simultaneous actors, negating the need for K-actor implementation. Thus, this code has no multiprocessing functionality. It could be easily added as part of the main.py script. In the original D4PG paper, it is suggested in the data that PER does not have significant (or perhaps any at all) effect on the speed or stability of learning. Thus, it too has been left out of this implementation but may be added as a future TODO item. """ def __init__(self, env, args, e_decay=1, e_min=0.05, l2_decay=0.0001, update_type="hard"): """ Initialize a D4PG Agent. """ self.device = args.device self.framework = "D4PG" self.eval = args.eval self.agent_count = env.agent_count self.actor_learn_rate = args.actor_learn_rate self.critic_learn_rate = args.critic_learn_rate self.batch_size = args.batch_size self.buffer_size = args.buffer_size self.action_size = env.action_size self.state_size = env.state_size self.C = args.C self._e = args.e self.e_decay = e_decay self.e_min = e_min self.gamma = args.gamma self.rollout = args.rollout self.tau = args.tau self.update_type = update_type self.num_atoms = args.num_atoms self.vmin = args.vmin self.vmax = args.vmax self.atoms = torch.linspace(self.vmin, self.vmax, self.num_atoms).to(self.device) self.t_step = 0 self.episode = 0 # Set up memory buffers, currently only standard replay is implemented # self.memory = ReplayBuffer(self.device, self.buffer_size, self.gamma, self.rollout) # Initialize ACTOR networks # self.actor = ActorNet(args.layer_sizes, self.state_size, self.action_size).to(self.device) self.actor_target = ActorNet(args.layer_sizes, self.state_size, self.action_size).to(self.device) self._hard_update(self.actor, self.actor_target) self.actor_optim = optim.Adam(self.actor.parameters(), lr=self.actor_learn_rate, weight_decay=l2_decay) # Initialize CRITIC networks # self.critic = CriticNet(args.layer_sizes, self.state_size, self.action_size, self.num_atoms).to(self.device) self.critic_target = CriticNet(args.layer_sizes, self.state_size, self.action_size, self.num_atoms).to(self.device) self._hard_update(self.actor, self.actor_target) self.critic_optim = optim.Adam(self.critic.parameters(), lr=self.critic_learn_rate, weight_decay=l2_decay) self.new_episode() def act(self, states, eval=False): """ Predict an action using a policy/ACTOR network π. Scaled noise N (gaussian distribution) is added to all actions todo encourage exploration. """ states = states.to(self.device) with torch.no_grad(): actions = self.actor(states).detach().cpu().numpy() if not eval: noise = self._gauss_noise(actions.shape) actions += noise return np.clip(actions, -1, 1) def step(self, states, actions, rewards, next_states, pretrain=False): """ Add the current SARS' tuple into the short term memory, then learn """ # Current SARS' stored in short term memory, then stacked for NStep experience = list(zip(states, actions, rewards, next_states)) self.memory.store_experience(experience) self.t_step += 1 # Learn after done pretraining if not pretrain: self.learn() def learn(self): """ Performs a distributional Actor/Critic calculation and update. Actor πθ and πθ' Critic Zw and Zw' (categorical distribution) """ # Sample from replay buffer, REWARDS are sum of ROLLOUT timesteps # Already calculated before storing in the replay buffer. # NEXT_STATES are ROLLOUT steps ahead of STATES batch = self.memory.sample(self.batch_size) states, actions, rewards, next_states = batch atoms = self.atoms.unsqueeze(0) # Calculate Yᵢ from target networks using πθ' and Zw' # These tensors are not needed for backpropogation, so detach from the # calculation graph (literally doubles runtime if this is not detached) target_dist = self._get_targets(rewards, next_states).detach() # Calculate log probability DISTRIBUTION using Zw w.r.t. stored actions log_probs = self.critic(states, actions, log=True) # Calculate the critic network LOSS (Cross Entropy), CE-loss is ideal # for categorical value distributions as utilized in D4PG. # estimates distance between target and projected values critic_loss = -(target_dist * log_probs).sum(-1).mean() # Predict action for actor network loss calculation using πθ predicted_action = self.actor(states) # Predict value DISTRIBUTION using Zw w.r.t. action predicted by πθ probs = self.critic(states, predicted_action) # Multiply probabilities by atom values and sum across columns to get # Q-Value expected_reward = (probs * atoms).sum(-1) # Calculate the actor network LOSS (Policy Gradient) # Take the mean across the batch and multiply in the negative to # perform gradient ascent actor_loss = -expected_reward.mean() # Perform gradient ascent self.actor_optim.zero_grad() actor_loss.backward() self.actor_optim.step() # Perform gradient descent self.critic_optim.zero_grad() critic_loss.backward() self.critic_optim.step() self._update_networks() self.actor_loss = actor_loss.item() self.critic_loss = critic_loss.item() def initialize_memory(self, pretrain_length, env): """ Fills up the ReplayBuffer memory with PRETRAIN_LENGTH number of experiences before training begins. """ if len(self.memory) >= pretrain_length: print("Memory already filled, length: {}".format(len(self.memory))) return print("Initializing memory buffer.") states = env.states while len(self.memory) < pretrain_length: actions = np.random.uniform(-1, 1, (self.agent_count, self.action_size)) next_states, rewards, dones = env.step(actions) self.step(states, actions, rewards, next_states, pretrain=True) if self.t_step % 10 == 0 or len(self.memory) >= pretrain_length: print("Taking pretrain step {}... memory filled: {}/{}\ ".format(self.t_step, len(self.memory), pretrain_length)) states = next_states print("Done!") self.t_step = 0 def _get_targets(self, rewards, next_states): """ Calculate Yᵢ from target networks using πθ' and Zw' """ target_actions = self.actor_target(next_states) target_probs = self.critic_target(next_states, target_actions) # Project the categorical distribution onto the supports projected_probs = self._categorical(rewards, target_probs) return projected_probs def _categorical(self, rewards, probs): """ Returns the projected value distribution for the input state/action pair While there are several very similar implementations of this Categorical Projection methodology around github, this function owes the most inspiration to Zhang Shangtong and his excellent repository located at: https://github.com/ShangtongZhang """ # Create local vars to keep code more concise vmin = self.vmin vmax = self.vmax atoms = self.atoms num_atoms = self.num_atoms gamma = self.gamma rollout = self.rollout rewards = rewards.unsqueeze(-1) delta_z = (vmax - vmin) / (num_atoms - 1) # Rewards were stored with 0->(N-1) summed, take Reward and add it to # the discounted expected reward at N (ROLLOUT) timesteps projected_atoms = rewards + gamma**rollout * atoms.unsqueeze(0) projected_atoms.clamp_(vmin, vmax) b = (projected_atoms - vmin) / delta_z # It seems that on professional level GPUs (for instance on AWS), the # floating point math is accurate to the degree that a tensor printing # as 99.00000 might in fact be 99.000000001 in the backend, perhaps due # to binary imprecision, but resulting in 99.00000...ceil() evaluating # to 100 instead of 99. Forcibly reducing the precision to the minimum # seems to be the only solution to this problem, and presents no issues # to the accuracy of calculating lower/upper_bound correctly. precision = 1 b = torch.round(b * 10**precision) / 10**precision lower_bound = b.floor() upper_bound = b.ceil() m_lower = (upper_bound + (lower_bound == upper_bound).float() - b) * probs m_upper = (b - lower_bound) * probs projected_probs = torch.tensor(np.zeros(probs.size())).to(self.device) for idx in range(probs.size(0)): projected_probs[idx].index_add_(0, lower_bound[idx].long(), m_lower[idx].double()) projected_probs[idx].index_add_(0, upper_bound[idx].long(), m_upper[idx].double()) return projected_probs.float() @property def e(self): """ This property ensures that the annealing process is run every time that E is called. Anneals the epsilon rate down to a specified minimum to ensure there is always some noisiness to the policy actions. Returns as a property. """ self._e = max(self.e_min, self._e * self.e_decay) return self._e def _gauss_noise(self, shape): """ Returns the epsilon scaled noise distribution for adding to Actor calculated action policy. """ n = np.random.normal(0, 1, shape) return self.e * n def new_episode(self): """ Handle any cleanup or steps to begin a new episode of training. """ self.memory.init_n_step() self.episode += 1 def _update_networks(self): """ Updates the network using either DDPG-style soft updates (w/ param \ TAU), or using a DQN/D4PG style hard update every C timesteps. """ if self.update_type == "soft": self._soft_update(self.actor, self.actor_target) self._soft_update(self.critic, self.critic_target) elif