def __init__(self, nb_state, nb_action):
self.nb_state = nb_state
self.nb_action = nb_action
self.actor = Actor(self.nb_state, self.nb_action)
self.actor_target = Actor(self.nb_state, self.nb_action)
self.actor_optim = Adam(self.actor.parameters(), lr=LEARNING_RATE)
self.critic = Critic(self.nb_state, self.nb_action)
self.critic_target = Critic(self.nb_state, self.nb_action)
self.critic_optim = Adam(self.critic.parameters(), lr=LEARNING_RATE)
hard_update(self.actor_target,
self.actor) # Make sure target is with the same weight
hard_update(self.critic_target, self.critic)
#Create replay buffer
self.memory = SequentialMemory(limit=MEMORY_SIZE, window_length=1)
self.random_process = OrnsteinUhlenbeckProcess(size=nb_action,
theta=OU_THETA,
mu=OU_MU,
sigma=OU_SIGMA)
self.is_training = True
self.epsilon = 1.0
self.a_t = None
self.s_t = None
if USE_CUDA: self.cuda()
def __init__(self, nb_states, nb_actions):
self.critic = Critic(nb_states, nb_actions) # Q
self.critic_target = Critic(nb_states, nb_actions)
self.actor = Actor(nb_states, nb_actions) # policy mu
self.actor_target = Actor(nb_states, nb_actions)
hard_update(self.critic_target, self.critic)
hard_update(self.actor_target, self.actor)
self.critic_optimizer = torch.optim.Adam(self.critic.parameters(),
lr=0.001)
self.actor_optimizer = torch.optim.Adam(self.actor.parameters(),
lr=0.0001)
self.criterion = nn.MSELoss()
self.random_process = OrnsteinUhlenbeckProcess(size=nb_actions,
theta=0.15,
mu=0,
sigma=0.2)
self.gamma = 0.99
self.batch_size = 64
if USE_CUDA:
self.actor.cuda()
self.actor_target.cuda()
self.critic.cuda()
self.critic_target.cuda()
def __init__(self,
state_size,
action_size,
memory_size,
batch_size=128,
tan=0.001,
actor_lr=0.001,
critic_lr=0.001,
epsilon=1.):
self.state_size = state_size
self.action_size = action_size
self.batch_size = batch_size
self.tan = tan
self.warmup = WARM_UP
self.epsilon = epsilon
self.epsilon_decay = hyperparameters['D_EPSILON']
self.actor = Actor(state_size, action_size)
self.actor_target = Actor(state_size, action_size)
self.actor_optimizer = optim.Adam(self.actor.parameters(), lr=actor_lr)
self.critic = Critic(state_size, action_size)
self.critic_target = Critic(state_size, action_size)
self.critic_optimizer = optim.Adam(self.critic.parameters(),
lr=critic_lr)
self.memory = Memory(memory_size)
self.criterion = nn.MSELoss()
self.random_process = OrnsteinUhlenbeckProcess(size=action_size,
theta=0.15,
mu=0.,
sigma=0.2)
copy_parameter(self.actor, self.actor_target)
copy_parameter(self.critic, self.critic_target)
def __init__(self, s_dim, a_dim, num_agent, **kwargs):
self.s_dim = s_dim
self.a_dim = a_dim
self.config = kwargs['config']
self.num_agent = num_agent
self.actor = Actor(s_dim, a_dim)
self.actor_target = Actor(s_dim, a_dim)
self.critic = Critic(s_dim, a_dim, num_agent)
self.critic_target = Critic(s_dim, a_dim, num_agent)
self.actor_optimizer = torch.optim.Adam(self.actor.parameters(),
lr=self.config.a_lr)
self.critic_optimizer = torch.optim.Adam(self.critic.parameters(),
lr=self.config.c_lr)
self.a_loss = 0
self.c_loss = 0
if self.config.use_cuda:
self.actor.cuda()
self.actor_target.cuda()
self.critic.cuda()
self.critic_target.cuda()
hard_update(self.actor, self.actor_target)
hard_update(self.critic, self.critic_target)
self.random_process = OrnsteinUhlenbeckProcess(
size=self.a_dim,
theta=self.config.ou_theta,
mu=self.config.ou_mu,
sigma=self.config.ou_sigma)
def __init__(self, s_dim, a_dim, **kwargs):
self.s_dim = s_dim
self.a_dim = a_dim
self.config = kwargs['config']
self.device = 'cuda' if self.config.use_cuda else 'cpu'
self.actor = Actor(s_dim, a_dim)
self.actor_target = Actor(s_dim, a_dim)
self.critic = Critic(s_dim, a_dim, 1)
self.critic_target = Critic(s_dim, a_dim, 1)
self.actor_optimizer = torch.optim.Adam(self.actor.parameters(),
lr=self.config.a_lr)
self.critic_optimizer = torch.optim.Adam(self.critic.parameters(),
lr=self.config.c_lr)
self.c_loss = 0
self.a_loss = 0
if self.config.use_cuda:
self.actor.cuda()
self.actor_target.cuda()
self.critic.cuda()
self.critic_target.cuda()
hard_update(self.actor, self.actor_target)
hard_update(self.critic, self.critic_target)
self.random_process = OrnsteinUhlenbeckProcess(
size=self.a_dim,
theta=self.config.ou_theta,
mu=self.config.ou_mu,
sigma=self.config.ou_sigma)
self.replay_buffer = list()
self.epsilon = 1.
