def evaluate(weights, args, NUM_EVALS=10, render=False): """Rollout Worker runs a simulation in the environment to generate experiences and fitness values Parameters: args (object): Parameter class id (int): Specific Id unique to each worker spun task_pipe (pipe): Receiver end of the task pipe used to receive signal to start on a task result_pipe (pipe): Sender end of the pipe used to report back results data_bucket (shared list object): A shared list object managed by a manager that is used to store experience tuples models_bucket (shared list object): A shared list object managed by a manager used to store all the models (actors) store_transition (bool): Log experiences to exp_list? random_baseline (bool): Test with random action for baseline purpose? Returns: None """ env = RoverDomainPython(args, NUM_EVALS) model = MultiHeadActor(args.state_dim, args.action_dim, args.hidden_size, args.config.num_agents) for i, param in enumerate(model.parameters()): try: param.data = weights[i] except: param.data = weights[i].data fitness = [None for _ in range(NUM_EVALS)]; frame=0 joint_state = env.reset() joint_state = utils.to_tensor(np.array(joint_state)) while True: #unless done joint_action = [model.clean_action(joint_state[i, :], head=i).detach().numpy() for i in range(args.config.num_agents)] #JOINT ACTION [agent_id, universe_id, action] #Bound Action joint_action = np.array(joint_action).clip(-1.0, 1.0) next_state, reward, done, global_reward = env.step(joint_action) # Simulate one step in environment #State --> [agent_id, universe_id, obs] #reward --> [agent_id, universe_id] #done --> [universe_id] #info --> [universe_id] next_state = utils.to_tensor(np.array(next_state)) #Grab global reward as fitnesses for i, grew in enumerate(global_reward): if grew != None: fitness[i] = grew joint_state = next_state frame+=NUM_EVALS #DONE FLAG IS Received if sum(done)==len(done): break return sum(fitness)/len(fitness)
class MATD3(object):
"""Classes implementing TD3 and DDPG off-policy learners
Parameters:
args (object): Parameter class
"""
def __init__(self, id, algo_name, state_dim, action_dim, hidden_size, actor_lr, critic_lr, gamma, tau, savetag, foldername, actualize, use_gpu, num_agents, init_w = True):
self.algo_name = algo_name; self.gamma = gamma; self.tau = tau; self.total_update = 0; self.agent_id = id;self.use_gpu = use_gpu
self.tracker = utils.Tracker(foldername, ['q_'+savetag, 'qloss_'+savetag, 'policy_loss_'+savetag], '.csv', save_iteration=1000, conv_size=1000)
self.num_agents = num_agents
#Initialize actors
self.policy = MultiHeadActor(state_dim, action_dim, hidden_size, num_agents)
if init_w: self.policy.apply(utils.init_weights)
self.policy_target = MultiHeadActor(state_dim, action_dim, hidden_size, num_agents)
utils.hard_update(self.policy_target, self.policy)
self.policy_optim = Adam(self.policy.parameters(), actor_lr)
self.critics = [QNetwork(state_dim*num_agents, action_dim*num_agents, hidden_size*2) for _ in range(num_agents)]
self.critics_target = [QNetwork(state_dim*num_agents, action_dim*num_agents, hidden_size*2) for _ in range(num_agents)]
if init_w:
for critic, critic_target in zip(self.critics, self.critics_target):
critic.apply(utils.init_weights)
utils.hard_update(critic_target, critic)
self.critic_optims = [Adam(critic.parameters(), critic_lr) for critic in self.critics]
self.loss = nn.MSELoss()
if use_gpu:
self.policy_target.cuda(); self.policy.cuda()
for critic, critic_target in zip(self.critics, self.critics_target):
critic.cuda()