self.t_step % self.C == 0: self._hard_update(self.actor, self.actor_target) self._hard_update(self.critic, self.critic_target) def _soft_update(self, active, target): """ Slowly updated the network using every-step partial network copies modulated by parameter TAU. """ for t_param, param in zip(target.parameters(), active.parameters()): t_param.data.copy_(self.tau * param.data + (1 - self.tau) * t_param.data) def _hard_update(self, active, target): """ Fully copy parameters from active network to target network. To be used in conjunction with a parameter "C" that modulated how many timesteps between these hard updates. """ target.load_state_dict(active.state_dict())
class ACDQNGM(DQNG): def __init__(self, args, env, env_test, logger): super(ACDQNGM, self).__init__(args, env, env_test, logger) def init(self, args, env): names = ['s0', 'a', 's1', 'r', 't', 'g', 'm', 'task', 'mcr'] metrics = ['loss_dqn', 'qval', 'val'] self.buffer = ReplayBuffer(limit=int(1e6), names=names.copy(), args=args) self.actorCritic = ActorCriticDQNGM(args, env) for metric in metrics: self.metrics[metric] = 0 self.goalcounts = np.zeros((len(self.env.goals), )) def train(self): if self.buffer.nb_entries > 100 * self.batch_size: samples = self.buffer.sample(self.batch_size) samples = self.env.augment_samples(samples) targets = self.actorCritic.get_targets_dqn(samples['r'], samples['t'], samples['s1'], samples['g'], samples['m']) inputs = [ samples['s0'], samples['a'], samples['g'], samples['m'], targets ] metricsCritic = self.actorCritic.trainCritic(inputs) self.metrics['loss_dqn'] += np.squeeze(metricsCritic[0]) self.metrics['qval'] += np.mean(metricsCritic[1]) self.goalcounts += np.bincount(samples['task'], minlength=len(self.env.goals)) metricsActor = self.actorCritic.trainActor( [samples['s0'], samples['g'], samples['m']]) if self.env_step % 1000 == 0: print(metricsActor[0], metricsActor[1]) self.metrics['val'] += np.mean(metricsActor[2]) self.actorCritic.target_train() def get_stats(self): sumsamples = np.sum(self.goalcounts) if sumsamples != 0: for i, goal in enumerate(self.env.goals): self.stats['samplecount_{}'.format(goal)] = float( "{0:.3f}".format(self.goalcounts[i] / sumsamples)) def make_input(self, state, mode): if mode == 'train': input = [ np.expand_dims(i, axis=0) for i in [state, self.env.goal, self.env.mask] ] else: input = [ np.expand_dims(i, axis=0) for i in [state, self.env_test.goal, self.env_test.mask] ] return input def act(self, exp, mode='train'): input = self.make_input(exp['s0'], mode) actionProbs = self.actorCritic.probs(input)[0].squeeze() # if self.env_step % 1000 == 0: print(actionProbs) if mode == 'train': action = np.random.choice(range(self.env.action_dim), p=actionProbs) else: action = np.argmax(actionProbs[0]) prob = actionProbs[action] action = np.expand_dims(action, axis=1) exp['a'] = action # exp['p_a'] = prob return exp def reset(self): if self.trajectory: augmented_episode = self.env.end_episode(self.trajectory) for expe in augmented_episode: self.buffer.append(expe) # for expe in self.trajectory: # self.buffer.append(expe.copy()) # augmented_ep = self.env.augment_episode(self.trajectory) # for e in augmented_ep: # self.buffer.append(e) self.trajectory.clear() state = self.env.reset() self.episode_step = 0 return state def get_demo(self, rndprop): demo = [] exp = {} exp['s0'] = self.env_test.env.reset() # obj = np.random.choice(self.env_test.env.objects) # goal = np.random.randint(obj.high[2]+1) obj = self.env_test.env.light goal = 1 while True: if np.random.rand() < rndprop: a = np.random.randint(self.env_test.action_dim) done = False else: a, done = self.env_test.env.opt_action(obj, goal) if not done: exp['a'] = np.expand_dims(a, axis=1) exp['s1'] = self.env_test.env.step(exp['a'])[0] demo.append(exp.copy()) exp['s0'] = exp['s1'] else: break return demo def demo(self): if self.env_step % self.demo_freq == 0: for i in range(5): demo = self.get_demo(rndprop=0.) augmented_demo = self.env.augment_demo(demo) for exp in augmented_demo: self.buffer.append(exp)