self.depsilon = self.epsilon / self.config.epsilon_decay
def __init__(self, conf, device):
self.conf = conf
self.state_dim = conf['state_dim']
self.action_dim = conf['action_dim']
self.device = device
# create actor and critic network
self.actor = Actor_RDPG(self.state_dim,
self.action_dim).to(self.device)
self.actor_target = Actor_RDPG(self.state_dim,
self.action_dim).to(self.device)
self.critic = Critic_RDPG(self.state_dim,
self.action_dim).to(self.device)
self.critic_target = Critic_RDPG(self.state_dim,
self.action_dim).to(self.device)
hard_update(self.actor_target,
self.actor) # Make sure target is with the same weight
hard_update(self.critic_target, self.critic)
self.critic_optim = optim.Adam(self.critic.parameters(), lr=q_lr)
self.actor_optim = optim.Adam(self.actor.parameters(), lr=policy_lr)
#Create replay buffer
self.random_process = OrnsteinUhlenbeckProcess(size=self.action_dim,
theta=0.15,
mu=0.0,
sigma=0.2)
# args.ou_theta:0.15 (noise theta), args.ou_sigma:0.2 (noise sigma), args.out_mu:0.0 (noise mu)
self.epsilon = 1.0
self.depsilon = 1.0 / 50000
self.is_training = True
self.tau = 0.001 # moving average for target network
def __init__(self, args, nb_states, nb_actions):
USE_CUDA = torch.cuda.is_available()
if args.seed > 0:
self.seed(args.seed)
self.nb_states = nb_states
self.nb_actions = nb_actions
self.gpu_ids = [i for i in range(args.gpu_nums)
] if USE_CUDA and args.gpu_nums > 0 else [-1]
self.gpu_used = True if self.gpu_ids[0] >= 0 else False
net_cfg = {
'hidden1': args.hidden1,
'hidden2': args.hidden2,
'init_w': args.init_w
}
self.actor = Actor(self.nb_states, self.nb_actions, **net_cfg).double()
self.actor_target = Actor(self.nb_states, self.nb_actions,
**net_cfg).double()
self.actor_optim = Adam(self.actor.parameters(),
lr=args.p_lr,
weight_decay=args.weight_decay)
self.critic = Critic(self.nb_states, self.nb_actions,
**net_cfg).double()
self.critic_target = Critic(self.nb_states, self.nb_actions,
**net_cfg).double()
self.critic_optim = Adam(self.critic.parameters(),
lr=args.c_lr,
weight_decay=args.weight_decay)
hard_update(self.actor_target,
self.actor) # Make sure target is with the same weight
hard_update(self.critic_target, self.critic)
#Create replay buffer
self.memory = SequentialMemory(limit=args.rmsize,
window_length=args.window_length)
self.random_process = OrnsteinUhlenbeckProcess(size=self.nb_actions,
theta=args.ou_theta,
mu=args.ou_mu,
sigma=args.ou_sigma)
# Hyper-parameters
self.batch_size = args.bsize
self.tau_update = args.tau_update
self.gamma = args.gamma
# Linear decay rate of exploration policy
self.depsilon = 1.0 / args.epsilon
# initial exploration rate
self.epsilon = 1.0
self.s_t = None # Most recent state
self.a_t = None # Most recent action
self.is_training = True
self.continious_action_space = False
def __init__(self, nb_states, nb_actions, args):
if args.seed > 0:
self.seed(args.seed)
self.nb_states = nb_states
self.nb_actions = nb_actions
actor_net_cfg = {
'hidden1': 32,
'hidden2': 32,
'hidden3': 32,
'init_w': args.init_w
}
critic_net_cfg = {
'hidden1': 64,
'hidden2': 64,
'hidden3': 64,
'init_w': args.init_w
}
self.actor = Actor(self.nb_states, self.nb_actions, **actor_net_cfg)
self.actor_target = Actor(self.nb_states, self.nb_actions,
**actor_net_cfg)
self.actor_optim = Adam(self.actor.parameters(), lr=args.prate)
self.critic = Critic(self.nb_states, self.nb_actions, **critic_net_cfg)
self.critic_target = Critic(self.nb_states, self.nb_actions,
**critic_net_cfg)
self.critic_optim = Adam(self.critic.parameters(), lr=args.rate)
hard_update(self.actor_target,
self.actor) # Make sure target is with the same weight
hard_update(self.critic_target, self.critic)
#Create replay buffer
self.memory = SequentialMemory(limit=args.rmsize,
window_length=args.window_length)
self.random_process = OrnsteinUhlenbeckProcess(size=nb_actions,
theta=args.ou_theta,
mu=args.ou_mu,
sigma=args.ou_sigma)
# Hyper-parameters
self.batch_size = args.bsize
self.tau = args.tau
self.discount = args.discount
self.depsilon = 1.0 / args.epsilon
self.epsilon = 1.0
self.s_t = None # Most recent state
self.a_t = None # Most recent action
self.is_training = True
self.best_reward = -10
def __init__(self,
env,
mem_size=7 * int(1e3),
lr_critic=1e-3,
lr_actor=1e-4,
epsilon=1.,
max_epi=1500,
epsilon_decay=1. / (1e5),
gamma=.99,
target_update_frequency=200,
batch_size=64,
random_process=True,
max_step=None):
self.CUDA = torch.cuda.is_available()
self.orig_env = env #for recording
if max_step is not None:
self.orig_env._max_episode_steps = max_step
self.env = self.orig_env
self.N_S = self.env.observation_space.shape[0]
self.N_A = self.env.action_space.shape[0]