critic_target.cuda()
self.num_critic_updates = 0
#Statistics Tracker
#self.action_loss = {'min':None, 'max': None, 'mean':None, 'std':None}
self.policy_loss = {'min':None, 'max': None, 'mean':None, 'std':None}
self.q_loss = {'min':None, 'max': None, 'mean':None, 'std':None}
self.q = {'min':None, 'max': None, 'mean':None, 'std':None}
def update_parameters(self, state_batch, next_state_batch, action_batch, reward_batch, done_batch, agent_id, num_epoch=1, **kwargs):
"""Runs a step of Bellman upodate and policy gradient using a batch of experiences
Parameters:
state_batch (tensor): Current States
next_state_batch (tensor): Next States
action_batch (tensor): Actions
reward_batch (tensor): Rewards
done_batch (tensor): Done batch
num_epoch (int): Number of learning iteration to run with the same data
Returns:
None
"""
if isinstance(state_batch, list): state_batch = torch.cat(state_batch); next_state_batch = torch.cat(next_state_batch); action_batch = torch.cat(action_batch); reward_batch = torch.cat(reward_batch). done_batch = torch.cat(done_batch)
batch_size = len(state_batch)
for _ in range(num_epoch):
########### CRITIC UPDATE ####################
#Compute next q-val, next_v and target
with torch.no_grad():
#Compute next action_bacth
next_action_batch = torch.cat([self.policy_target.clean_action(next_state_batch[:, id, :], id) for id in range(self.num_agents)], 1)
if self.algo_name == 'TD3':
# Policy Noise
policy_noise = np.random.normal(0, kwargs['policy_noise'], (action_batch.size()[0], action_batch.size()[1] * action_batch.size()[2]))
policy_noise = torch.clamp(torch.Tensor(policy_noise), -kwargs['policy_noise_clip'], kwargs['policy_noise_clip'])
next_action_batch += policy_noise.cuda() if self.use_gpu else policy_noise
next_action_batch = torch.clamp(next_action_batch, -1, 1)
#Compute Q-val and value of next state masking by done
q1, q2 = self.critics_target[agent_id].forward(next_state_batch.view(batch_size, -1), next_action_batch)
q1 = (1 - done_batch) * q1
q2 = (1 - done_batch) * q2
#next_val = (1 - done_batch) * next_val
#Select which q to use as next-q (depends on algo)
if self.algo_name == 'TD3':next_q = torch.min(q1, q2)
elif self.algo_name == 'DDPG': next_q = q1
#Compute target q and target val
target_q = reward_batch[:,agent_id].unsqueeze(1) + (self.gamma * next_q)
#if self.args.use_advantage: target_val = reward_batch + (self.gamma * next_val)
self.critic_optims[agent_id].zero_grad()
current_q1, current_q2 = self.critics[agent_id].forward((state_batch.view(batch_size, -1)), (action_batch.view(batch_size, -1)))
utils.compute_stats(current_q1, self.q)
dt = self.loss(current_q1, target_q)
# if self.args.use_advantage:
# dt = dt + self.loss(current_val, target_val)
# utils.compute_stats(current_val, self.val)
if self.algo_name == 'TD3': dt = dt + self.loss(current_q2, target_q)
utils.compute_stats(dt, self.q_loss)
# if self.args.critic_constraint:
# if dt.item() > self.args.critic_constraint_w:
# dt = dt * (abs(self.args.critic_constraint_w / dt.item()))
dt.backward()
self.critic_optims[agent_id].step()
self.num_critic_updates += 1
#Delayed Actor Update
if self.num_critic_updates % kwargs['policy_ups_freq'] == 0 or self.algo_name == 'DDPG':
agent_action = self.policy.clean_action(state_batch[:,agent_id,:], agent_id)
joint_action = action_batch.clone()
joint_action[:,agent_id,:] = agent_action[:]
#print(np.max(torch.abs(joint_action - action_batch).detach().cpu().numpy()), np.max(torch.abs(joint_action[:,agent_id,:] - agent_action).detach().cpu().numpy()))