class DQN_Agent: """ PyTorch Implementation of DQN/DDQN. """ def __init__(self, state_size, action_size, args): """ Initialize a D4PG Agent. """ self.device = torch.device( "cuda:0" if torch.cuda.is_available() else "cpu") self.action_size = action_size self.state_size = state_size self.framework = args.framework self.eval = args.eval self.agent_count = 1 self.learn_rate = args.learn_rate self.batch_size = args.batch_size self.buffer_size = args.buffer_size self.C = args.C self._epsilon = args.epsilon self.epsilon_decay = args.epsilon_decay self.epsilon_min = args.epsilon_min self.gamma = 0.99 self.rollout = args.rollout self.tau = args.tau self.momentum = 1 self.l2_decay = 0.0001 self.update_type = "hard" self.t_step = 0 self.episode = 0 self.seed = 0 # Set up memory buffers if args.prioritized_experience_replay: self.memory = PERBuffer(args.buffersize, self.batchsize, self.framestack, self.device, args.alpha, args.beta) self.criterion = WeightedLoss() else: self.memory = ReplayBuffer(self.device, self.buffer_size, self.gamma, self.rollout) # Initialize Q networks # self.q = self._make_model(state_size, action_size, args.pixels) self.q_target = self._make_model(state_size, action_size, args.pixels) self._hard_update(self.q, self.q_target) self.q_optimizer = self._set_optimizer(self.q.parameters(), lr=self.learn_rate, decay=self.l2_decay, momentum=self.momentum) self.new_episode() @property def epsilon(self): """ This property ensures that the annealing process is run every time that E is called. Anneals the epsilon rate down to a specified minimum to ensure there is always some noisiness to the policy actions. Returns as a property. """ self._epsilon = max(self.epsilon_min, self.epsilon_decay**self.t_step) return self._epsilon def act(self, state, eval=False, pretrain=False): """ Select an action using epsilon-greedy π. Always use greedy if not training. """ if np.random.random() > self.epsilon or not eval and not pretrain: state = state.to(self.device) with torch.no_grad(): action_values = self.q(state).detach().cpu() action = action_values.argmax(dim=1).unsqueeze(0).numpy() else: action = np.random.randint(self.action_size, size=(1, 1)) return action.astype(np.long) def step(self, state, action, reward, next_state, pretrain=False): """ Add the current SARS' tuple into the short term memory, then learn """ # Current SARS' stored in short term memory, then stacked for NStep experience = (state, action, reward, next_state) if self.rollout == 1: self.memory.store_trajectory(state, torch.from_numpy(action), torch.tensor([reward]), next_state) else: self.memory.store_experience(experience) self.t_step += 1 # Learn after done pretraining if not pretrain: self.learn() def learn(self): """ Trains the Deep QNetwork and returns action values. Can use multiple frameworks. """ # Sample from replay buffer, REWARDS are sum of (ROLLOUT - 1) timesteps # Already calculated before storing in the replay buffer. # NEXT_STATES are ROLLOUT steps ahead of STATES batch, is_weights, tree_idx = self.memory.sample(self.batch_size) states, actions, rewards, next_states, terminal_mask = batch q_values = torch.zeros(self.batch_size).to(self.device) if self.framework == 'DQN': # Max predicted Q values for the next states from the target model q_values[terminal_mask] = self.q_target(next_states).detach().max( dim=1)[0] if self.framework == 'DDQN': # Get maximizing ACTION under Q, evaluate actionvalue # under q_target # Max valued action under active network max_actions = self.q(next_states).detach().argmax(1).unsqueeze(1) # Use the active network action to get the value of the stable # target network q_values[terminal_mask] = self.q_target( next_states).detach().gather(1, max_actions).squeeze(1) targets = rewards + (self.gamma**self.rollout * q_values) targets = targets.unsqueeze(1) values = self.q(states).gather(1, actions) #Huber Loss provides better results than MSE if is_weights is None: loss = F.smooth_l1_loss(values, targets) #Compute Huber Loss manually to utilize is_weights with Prioritization else: loss, td_errors = self.criterion.huber(values, targets, is_weights) self.memory.batch_update(tree_idx, td_errors) # Perform gradient descent self.q_optimizer.zero_grad() loss.backward() self.q_optimizer.step() self._update_networks() self.loss = loss.item() def initialize_memory(self, pretrain_length, env): """ Fills up the ReplayBuffer memory with PRETRAIN_LENGTH number of experiences before training begins. """ if len(self.memory) >= pretrain_length: print("Memory already filled, length: {}".format(len(self.memory))) return print("Initializing memory buffer.") while True: done = False env.reset() state = env.state while not done: action = self.act(state, pretrain=True) next_state, reward, done = env.step(action) if done: next_state = None self.step(state, action, reward, next_state, pretrain=True) states = next_state if self.t_step % 50 == 0 or len( self.memory) >= pretrain_length: print("Taking pretrain step {}... memory filled: {}/{}\ ".format(self.t_step, len(self.memory), pretrain_length)) if len(self.memory) >= pretrain_length: print("Done!") self.t_step = 0 self._epsilon = 1 return def new_episode(self): """ Handle any cleanup or steps to begin a new episode of training. """ self.memory.init_n_step() self.episode += 1 def _update_networks(self): """ Updates the network using either DDPG-style soft updates (w/ param \ TAU), or using a DQN/D4PG style hard update every C timesteps. """ if self.update_type == "soft": self._soft_update(self.q, self.q_target) elif self.t_step % self.C == 0: self._hard_update(self.q, self.q_target) def _soft_update(self, active, target): """ Slowly updated the network using every-step partial network copies modulated by parameter TAU. """ for t_param, param in zip(target.parameters(), active.parameters()): t_param.data.copy_(self.tau * param.data + (1 - self.tau) * t_param.data) def _hard_update(self, active, target): """ Fully copy parameters from active network to target network. To be used in conjunction with a parameter "C" that modulated how many timesteps between these hard updates. """ target.load_state_dict(active.state_dict()) def _set_optimizer(self, params, lr, decay, momentum, optimizer="Adam"): """ Sets the optimizer based on command line choice. Defaults to Adam. """ if optimizer == "RMSprop": return optim.RMSprop(params, lr=lr, momentum=momentum) elif optimizer == "SGD": return optim.SGD(params, lr=lr, momentum=momentum) else: return optim.Adam(params, lr=lr, weight_decay=decay) def _make_model(self, state_size, action_size, use_cnn): """ Sets up the network model based on whether state data or pixel data is provided. """ if use_cnn: return QCNNetwork(state_size, action_size, self.seed).to(self.device) else: return QNetwork(state_size, action_size, self.seed).to(self.device)
class DDPG: def __init__( self, env, gamma=0.99, tau=1e-3, pol_lr=1e-4, q_lr=1e-3, batch_size=64, buffer_size=10000, ): # environment stuff self.env = env self.num_act = env.action_space.shape[0] self.num_obs = env.observation_space.shape[0] self.eval_env = copy.deepcopy(env) # hyper parameters self.gamma = gamma self.tau = tau self.pol_lr = pol_lr self.q_lr = q_lr self.batch_size = batch_size self.buffer_size = buffer_size # networks self.pol = Actor(self.num_obs, self.num_act, [400, 300]).double() self.q = Critic(self.num_obs, self.num_act, [400, 300]).double() self.pol.init_weights() self.q.init_weights() self.target_pol = copy.deepcopy(self.pol).double() self.target_q = copy.deepcopy(self.q).double() # optimizers, buffer self.pol_opt = torch.optim.Adam(self.pol.parameters(), lr=self.pol_lr) self.q_opt = torch.optim.Adam( self.q.parameters(), lr=self.q_lr, ) # weight_decay=1e-2) self.buffer = ReplayBuffer(self.buffer_size, 1000) self.mse_loss = torch.nn.MSELoss() self.cum_loss = 0 self.cum_obj = 0 # fill up buffer first def prep_buffer(self): obs = self.env.reset() while not self.buffer.ready: pre_obs = obs action = self.env.action_space.sample() obs, reward, done, _ = self.env.step(action) self.buffer.insert((pre_obs, action, reward, obs, done)) if done: obs = self.env.reset() # update neural net def update_networks(self): # (pre_obs, action, reward, obs, done) pre_obs = torch.tensor(self.batch[0], dtype=torch.double) actions = torch.tensor(self.batch[1], dtype=torch.double) rewards = torch.tensor(self.batch[2], dtype=torch.double) obs = torch.tensor(self.batch[3], dtype=torch.double) done = torch.tensor(self.batch[4], dtype=torch.double).unsqueeze(1) self.q_opt.zero_grad() y = rewards + (self.gamma * (1.0 - done) * self.target_q(obs, self.target_pol(obs))) # loss = torch.sum((y - self.q(pre_obs, actions)) ** 2) / self.batch_size loss = self.mse_loss(self.q(pre_obs, actions), y) loss.backward() self.q_opt.step() self.cum_loss += loss self.pol_opt.zero_grad() objective = -self.q(pre_obs, self.pol(pre_obs)).mean() objective.backward() self.pol_opt.step() self.cum_obj += objective # update target networks with tau def update_target_networks(self): for target, actual in zip(self.target_q.named_parameters(), self.q.named_parameters()): target[1].data.copy_(self.tau * actual[1].data + (1 - self.tau) * target[1].data) for target, actual in zip(self.target_pol.named_parameters(), self.pol.named_parameters()): target[1].data.copy_(self.tau * actual[1].data + (1 - self.tau) * target[1].data) def policy_eval(self): state = self.eval_env.reset() done = False rewards = [] while not done: inp = torch.tensor(state, dtype=torch.double) action = self.pol(inp) action = action.detach().numpy() next_state, r, done, _ = self.eval_env.step(action) rewards.append(r) # self.eval_env.render() # time.sleep(0.1) state = next_state total = sum(rewards) return total def train(self, num_iters=200000, eval_len=1000, render=False): print("Start") if render: self.env.render('human') self.prep_buffer() obs = self.env.reset() iter_info = [] # train for num_iters for i in range(int(num_iters / eval_len)): for j in trange(eval_len): # one step and put into buffer pre_obs = obs inp = torch.tensor(obs, dtype=torch.double) action = self.pol(inp) action = action.detach().numpy( ) + np.random.multivariate_normal(mean=np.array([0.0, 0.0]), cov=np.array([[0.1, 0.0], [0.0, 0.1]])) obs, reward, done, _ = self.env.step(action) self.buffer.insert((pre_obs, action, reward, obs, done)) if render: self.env.render('human') time.sleep(0.000001) if done: obs = self.env.reset() # sample from buffer, train one step, update target networks self.batch = self.buffer.sample(self.batch_size) self.update_networks() self.update_target_networks() iter_reward = self.policy_eval() avg_loss = self.cum_loss / ((i + 1) * eval_len) avg_obj = self.cum_obj / ((i + 1) * eval_len) print("Iteration {}/{}".format((i + 1) * eval_len, num_iters)) print("Rewards: {} | Q Loss: {} | Policy Objective: {}".format( iter_reward, avg_loss, avg_obj)) iter_info.append((iter_reward, avg_loss, avg_obj)) return iter_info
class Agent(): """Interacts with and learns from the environment.""" def __init__(self, state_size, action_size, seed, framework, buffer_type): """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) self.framework = framework self.buffer_type = buffer_type # 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 # def __init__(self, device, buffer_size, batch_size, alpha, beta): if self.buffer_type == 'PER_ReplayBuffer': self.memory = PER_ReplayBuffer(device, BUFFER_SIZE, BATCH_SIZE, ALPHA, BETA) if self.buffer_type == 'ReplayBuffer': self.memory = ReplayBuffer(device, 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: if self.buffer_type == 'ReplayBuffer': experiences = self.memory.sample() is_weights = None idxs = None if self.buffer_type == 'PER_ReplayBuffer': experiences, is_weights, idxs = self.memory.sample() self.criterion = WeightedLoss() #print('debugging:', experiences ) self.learn(experiences, is_weights, idxs, 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() # use local network # 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, is_weights, idxs, 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 if self.framework == 'DQN': # 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 ) #target network: which is uploaded slower than the local network #print('Q_targets_next is ', Q_targets_next) if self.framework == "DDQN": #print("DDQN") max_actions = self.qnetwork_local(next_states).detach().argmax( 1).unsqueeze(1) #print('max_actions is ', max_actions) #print('self.qnetwork_target(next_states).detach() is ',self.qnetwork_target(next_states).detach()) #Q_targets_next = self.qnetwork_target(next_states).detach().gather(1, max_actions).squeeze(1) Q_targets_next = self.qnetwork_target(next_states).detach().gather( 1, max_actions) #print('DDQN, Q', Q_targets_next) #print('rewards is ', rewards) # 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) #------------------------------------------------------------------------------ #Huber Loss provides better results than MSE if is_weights is None: loss = F.smooth_l1_loss(Q_expected, Q_targets) #Compute Huber Loss manually to utilize is_weights with Prioritization else: loss, td_errors = self.criterion.huber(Q_expected, Q_targets, is_weights) self.memory.batch_update(idxs, td_errors) # Perform gradient descent 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)