self.MAX_EPI = max_epi
self.LOW = self.env.action_space.low
self.HIGH = self.env.action_space.high
self.actor = Actor(self.N_S, self.N_A)
self.critic = Critic(self.N_S, self.N_A)
self.target_actor = Actor(self.N_S, self.N_A)
self.target_critic = Critic(self.N_S, self.N_A)
self.target_actor.eval()
self.target_critic.eval()
self.target_actor.load_state_dict(self.actor.state_dict())
self.target_critic.load_state_dict(self.critic.state_dict())
if self.CUDA:
self.actor.cuda()
self.critic.cuda()
self.target_actor.cuda()
self.target_critic.cuda()
self.exp = Experience(mem_size)
self.optim_critic = optim.Adam(self.critic.parameters(), lr=lr_critic)
self.optim_actor = optim.Adam(self.actor.parameters(), lr=-lr_actor)
self.random_process = OrnsteinUhlenbeckProcess(\
size=self.N_A, theta=.15, mu=0, sigma=.2)
self.EPSILON = epsilon
self.EPSILON_DECAY = epsilon_decay
self.GAMMA = gamma
self.TARGET_UPDATE_FREQUENCY = target_update_frequency
self.BATCH_SIZE = batch_size
title = {common.S_EPI: [], common.S_TOTAL_R: []}
self.data = pd.DataFrame(title)
self.RAND_PROC = random_process
class Predator:
def __init__(self, s_dim, a_dim, num_agent, **kwargs):
self.s_dim = s_dim
self.a_dim = a_dim
self.config = kwargs['config']
self.num_agent = num_agent
self.actor = Actor(s_dim, a_dim)
self.actor_target = Actor(s_dim, a_dim)
self.critic = Critic(s_dim, a_dim, num_agent)
self.critic_target = Critic(s_dim, a_dim, num_agent)
self.actor_optimizer = torch.optim.Adam(self.actor.parameters(),
lr=self.config.a_lr)
self.critic_optimizer = torch.optim.Adam(self.critic.parameters(),
lr=self.config.c_lr)
self.a_loss = 0
self.c_loss = 0
if self.config.use_cuda:
self.actor.cuda()
self.actor_target.cuda()
self.critic.cuda()
self.critic_target.cuda()
hard_update(self.actor, self.actor_target)
hard_update(self.critic, self.critic_target)
self.random_process = OrnsteinUhlenbeckProcess(
size=self.a_dim,
theta=self.config.ou_theta,
mu=self.config.ou_mu,
sigma=self.config.ou_sigma)
def get_batches(self):
experiences = random.sample(self.replay_buffer, self.batch_size)
state_batches = np.array([_[0] for _ in experiences])
action_batches = np.array([_[1] for _ in experiences])
reward_batches = np.array([_[2] for _ in experiences])
next_state_batches = np.array([_[3] for _ in experiences])
done_batches = np.array([_[4] for _ in experiences])
return state_batches, action_batches, reward_batches, next_state_batches, done_batches
def random_action(self):
action = np.random.uniform(low=-1.,
high=1.,
size=(self.num_agent, self.a_dim))
return action
def reset(self):
self.random_process.reset_states()
def __init__(self, env, args): #(self, nb_states, nb_actions, args):
if args.seed > 0:
self.seed(args.seed)
self.env = env
self.nb_states = self.env.observation_space.shape[0]
self.nb_actions = self.env.action_space.shape[0]
# Create Actor and Critic Network
net_cfg = {
'hidden1': args.hidden1,
'hidden2': args.hidden2,
'init_w': args.init_w
}
self.actor = Actor(self.nb_states, self.nb_actions, **net_cfg)
self.actor_target = Actor(self.nb_states, self.nb_actions, **net_cfg)
self.actor_optim = Adam(self.actor.parameters(), lr=args.prate)
self.critic = Critic(self.nb_states, self.nb_actions, **net_cfg)
self.critic_target = Critic(self.nb_states, self.nb_actions, **net_cfg)
self.critic_optim = Adam(self.critic.parameters(), lr=args.rate)
self.load_weights(args.output)
hard_update(self.actor_target,
self.actor) # Make sure target is with the same weight
hard_update(self.critic_target, self.critic)
#Create replay buffer
self.memory = SequentialMemory(limit=args.rmsize,
window_length=args.window_length)
self.random_process = OrnsteinUhlenbeckProcess(size=self.nb_actions,
theta=args.ou_theta,
mu=args.ou_mu,
sigma=args.ou_sigma)
# Hyper-parameters
self.batch_size = args.bsize
self.tau = args.tau
self.discount = args.discount
self.depsilon = 1.0 / args.epsilon
#
self.epsilon = 1.0
self.s_t = None # Most recent state
self.a_t = None # Most recent action
self.is_training = True
#
if USE_CUDA: self.cuda()
def __init__(self,
env,
actor_model,
critic_model,
memory=10000,
batch_size=64,
gamma=0.99,
tau=0.001,
actor_lr=1e-4,
critic_lr=1e-3,
critic_decay=1e-2,
ou_theta=0.15,
ou_sigma=0.2,
render=None,
evaluate=None,
save_path=None,
save_every=10,
render_every=10,
train_per_step=True):
self.env = env
self.actor = actor_model
self.actor_target = actor_model.clone()
self.critic = critic_model
self.critic_target = critic_model.clone()
if use_cuda:
for net in [
self.actor, self.actor_target, self.critic,
self.critic_target
]:
net.cuda()
self.memory = ReplayMemory(memory)
self.batch_size = batch_size
self.gamma = gamma