# # Trust Region constraint
# if self.args.trust_region_actor:
# with torch.no_grad(): old_actor_actions = self.actor_target.forward(state_batch)
# actor_actions = action_batch - old_actor_actions
Q1, Q2 = self.critics[agent_id].forward(state_batch.view(batch_size, -1), joint_action.view(batch_size, -1))
# if self.args.use_advantage: policy_loss = -(Q1 - val)
policy_loss = -Q1
utils.compute_stats(-policy_loss,self.policy_loss)
policy_loss = policy_loss.mean()
self.policy_optim.zero_grad()
policy_loss.backward(retain_graph=True)
#nn.utils.clip_grad_norm_(self.actor.parameters(), 10)
# if self.args.action_loss:
# action_loss = torch.abs(actor_actions-0.5)
# utils.compute_stats(action_loss, self.action_loss)
# action_loss = action_loss.mean() * self.args.action_loss_w
# action_loss.backward()
# #if self.action_loss[-1] > self.policy_loss[-1]: self.args.action_loss_w *= 0.9 #Decay action_w loss if action loss is larger than policy gradient loss
self.policy_optim.step()
# if self.args.hard_update:
# if self.num_critic_updates % self.args.hard_update_freq == 0:
# if self.num_critic_updates % self.args.policy_ups_freq == 0: self.hard_update(self.actor_target, self.actor)
# self.hard_update(self.critic_target, self.critic)
if self.num_critic_updates % kwargs['policy_ups_freq'] == 0 or self.algo_name == 'DDPG': utils.soft_update(self.policy_target, self.policy, self.tau)
for critic, critic_target in zip(self.critics, self.critics_target):
utils.soft_update(critic_target, critic, self.tau)
self.total_update += 1
if self.agent_id == 0:
self.tracker.update([self.q['mean'], self.q_loss['mean'], self.policy_loss['mean']] ,self.total_update)
class MultiTD3(object):
"""Classes implementing TD3 and DDPG off-policy learners
Parameters:
args (object): Parameter class
"""
def __init__(self, id, algo_name, state_dim, action_dim, hidden_size, actor_lr, critic_lr, gamma, tau, savetag, foldername, actualize, use_gpu, num_agents, init_w = True):
self.algo_name = algo_name; self.gamma = gamma; self.tau = tau; self.total_update = 0; self.agent_id = id; self.actualize = actualize; self.use_gpu = use_gpu
self.tracker = utils.Tracker(foldername, ['q_'+savetag, 'qloss_'+savetag, 'policy_loss_'+savetag, 'alz_score'+savetag,'alz_policy'+savetag], '.csv', save_iteration=1000, conv_size=1000)
#Initialize actors
self.policy = MultiHeadActor(state_dim, action_dim, hidden_size, num_agents)
if init_w: self.policy.apply(utils.init_weights)
self.policy_target = MultiHeadActor(state_dim, action_dim, hidden_size, num_agents)
utils.hard_update(self.policy_target, self.policy)
self.policy_optim = Adam(self.policy.parameters(), actor_lr)
self.critic = QNetwork(state_dim, action_dim,hidden_size)
if init_w: self.critic.apply(utils.init_weights)
self.critic_target = QNetwork(state_dim, action_dim, hidden_size)
utils.hard_update(self.critic_target, self.critic)
self.critic_optim = Adam(self.critic.parameters(), critic_lr)
if actualize:
self.ANetwork = ActualizationNetwork(state_dim, action_dim, hidden_size)
if init_w: self.ANetwork.apply(utils.init_weights)
self.actualize_optim = Adam(self.ANetwork.parameters(), critic_lr)
self.actualize_lr = 0.2
if use_gpu: self.ANetwork.cuda()
self.loss = nn.MSELoss()
if use_gpu:
self.policy_target.cuda(); self.critic_target.cuda(); self.policy.cuda(); self.critic.cuda()
self.num_critic_updates = 0
#Statistics Tracker
#self.action_loss = {'min':None, 'max': None, 'mean':None, 'std':None}
self.policy_loss = {'min':None, 'max': None, 'mean':None, 'std':None}