def learn(self, timesteps=10000, verbose=0, seed=None): if seed is not None: random.seed(seed) np.random.seed(seed) torch.manual_seed(seed) self.eps_range = self._eps_range(timesteps) replay_buffer = ReplayBuffer(self.buffer_size) self._init_model() obs = self.env.reset() for step in range(timesteps): # while not done: cur_eps = next(self.eps_range, None) if cur_eps is None: cur_eps = self.final_eps action = self._select_action(obs, cur_eps) new_obs, rewards, done, info = self.env.step(action) if done: new_obs = [ np.nan ] * self.obs_shape[0] # hacky way to keep dimensions correct replay_buffer.add(obs, action, rewards, new_obs) obs = new_obs # learn gradient if step > self.learning_starts: if len(replay_buffer.buffer ) < self.batch_size: # buffer too small continue samples = replay_buffer.sample(self.batch_size, self.device) obs_batch, actions_batch, rewards_batch, new_obs_batch = samples predicted_q_values = self._predictQValue( self.step_model, obs_batch, actions_batch) ys = self._expectedLabels(self.target_model, new_obs_batch, rewards_batch) loss = F.smooth_l1_loss(predicted_q_values, ys) self.optim.zero_grad() loss.backward() for i in self.step_model.parameters(): i.grad.clamp_(min=-1, max=1) # exploding gradient # i.grad.clamp_(min=-10, max=10) # exploding gradient self.optim.step() # update target if step % self.target_network_update_freq == 0: self.target_model.load_state_dict( self.step_model.state_dict()) if done: obs = self.env.reset() if verbose == 1: if step % (timesteps * 0.1) == 0: perc = int(step / (timesteps * 0.1)) print(f"At step {step}") print(f"{perc}% done")
class DDPG(Agent): def __init__(self, args, env, env_test, logger): super(DDPG, self).__init__(args, env, env_test, logger) self.args = args self.init(args, env) def init(self, args, env): names = ['s0', 'a', 's1', 'r', 't'] metrics = ['loss_dqn', 'loss_actor'] self.buffer = ReplayBuffer(limit=int(1e6), names=names.copy(), args=args) self.actorCritic = ActorCriticDDPG(args, env) for metric in metrics: self.metrics[metric] = 0 def train(self): if self.buffer.nb_entries > self.batch_size: exp = self.buffer.sample(self.batch_size) targets_dqn = self.actorCritic.get_targets_dqn( exp['r'], exp['t'], exp['s1']) inputs = [exp['s0'], exp['a'], targets_dqn] loss_dqn = self.actorCritic.trainQval(inputs) action, criticActionGrads, invertedCriticActionGrads = self.actorCritic.trainActor( [exp['s0']]) self.metrics['loss_dqn'] += np.squeeze(loss_dqn) # a2 = self.actor.model.predict_on_batch(s0) # grads = self.critic.gradsModel.predict_on_batch([s0, a2]) # low = self.env.action_space.low # high = self.env.action_space.high # for d in range(grads[0].shape[0]): # width = high[d] - low[d] # for k in range(self.batch_size): # if grads[k][d] >= 0: # grads[k][d] *= (high[d] - a2[k][d]) / width # else: # grads[k][d] *= (a2[k][d] - low[d]) / width # self.actor.train(s0, grads) self.actorCritic.target_train() def make_input(self, state, mode): input = [np.expand_dims(state, axis=0)] return input def reset(self): if self.trajectory: self.env.end_episode(self.trajectory) for expe in self.trajectory: self.buffer.append(expe.copy()) self.trajectory.clear() state = self.env.reset() self.episode_step = 0 return state def act(self, state, mode='train'): input = self.make_input(state, mode) action = self.actorCritic.action(input)[0] if mode == 'train': noise = np.random.normal(0., 0.1, size=action[0].shape) action = noise + action action = np.clip(action, self.env.action_space.low, self.env.action_space.high) action = action.squeeze() return action def save_model(self): self.actorCritic.actionModel.save(os.path.join(self.logger.get_dir(), 'actor_model'), overwrite=True) self.actorCritic.qvalModel.save(os.path.join(self.logger.get_dir(), 'qval_model'), overwrite=True)
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.): """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 # print("Rewards") # print(rewards.shape) ## TODO: compute and minimize the loss "*** YOUR CODE HERE ***" # self.qnetwork_local.train() local_output = self.qnetwork_local(states).gather(1, actions) selected_target_actions = torch.max(self.qnetwork_target(next_states).detach(),1)[0].unsqueeze(1) activated = torch.sub(torch.Tensor(np.ones(dones.shape)), dones) is_it_done = torch.mul(selected_target_actions, activated) target_output = torch.add(torch.mul(is_it_done, gamma), rewards) loss = F.mse_loss(local_output, target_output) 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, s_dim, num_actions, lr): self.step = 0 self.epStep = 0 self.ep = 0 self.tutorListened = True self.tutorInput = '' self.sDim = s_dim self.num_actions = num_actions self.learning_rate = lr self.names = ['state0', 'action', 'feedback', 'fWeight'] self.buffer = ReplayBuffer(limit=int(1e6), names=self.names) self.batchSize = 64 self.episode = deque(maxlen=400) self.model = self.create_model() def create_model(self): state = Input(shape=self.sDim) action = Input(shape=(1,), dtype='uint8') l1 = Dense(400, activation="relu")(state) feedback = Dense(self.num_actions, activation=None, kernel_initializer='random_uniform')(l1) feedback = Reshape((1, self.num_actions))(feedback) mask = Lambda(K.one_hot, arguments={'num_classes': self.num_actions}, output_shape=(self.num_actions,))(action) feedback = multiply([feedback, mask]) feedback = Lambda(K.sum, arguments={'axis': 2})(feedback) feedbackModel = Model(inputs=[state, action], outputs=feedback) feedbackModel.compile(loss='mse', optimizer=Adam(lr=self.learning_rate)) return feedbackModel def train(self): loss = 0 if self.buffer.nb_entries > self.batchSize: samples = self.buffer.sample(self.batchSize) s, a, targets, weights = [np.array(samples[name]) for name in self.names] loss = self.model.train_on_batch(x=[s,a], y=targets, sample_weight=weights) return loss def tutorListener(self): self.tutorInput = input("> ") print("maybe updating...the kbdInput variable is: {}".format(self.tutorInput)) self.tutorListened = True def run(self): state0 = np.random.randint(0, 4, size=(5,)) while self.step < 100000: if self.tutorInput != '': print("Received new keyboard Input. Setting playing ID to keyboard input value") for i in range(1,10): self.episode[-i]['fWeight'] = 1 self.episode[-i]['feedback'] = self.tutorInput self.tutorInput = '' else: action = np.random.randint(self.num_actions) state1 = np.random.randint(0, 4, size=(5,)) self.step += 1 self.epStep += 1 experience = {'state0': state0, 'action': action, 'fWeight': 0} self.episode.append(experience) self.loss = self.train() state0 = state1 time.sleep(0.001) if self.tutorListened: self.tutorListened = False self.listener = Thread(target=self.tutorListener) self.listener.start() if self.epStep >= 200: if self.ep > 0: for s in range(self.epStep): exp = self.episode.popleft() if exp['fWeight'] != 0: self.buffer.append(exp) self.epStep = 0 self.ep += 1 state0 = np.random.randint(0, 4, size=(5,)) if self.step % 1000 == 0: print(self.step, self.loss) def input(self): while True: if input() == '+': inputStep = self.step time.sleep(2) print('input +1, step = ', inputStep) elif input() == '-': inputStep = self.step time.sleep(2) print('input -1, step = ', inputStep) else: print('wrong input')
class DQN(Agent): def __init__(self, args, env, env_test, logger): super(DQN, self).__init__(args, env, env_test, logger) self.args = args self.init(args, env) def init(self, args, env): names = ['state0', 'action', 'state1', 'reward', 'terminal'] self.buffer = ReplayBuffer(limit=int(1e6), names=names.copy()) self.critic = CriticDQN(args, env) for metric_name in ['loss_dqn', 'qval', 'val']: self.metrics[metric_name] = 0 def train(self): if self.buffer.nb_entries > self.batch_size: exp = self.buffer.sample(self.batch_size) s0, a0, s1, r, t = [exp[name] for name in self.buffer.names] targets_dqn = self.critic.get_targets_dqn(r, t, s1) inputs = [s0, a0] loss = self.critic.criticModel.train_on_batch(inputs, targets_dqn) for i, metric in enumerate(self.critic.criticModel.metrics_names): self.metrics[metric] += loss[i] self.critic.target_train() def reset(self): if self.trajectory: R = np.sum([ self.env.unshape(exp['reward'], exp['terminal']) for exp in self.trajectory ]) self.env.processEp(R) for expe in reversed(self.trajectory): self.buffer.append(expe.copy()) if self.args['--imit'] != '0': Es = [0] for i, expe in enumerate(reversed(self.trajectory)): if self.trajectory[-1]['terminal']: Es[0] = Es[0] * self.critic.gamma + expe['reward'] expe['expVal'] = Es[0] else: expe['expVal'] = -self.ep_steps self.bufferImit.append(expe.copy()) self.trajectory.clear() state = self.env.reset() self.episode_step = 0 return state def make_input(self, state, mode): input = [np.reshape(state, (1, self.critic.s_dim[0]))] input.append(np.expand_dims([0.5], axis=0)) return input def act(self, state, mode='train'): input = self.make_input(state, mode) actionProbs = self.critic.actionProbs(input) if mode == 'train': action = np.random.choice(range(self.env.action_dim), p=actionProbs[0].squeeze()) else: action = np.argmax(actionProbs[0]) return np.expand_dims(action, axis=1)