self.tau = tau
self.random_process = OrnsteinUhlenbeckProcess(
env.action_space.shape[0], theta=ou_theta, sigma=ou_sigma)
self.optim_critic = optim.Adam(self.critic.parameters(),
lr=critic_lr,
weight_decay=critic_decay)
self.optim_actor = optim.Adam(self.actor.parameters(), lr=actor_lr)
self.render = render
self.render_every = render_every
self.evaluate = evaluate
self.save_path = save_path
self.save_every = save_every
self.train_per_step = train_per_step
def __init__(self, nb_states, nb_actions, args):
if args.seed > 0:
self.seed(args.seed)
self.nb_states = nb_states
self.nb_actions = nb_actions
# Create Actor and Critic Network
self.actor = Actor(self.nb_states, self.nb_actions, args.init_w)
self.actor_target = Actor(self.nb_states, self.nb_actions, args.init_w)
self.critic = Critic(self.nb_states, self.nb_actions, args.init_w)
self.critic_target = Critic(self.nb_states, self.nb_actions,
args.init_w)
self.reward_predictor = Critic(self.nb_states, self.nb_actions,
args.init_w)
hard_update(self.actor_target,
self.actor) # Make sure target is with the same weight
hard_update(self.critic_target, self.critic)
#Create replay buffer
self.random_process = OrnsteinUhlenbeckProcess(size=nb_actions,
theta=args.ou_theta,
mu=args.ou_mu,
sigma=args.ou_sigma)
# Hyper-parameters
self.batch_size = args.bsize
self.trajectory_length = args.trajectory_length
self.tau = args.tau
self.discount = args.discount
self.depsilon = 1.0 / args.epsilon
#
self.epsilon = 1.0
self.is_training = True
#
if USE_CUDA: self.cuda()
def __init__(self, s_dim, a_dim, n_agents, **kwargs):
self.s_dim = s_dim
self.a_dim = a_dim
self.config = kwargs['config']
self.n_agents = n_agents
self.device = 'cuda' if self.config.use_cuda else 'cpu'
# Networks
self.policy = Actor(s_dim, a_dim, n_agents)
self.policy_target = Actor(s_dim, a_dim, n_agents)
self.critic = Critic(s_dim, a_dim, n_agents)
self.critic_target = Critic(s_dim, a_dim, n_agents)
if self.config.use_cuda:
self.policy.cuda()
self.policy_target.cuda()
self.critic.cuda()
self.critic_target.cuda()
self.policy_optimizer = torch.optim.Adam(self.policy.parameters(),
lr=self.config.a_lr)
self.critic_optimizer = torch.optim.Adam(self.critic.parameters(),
lr=self.config.c_lr)
hard_update(self.policy, self.policy_target)
hard_update(self.critic, self.critic_target)
self.random_process = OrnsteinUhlenbeckProcess(
size=self.a_dim,
theta=self.config.ou_theta,
mu=self.config.ou_mu,
sigma=self.config.ou_sigma)
self.replay_buffer = list()
self.epsilon = 1.
self.depsilon = self.epsilon / self.config.epsilon_decay
self.c_loss = None
self.a_loss = None
self.action_log = list()
def __init__(self, nb_states, nb_actions):
self.nb_states = nb_states
self.nb_actions = nb_actions
# Create Actor and Critic Network
self.actor = Actor(self.nb_states, self.nb_actions)
self.actor_target = Actor(self.nb_states, self.nb_actions)
self.actor_optim = Adam(self.actor.parameters(), lr=ACTOR_LR)
self.critic = Critic(self.nb_states, self.nb_actions)
self.critic_target = Critic(self.nb_states, self.nb_actions)
self.critic_optim = Adam(self.critic.parameters(), lr=CRITIC_LR)
hard_update(self.actor_target,
self.actor) # Make sure target is with the same weight
hard_update(self.critic_target, self.critic)
#Create replay buffer
self.memory = SequentialMemory(limit=MEMORY_SIZE,
window_length=HISTORY_LEN)
self.random_process = OrnsteinUhlenbeckProcess(size=nb_actions,
theta=OU_THETA,
mu=OU_MU,
sigma=OU_SIGMA)
# Hyper-parameters
self.batch_size = BATCH_SIZE
self.tau = TAU
self.discount = GAMMA
self.depsilon = 1.0 / DEPSILON
self.epsilon = 1.0
self.s_t = None # Most recent state
self.a_t = None # Most recent action
self.is_training = True
if USE_CUDA: self.cuda()
class Preyer:
def __init__(self, s_dim, a_dim, **kwargs):
self.s_dim = s_dim
self.a_dim = a_dim
self.config = kwargs['config']
self.device = 'cuda' if self.config.use_cuda else 'cpu'
self.actor = Actor(s_dim, a_dim)
self.actor_target = Actor(s_dim, a_dim)
self.critic = Critic(s_dim, a_dim, 1)
self.critic_target = Critic(s_dim, a_dim, 1)
self.actor_optimizer = torch.optim.Adam(self.actor.parameters(),
lr=self.config.a_lr)
self.critic_optimizer = torch.optim.Adam(self.critic.parameters(),
lr=self.config.c_lr)
self.c_loss = 0
self.a_loss = 0
if self.config.use_cuda:
self.actor.cuda()
self.actor_target.cuda()
self.critic.cuda()
self.critic_target.cuda()
hard_update(self.actor, self.actor_target)
hard_update(self.critic, self.critic_target)
self.random_process = OrnsteinUhlenbeckProcess(
size=self.a_dim,
theta=self.config.ou_theta,
mu=self.config.ou_mu,
sigma=self.config.ou_sigma)
self.replay_buffer = list()
self.epsilon = 1.