self.q_loss = {'min':None, 'max': None, 'mean':None, 'std':None}
self.q = {'min':None, 'max': None, 'mean':None, 'std':None}
self.alz_score = {'min':None, 'max': None, 'mean':None, 'std':None}
self.alz_policy = {'min':None, 'max': None, 'mean':None, 'std':None}
#self.val = {'min':None, 'max': None, 'mean':None, 'std':None}
#self.value_loss = {'min':None, 'max': None, 'mean':None, 'std':None}
def update_parameters(self, state_batch, next_state_batch, action_batch, reward_batch, done_batch, global_reward, agent_id, num_epoch=1, **kwargs):
"""Runs a step of Bellman upodate and policy gradient using a batch of experiences
Parameters:
state_batch (tensor): Current States
next_state_batch (tensor): Next States
action_batch (tensor): Actions
reward_batch (tensor): Rewards
done_batch (tensor): Done batch
num_epoch (int): Number of learning iteration to run with the same data
Returns:
None
"""
if isinstance(state_batch, list): state_batch = torch.cat(state_batch); next_state_batch = torch.cat(next_state_batch); action_batch = torch.cat(action_batch); reward_batch = torch.cat(reward_batch). done_batch = torch.cat(done_batch); global_reward = torch.cat(global_reward)
for _ in range(num_epoch):
########### CRITIC UPDATE ####################
#Compute next q-val, next_v and target
with torch.no_grad():
#Policy Noise
policy_noise = np.random.normal(0, kwargs['policy_noise'], (action_batch.size()[0], action_batch.size()[1]))
policy_noise = torch.clamp(torch.Tensor(policy_noise), -kwargs['policy_noise_clip'], kwargs['policy_noise_clip'])
#Compute next action_bacth
next_action_batch = self.policy_target.clean_action(next_state_batch, agent_id) + policy_noise.cuda() if self.use_gpu else policy_noise
next_action_batch = torch.clamp(next_action_batch, -1, 1)
#Compute Q-val and value of next state masking by done
q1, q2 = self.critic_target.forward(next_state_batch, next_action_batch)
q1 = (1 - done_batch) * q1
q2 = (1 - done_batch) * q2
#next_val = (1 - done_batch) * next_val
#Select which q to use as next-q (depends on algo)
if self.algo_name == 'TD3' or self.algo_name == 'TD3_actor_min': next_q = torch.min(q1, q2)
elif self.algo_name == 'DDPG': next_q = q1
elif self.algo_name == 'TD3_max': next_q = torch.max(q1, q2)
#Compute target q and target val
target_q = reward_batch + (self.gamma * next_q)
#if self.args.use_advantage: target_val = reward_batch + (self.gamma * next_val)
if self.actualize:
##########Actualization Network Update
current_Ascore = self.ANetwork.forward(state_batch, action_batch)
utils.compute_stats(current_Ascore, self.alz_score)
target_Ascore = (self.actualize_lr) * (global_reward * 10.0) + (1 - self.actualize_lr) * current_Ascore.detach()
actualize_loss = self.loss(target_Ascore, current_Ascore).mean()
self.critic_optim.zero_grad()
current_q1, current_q2 = self.critic.forward((state_batch), (action_batch))
utils.compute_stats(current_q1, self.q)
dt = self.loss(current_q1, target_q)
# if self.args.use_advantage:
# dt = dt + self.loss(current_val, target_val)
# utils.compute_stats(current_val, self.val)
if self.algo_name == 'TD3' or self.algo_name == 'TD3_max': dt = dt + self.loss(current_q2, target_q)
utils.compute_stats(dt, self.q_loss)
# if self.args.critic_constraint:
# if dt.item() > self.args.critic_constraint_w:
# dt = dt * (abs(self.args.critic_constraint_w / dt.item()))
dt.backward()
self.critic_optim.step()
self.num_critic_updates += 1
if self.actualize:
self.actualize_optim.zero_grad()
actualize_loss.backward()