self.depsilon = self.epsilon / self.config.epsilon_decay
def memory(self, s, a, r, s_, done):
self.replay_buffer.append((s, a, r, s_, done))
if len(self.replay_buffer) >= self.config.memory_length:
self.replay_buffer.pop(0)
def get_batches(self):
experiences = random.sample(self.replay_buffer, self.config.batch_size)
state_batches = np.array([_[0] for _ in experiences])
action_batches = np.array([_[1] for _ in experiences])
reward_batches = np.array([_[2] for _ in experiences])
next_state_batches = np.array([_[3] for _ in experiences])
done_batches = np.array([_[4] for _ in experiences])
return state_batches, action_batches, reward_batches, next_state_batches, done_batches
def choose_action(self, s, noisy=True):
if self.config.use_cuda:
s = Variable(torch.cuda.FloatTensor(s))
else:
s = Variable(torch.FloatTensor(s))
a = self.actor.forward(s).cpu().detach().numpy()
if noisy:
a += max(self.epsilon, 0.001) * self.random_process.sample()
self.epsilon -= self.depsilon
a = np.clip(a, -1., 1.)
return np.array([a])
def random_action(self):
action = np.random.uniform(low=-1., high=1., size=(1, self.a_dim))
return action
def reset(self):
self.random_process.reset_states()
def train(self):
state_batches, action_batches, reward_batches, next_state_batches, done_batches = self.get_batches(
)
state_batches = Variable(torch.Tensor(state_batches).to(self.device))
action_batches = Variable(
torch.Tensor(action_batches).reshape(-1, 1).to(self.device))
reward_batches = Variable(
torch.Tensor(reward_batches).reshape(-1, 1).to(self.device))
next_state_batches = Variable(
torch.Tensor(next_state_batches).to(self.device))
done_batches = Variable(
torch.Tensor(
(done_batches == False) * 1).reshape(-1, 1).to(self.device))
target_next_actions = self.actor_target.forward(
next_state_batches).detach()
target_next_q = self.critic_target.forward(
next_state_batches, target_next_actions).detach()
main_q = self.critic(state_batches, action_batches)
# Critic Loss
self.critic.zero_grad()
baselines = reward_batches + done_batches * self.config.gamma * target_next_q
loss_critic = torch.nn.MSELoss()(main_q, baselines)
loss_critic.backward()
self.critic_optimizer.step()
# Actor Loss
self.actor.zero_grad()
clear_action_batches = self.actor.forward(state_batches)
loss_actor = (
-self.critic.forward(state_batches, clear_action_batches)).mean()
loss_actor.backward()
self.actor_optimizer.step()
# This is for logging
self.c_loss = loss_critic.item()
self.a_loss = loss_actor.item()
soft_update(self.actor, self.actor_target, self.config.tau)
soft_update(self.critic, self.critic_target, self.config.tau)
def getLoss(self):
return self.c_loss, self.a_loss
class Agent():
def __init__(self, nb_states, nb_actions):
self.critic = Critic(nb_states, nb_actions) # Q
self.critic_target = Critic(nb_states, nb_actions)
self.actor = Actor(nb_states, nb_actions) # policy mu
self.actor_target = Actor(nb_states, nb_actions)
hard_update(self.critic_target, self.critic)
hard_update(self.actor_target, self.actor)
self.critic_optimizer = torch.optim.Adam(self.critic.parameters(),
lr=0.001)
self.actor_optimizer = torch.optim.Adam(self.actor.parameters(),
lr=0.0001)
self.criterion = nn.MSELoss()
self.random_process = OrnsteinUhlenbeckProcess(size=nb_actions,
theta=0.15,
mu=0,
sigma=0.2)
self.gamma = 0.99
self.batch_size = 64
if USE_CUDA:
self.actor.cuda()
self.actor_target.cuda()
self.critic.cuda()
self.critic_target.cuda()
def act(self, obs, epsilon=0.1): # epsilon -> tunning paramter
if (random.random() < epsilon): # choose random action
action = np.random.uniform(-1., 1., nb_actions)
return action
else: # the action is the output of actor network + Exploration Noise
action = self.actor(obs).cpu().data.numpy()
action += self.random_process.sample()
action = np.clip(action, -1., 1.) # to stay in interval [-1,1]
return action
def backward(self, transitions):
transitions = memory.sample(self.batch_size)
batch = Transition(*zip(*transitions))
state_batch = Variable(torch.cat(batch.state)).type(
FLOAT) # size 64 x 3
action_batch = Variable(torch.cat(batch.action)).type(FLOAT) # size 64
next_state_batch = Variable(torch.cat(batch.next_state)).type(
FLOAT) # size 64 x 3
reward_batch = Variable(torch.cat(batch.reward)).type(FLOAT) # size 64
done_batch = Variable(torch.cat(batch.done)).type(FLOAT)
#### Q - CRITIC UPDATE ####
# Q(s_t,a_t)
action_batch.unsqueeze_(1) # size 64x1
state_action_value = self.critic(state_batch, action_batch) # 64x1
# a_{t+1} = mu_target(s_{t+1})
next_action = self.actor_target(
next_state_batch).detach() # 64 x nb_actions
# Q'(s_{t+1},a_{t+1})
next_state_action_value = self.critic_target(next_state_batch,
next_action).detach()
next_state_action_value.squeeze_() # 64
# mask to consider next_state_values to 0 if state is terminal
mask = Variable(
np.logical_not(done_batch.data).type(
torch.FloatTensor)).type(FLOAT)
# mask = 1,1,1 ..
# Compare Q(s_t,a_t) with r_t + gamma * Q'(s_{t+1},a_{t+1})
expected_state_action_value = reward_batch + (
self.gamma * next_state_action_value * mask)
# Compute Huber loss
# loss = F.smooth_l1_loss(state_action_values, expected_state_action_values)
loss = self.criterion(state_action_value, expected_state_action_value)
# Optimize the nn by updating weights with adam descent
self.critic_optimizer.zero_grad()
loss.backward()
self.critic_optimizer.step()
#### mu - ACTOR UPDATE ####
# a_t = mu(s_t)
action = self.actor(state_batch)
# J = esperance[Q(s_t,mu(s_t))] -> a maximiser
# -J = policy_loss -> a minimiser
policy_loss = -self.critic(state_batch, action)
policy_loss = policy_loss.mean()
self.actor_optimizer.zero_grad()
policy_loss.backward()
self.actor_optimizer.step()
#### update target network with polyak averaging
soft_update(self.critic_target, self.critic, tau=0.001)
soft_update(self.actor_target, self.actor, tau=0.001)
return
class Agent(object):
def __init__(self, nb_states, nb_actions, args):
if args.seed > 0:
self.seed(args.seed)
self.nb_states = nb_states
self.nb_actions = nb_actions
# Create Actor and Critic Network
self.actor = Actor(self.nb_states, self.nb_actions, args.init_w)
self.actor_target = Actor(self.nb_states, self.nb_actions, args.init_w)
self.critic = Critic(self.nb_states, self.nb_actions, args.init_w)
self.critic_target = Critic(self.nb_states, self.nb_actions,
args.init_w)
hard_update(self.actor_target,
self.actor) # Make sure target is with the same weight
hard_update(self.critic_target, self.critic)
#Create replay buffer
self.random_process = OrnsteinUhlenbeckProcess(size=nb_actions,
theta=args.ou_theta,
mu=args.ou_mu,
sigma=args.ou_sigma)
# Hyper-parameters
self.batch_size = args.bsize
self.trajectory_length = args.trajectory_length
self.tau = args.tau
self.discount = args.discount
self.depsilon = 1.0 / args.epsilon
#
self.epsilon = 1.0
self.is_training = True
#
if USE_CUDA: self.cuda()
def eval(self):
self.actor.eval()
self.actor_target.eval()
self.critic.eval()
self.critic_target.eval()
def random_action(self):
action = np.random.uniform(-1., 1., self.nb_actions)
return action
def select_action(self, state, noise_enable=True, decay_epsilon=True):
action, _ = self.actor(to_tensor(np.array([state])))
action = to_numpy(action).squeeze(0)
if noise_enable == True:
action += self.is_training * max(self.epsilon,
0) * self.random_process.sample()
action = np.clip(action, -1., 1.)
if decay_epsilon:
self.epsilon -= self.depsilon
return action
def reset_lstm_hidden_state(self, done=True):
self.actor.reset_lstm_hidden_state(done)
def reset(self):
self.random_process.reset_states()
def cuda(self):
self.actor.cuda()
self.actor_target.cuda()
self.critic.cuda()
self.critic_target.cuda()
def load_weights(self, output):
if output is None: return False
self.actor.load_state_dict(torch.load('{}/actor.pkl'.format(output)))
self.critic.load_state_dict(torch.load('{}/critic.pkl'.format(output)))
return True
def save_model(self, output):
if not os.path.exists(output):
os.mkdir(output)
torch.save(self.actor.state_dict(), '{}/actor.pkl'.format(output))
torch.save(self.critic.state_dict(), '{}/critic.pkl'.format(output))
def __init__(self, env, config):
self.name = 'HierarchicalNet'
self.save_folder = None
self.test_record = {}
self.train_record = {}
self.config = config
self.env = env
self.epsilon = config.EPSILON
self.commander_memory = Commander_Memory(config.MEMORY_SIZE,config.BATCH_SIZE)
self.unit_memory = Unit_Memory(2*config.MEMORY_SIZE,config.UNIT_BATCH_SIZE)
self.commander_actor = Commander_Actor(config.STATE_DIM,config.COMMAND_DIM,config.RNN_INSIZE)
self.commander_actor_target = Commander_Actor(config.STATE_DIM,config.COMMAND_DIM,config.RNN_INSIZE)
self.commander_critic = Commander_Critic(config.STATE_DIM,config.COMMAND_DIM,config.BATCH_SIZE,config.RNN_INSIZE)
self.commander_critic_target = Commander_Critic(config.STATE_DIM,config.COMMAND_DIM,config.BATCH_SIZE,config.RNN_INSIZE)
self.unit_actor = Unit_Actor(config.STATE_DIM,config.COMMAND_DIM,config.ACTION_DIM)
self.unit_actor_target = Unit_Actor(config.STATE_DIM,config.COMMAND_DIM,config.ACTION_DIM)
self.unit_critic = Unit_Critic(config.STATE_DIM,config.COMMAND_DIM,config.ACTION_DIM,config.HIDDEN_SIZE)
self.unit_critic_target = Unit_Critic(config.STATE_DIM,config.COMMAND_DIM,config.ACTION_DIM,config.HIDDEN_SIZE)