self.actualize_optim.step()
#Delayed Actor Update
if self.num_critic_updates % kwargs['policy_ups_freq'] == 0:
actor_actions = self.policy.clean_action(state_batch, agent_id)
# # Trust Region constraint
# if self.args.trust_region_actor:
# with torch.no_grad(): old_actor_actions = self.actor_target.forward(state_batch)
# actor_actions = action_batch - old_actor_actions
Q1, Q2 = self.critic.forward(state_batch, actor_actions)
# if self.args.use_advantage: policy_loss = -(Q1 - val)
policy_loss = -Q1
utils.compute_stats(-policy_loss,self.policy_loss)
policy_loss = policy_loss.mean()
###Actualzie Policy Update
if self.actualize:
A1 = self.ANetwork.forward(state_batch, actor_actions)
utils.compute_stats(A1, self.alz_policy)
policy_loss += -A1.mean()
self.policy_optim.zero_grad()
policy_loss.backward(retain_graph=True)
#nn.utils.clip_grad_norm_(self.actor.parameters(), 10)
# if self.args.action_loss:
# action_loss = torch.abs(actor_actions-0.5)
# utils.compute_stats(action_loss, self.action_loss)
# action_loss = action_loss.mean() * self.args.action_loss_w
# action_loss.backward()
# #if self.action_loss[-1] > self.policy_loss[-1]: self.args.action_loss_w *= 0.9 #Decay action_w loss if action loss is larger than policy gradient loss
self.policy_optim.step()
# if self.args.hard_update:
# if self.num_critic_updates % self.args.hard_update_freq == 0:
# if self.num_critic_updates % self.args.policy_ups_freq == 0: self.hard_update(self.actor_target, self.actor)
# self.hard_update(self.critic_target, self.critic)
if self.num_critic_updates % kwargs['policy_ups_freq'] == 0: utils.soft_update(self.policy_target, self.policy, self.tau)
utils.soft_update(self.critic_target, self.critic, self.tau)
self.total_update += 1
if self.agent_id == 0:
self.tracker.update([self.q['mean'], self.q_loss['mean'], self.policy_loss['mean'],self.alz_score['mean'], self.alz_policy['mean']] ,self.total_update)
class MultiTD3(object):
"""Classes implementing TD3 and DDPG off-policy learners
"""
def __init__(self,
id,
algo_name,
state_dim,
action_dim,
hidden_size,
actor_lr,
critic_lr,
gamma,
tau,
savetag,
foldername,
use_gpu,
num_agents,
init_w=True):
self.algo_name = algo_name
self.gamma = gamma
self.tau = tau
self.total_update = 0
self.agent_id = id
self.use_gpu = use_gpu
self.tracker = utils.Tracker(
foldername,
['q_' + savetag, 'qloss_' + savetag, 'policy_loss_' + savetag],
'.csv',
save_iteration=1000,
conv_size=1000)
#Initialize actors
self.policy = MultiHeadActor(state_dim, action_dim, hidden_size,
num_agents)
if init_w: self.policy.apply(utils.init_weights)
self.policy_target = MultiHeadActor(state_dim, action_dim, hidden_size,
num_agents)
utils.hard_update(self.policy_target, self.policy)
self.policy_optim = Adam(self.policy.parameters(), actor_lr)
self.critic = QNetwork(state_dim, action_dim, hidden_size)
if init_w: self.critic.apply(utils.init_weights)
self.critic_target = QNetwork(state_dim, action_dim, hidden_size)
utils.hard_update(self.critic_target, self.critic)
self.critic_optim = Adam(self.critic.parameters(), critic_lr)
self.loss = nn.MSELoss()
if use_gpu:
self.policy_target.cuda()
self.critic_target.cuda()
self.policy.cuda()
self.critic.cuda()
self.num_critic_updates = 0
#Statistics Tracker
self.policy_loss = {
'min': None,
'max': None,
'mean': None,
'std': None
}
self.q_loss = {'min': None, 'max': None, 'mean': None, 'std': None}
self.q = {'min': None, 'max': None, 'mean': None, 'std': None}
def update_parameters(self,
state_batch,
next_state_batch,
action_batch,
reward_batch,