self.commander_actor_h0 = Variable(torch.zeros(2, 1, config.RNN_OUTSIZE),requires_grad=False)
if config.GPU >= 0:
self.commander_actor.cuda(device=config.GPU)
self.commander_actor_target.cuda(device=config.GPU)
self.commander_critic.cuda(device=config.GPU)
self.commander_critic_target.cuda(device=config.GPU)
self.unit_actor.cuda(device=config.GPU)
self.unit_actor_target.cuda(device=config.GPU)
self.unit_critic.cuda(device=config.GPU)
self.unit_critic_target.cuda(device=config.GPU)
self.commander_critic.h0 = self.commander_critic.h0.cuda(device=config.GPU)
self.commander_critic_target.h0 = self.commander_critic_target.h0.cuda(device=config.GPU)
self.commander_actor_h0 = self.commander_actor_h0.cuda(device=config.GPU)
copy_parameter(self.commander_actor, self.commander_actor_target)
copy_parameter(self.commander_critic, self.commander_critic_target)
copy_parameter(self.unit_actor, self.unit_actor_target)
copy_parameter(self.unit_critic, self.unit_critic_target)
self.commander_actor_optimizer = optim.Adam(self.commander_actor.parameters(),lr=config.ACTOR_LR)
self.unit_actor_optimizer = optim.Adam(self.unit_actor.parameters(),lr=config.ACTOR_LR)
self.commander_critic_optimizer = optim.Adam(self.commander_critic.parameters(), lr=config.CRITIC_LR)
self.unit_critic_optimizer = optim.Adam(self.unit_critic.parameters(), lr=config.CRITIC_LR)
self.criterion = nn.MSELoss()
self.action_noise = OrnsteinUhlenbeckProcess(size=(config.MYSELF_NUM, config.ACTION_DIM), theta=10, mu=0., sigma=2)
self.command_noise = OrnsteinUhlenbeckProcess(size=(1,config.MYSELF_NUM, config.COMMAND_DIM), theta=10, mu=0., sigma=2)
# self.action_noise = OrnsteinUhlenbeckProcess(size=(config.MYSELF_NUM, config.ACTION_DIM), theta=30, mu=0., sigma=3)
# self.command_noise = OrnsteinUhlenbeckProcess(size=(1,config.MYSELF_NUM, config.COMMAND_DIM), theta=30, mu=0., sigma=3)
# normalize
state_normalization_myelf = [1,100,100,1,100,100,1]
state_normalization_enemy = [1,100,100,100,100,10,100,100,1,1,1,10]
self.state_normalization = state_normalization_myelf
for i in range(config.K):
self.state_normalization += state_normalization_enemy
self.state_normalization = np.asarray(self.state_normalization,dtype=np.float32)
def __init__(
self,
env,
mem_size=int(1e6),
lr_critic=1e-3,
lr_actor=1e-4,
max_epi=int(1e4),
epsilon_decay=1. / (1e5),
gamma=.99,
target_update_frequency=200,
batch_size=64,
random_process_mode='default',
max_step=None,
actor_update_mode='default',
popart=False,
actor='standard',
critic='43',
epsilon_start=1.,
epsilon_end=.01,
epsilon_rate=1. / 200,
partition_num=100,
env_log_freq=100,
model_log_freq=500,
target_update_mode='hard',
tau=1e-3,
grad_clip_mode=None,
grad_clip_norm=5.,
critic_weight_decay=0.,
exp_trunc=[],
exp_percent=[],
exp_rebalance_freq=None,
exp_type='rank',
):
# configuration log
frame = inspect.currentframe()
args, _, _, values = inspect.getargvalues(frame)
self.config = ['{}: {}'.format(arg, values[arg]) for arg in args]
self.CUDA = torch.cuda.is_available()
self.ENV_NORMALIZED = env.class_name() == 'NormalizedEnv'
self.POPART = popart
self.actor_update_mode = actor_update_mode
self.orig_env = (env) #for recording
if max_step is not None:
tmp_env = env
if isinstance(tmp_env, gym.Wrapper):
while (tmp_env.class_name() != 'TimeLimit'):
tmp_env = tmp_env.env
tmp_env._max_episode_steps = max_step
self.env = self.orig_env
self.N = 1
if hasattr(self.env.unwrapped, 'N'):
self.N = self.env.unwrapped.N
self.N_S = self.env.observation_space.shape[0]
self.N_A = self.env.action_space.shape[0]
self.n_s = self.N_S / self.N
self.n_a = self.N_A / self.N
self.MAX_EPI = max_epi
self.LOW = self.env.action_space.low
self.HIGH = self.env.action_space.high
self.actor = ActorRegistry[actor](self.n_s, self.n_a)
self.critic = CriticRegistry[critic](self.N_S, self.N_A)
self.target_actor = ActorRegistry[actor](self.n_s, self.n_a)
self.target_critic = CriticRegistry[critic](self.N_S, self.N_A)
self.target_actor.eval()
self.target_critic.eval()
self.target_actor.load_state_dict(self.actor.state_dict())
self.target_critic.load_state_dict(self.critic.state_dict())
if self.CUDA:
self.actor.cuda()
self.critic.cuda()
self.target_actor.cuda()
self.target_critic.cuda()
# pop-art
self.update_counter = 0
self.beta = .1
self.y_mean = 0.
self.y_square_mean = 0.