done_batch,
global_reward,
agent_id,
num_epoch=1,
**kwargs):
"""Runs a step of Bellman upodate and policy gradient using a batch of experiences
"""
if isinstance(state_batch, list):
state_batch = torch.cat(state_batch)
next_state_batch = torch.cat(next_state_batch)
action_batch = torch.cat(action_batch)
reward_batch = torch.cat(reward_batch).done_batch = torch.cat(
done_batch)
global_reward = torch.cat(global_reward)
for _ in range(num_epoch):
########### CRITIC UPDATE ####################
#Compute next q-val, next_v and target
with torch.no_grad():
#Policy Noise
policy_noise = np.random.normal(
0, kwargs['policy_noise'],
(action_batch.size()[0], action_batch.size()[1]))
policy_noise = torch.clamp(torch.Tensor(policy_noise),
-kwargs['policy_noise_clip'],
kwargs['policy_noise_clip'])
#Compute next action_bacth
next_action_batch = self.policy_target.clean_action(
next_state_batch, agent_id) + policy_noise.cuda(
) if self.use_gpu else policy_noise
next_action_batch = torch.clamp(next_action_batch, -1, 1)
#Compute Q-val and value of next state masking by done
q1, q2 = self.critic_target.forward(next_state_batch,
next_action_batch)
q1 = (1 - done_batch) * q1
q2 = (1 - done_batch) * q2
#Select which q to use as next-q (depends on algo)
if self.algo_name == 'TD3': next_q = torch.min(q1, q2)
elif self.algo_name == 'DDPG': next_q = q1
#Compute target q and target val
target_q = reward_batch + (self.gamma * next_q)
self.critic_optim.zero_grad()
current_q1, current_q2 = self.critic.forward((state_batch),
(action_batch))
utils.compute_stats(current_q1, self.q)
dt = self.loss(current_q1, target_q)
if self.algo_name == 'TD3':
dt = dt + self.loss(current_q2, target_q)
utils.compute_stats(dt, self.q_loss)
dt.backward()
self.critic_optim.step()
self.num_critic_updates += 1
#Delayed Actor Update
if self.num_critic_updates % kwargs['policy_ups_freq'] == 0:
actor_actions = self.policy.clean_action(state_batch, agent_id)
Q1, Q2 = self.critic.forward(state_batch, actor_actions)
# if self.args.use_advantage: policy_loss = -(Q1 - val)
policy_loss = -Q1
utils.compute_stats(-policy_loss, self.policy_loss)
policy_loss = policy_loss.mean()
self.policy_optim.zero_grad()
policy_loss.backward(retain_graph=True)
self.policy_optim.step()
if self.num_critic_updates % kwargs['policy_ups_freq'] == 0:
utils.soft_update(self.policy_target, self.policy, self.tau)
utils.soft_update(self.critic_target, self.critic, self.tau)
self.total_update += 1
if self.agent_id == 0:
self.tracker.update([
self.q['mean'], self.q_loss['mean'],
self.policy_loss['mean']
], self.total_update)
if __name__ == "__main__": args = Parameters() # Create the Parameters class test_tracker = utils.Tracker(args.metric_save, [args.log_fname], '.csv') # Initiate tracker torch.manual_seed(args.seed); np.random.seed(args.seed); random.seed(args.seed) # Seeds model = MultiHeadActor(args.state_dim, args.action_dim, args.hidden_size, args.config.num_agents) #model.cuda() # EvolutionModule runs the population in a ThreadPool, so # if you need to inject other arguments, you can do that # using the partial tool partial_func = partial(evaluate, args=args) mother_parameters = list(model.parameters()) es = EvolutionModule( mother_parameters, partial_func, population_size=args.popsize, sigma=args.sigma, learning_rate=args.lr, threadcount=100, cuda=False, render_test=False ) ###### TRAINING LOOP ######## total_frames = 0 while True: #Train es.run(iterations=2, print_step = 10) total_frames += 10*args.popsize*args.config.ep_len*2