self.target_y_mean = self.y_mean
self.target_y_square_mean = self.y_square_mean
# per
self.total_step = 0
self.PARTITION_NUM = partition_num
self.LEARN_START = mem_size / self.PARTITION_NUM + 1
self.exp_trunc = exp_trunc
self.exp_percent = exp_percent
self.exp_rebalance_freq = exp_rebalance_freq
self.exp_batch_sizes = []
self.exp_type = exp_type
#if len(self.exp_trunc)>0:
if len(self.exp_trunc) != len(self.exp_percent):
raise RuntimeError("different exp_trunc and exp_percent length")
self.exp = []
for i in range(len(self.exp_trunc) + 1):
tmp_batch_size = int(
batch_size * (1 - sum(self.exp_percent))) if i == len(
self.exp_trunc) else int(batch_size * self.exp_percent[i])
self.exp_batch_sizes.append(tmp_batch_size)
exp_conf = {
'size': mem_size,
'learn_start': self.LEARN_START,
'partition_num': self.PARTITION_NUM,
'total_step': self.MAX_EPI * 50,
'batch_size': tmp_batch_size
}
self.exp.append(Experience(
exp_conf)) if self.exp_type == 'rank' else self.exp.append(
PrioritizedReplayBuffer(mem_size, alpha=.7))
#else:
# exp_conf = {
# 'size': mem_size,
# 'learn_start': self.LEARN_START,
# 'partition_num': self.PARTITION_NUM,
# 'total_step': self.MAX_EPI * 50,
# 'batch_size': batch_size,
# }
# self.exp = Experience(exp_conf)
# uniform er
#self.exp = Experience(mem_size)
self.optim_critic = optim.Adam(self.critic.parameters(),
lr=lr_critic,
weight_decay=critic_weight_decay)
self.optim_actor = optim.Adam(self.actor.parameters(), lr=-lr_actor)
self.random_processes = []
for _ in xrange(self.N):
random_process = OrnsteinUhlenbeckProcess(\
size=self.n_a, theta=.15, mu=0, sigma=.2)
self.random_processes.append(random_process)
self.EPSILON_START = epsilon_start
self.EPSILON_END = epsilon_end
# only default random process mode will use epsilon decay
self.EPSILON_DECAY = epsilon_decay
# other random process mode will use epsilon rate
self.EPSILON_RATE = epsilon_rate
self.GAMMA = gamma
self.TARGET_UPDATE_FREQUENCY = target_update_frequency
self.BATCH_SIZE = batch_size
self.target_update_mode = target_update_mode
self.tau = tau
#title = {common.S_EPI:[], common.S_TOTAL_R:[]}
#self.data = pd.DataFrame(title)
self.RAND_PROC = random_process_mode
self.grad_clip_mode = grad_clip_mode
self.grad_clip_norm = grad_clip_norm
# logger
self.logger = None
self.env_log_freq = env_log_freq
self.model_log_freq = model_log_freq
self.step = 0
# random seed
self.seed = int(time.time())
random.seed(self.seed)
np.random.seed(self.seed)
def run_agent(model_params, weights, state_transform, data_queue, weights_queue,
process, global_step, updates, best_reward, param_noise_prob, save_dir,
max_steps=10000000):
train_fn, actor_fn, target_update_fn, params_actor, params_crit, actor_lr, critic_lr = \
build_model(**model_params)
actor = Agent(actor_fn, params_actor, params_crit)
actor.set_actor_weights(weights)
env = RunEnv2(state_transform, max_obstacles=config.num_obstacles, skip_frame=config.skip_frames)
random_process = OrnsteinUhlenbeckProcess(theta=.1, mu=0., sigma=.2, size=env.noutput,
sigma_min=0.05, n_steps_annealing=1e6)
# prepare buffers for data
states = []
actions = []
rewards = []
terminals = []
total_episodes = 0
start = time()
action_noise = True
while global_step.value < max_steps:
seed = random.randrange(2**32-2)
state = env.reset(seed=seed, difficulty=2)
random_process.reset_states()
total_reward = 0.
total_reward_original = 0.
terminal = False
steps = 0
while not terminal:
state = np.asarray(state, dtype='float32')
action = actor.act(state)
if action_noise:
action += random_process.sample()
next_state, reward, next_terminal, info = env.step(action)
total_reward += reward
total_reward_original += info['original_reward']
steps += 1
global_step.value += 1
# add data to buffers
states.append(state)
actions.append(action)
rewards.append(reward)
terminals.append(terminal)
state = next_state
terminal = next_terminal
if terminal:
break
total_episodes += 1
# add data to buffers after episode end
states.append(state)
actions.append(np.zeros(env.noutput))
rewards.append(0)
terminals.append(terminal)
states_np = np.asarray(states).astype(np.float32)
data = (states_np,
np.asarray(actions).astype(np.float32),
np.asarray(rewards).astype(np.float32),
np.asarray(terminals),
)
weight_send = None
if total_reward > best_reward.value:
weight_send = actor.get_actor_weights()
# send data for training
data_queue.put((process, data, weight_send, total_reward))
# receive weights and set params to weights
weights = weights_queue.get()
report_str = 'Global step: {}, steps/sec: {:.2f}, updates: {}, episode len {}, ' \
'reward: {:.2f}, original_reward {:.4f}; best reward: {:.2f} noise {}'. \
format(global_step.value, 1. * global_step.value / (time() - start), updates.value, steps,
total_reward, total_reward_original, best_reward.value, 'actions' if action_noise else 'params')
print(report_str)
with open(os.path.join(save_dir, 'train_report.log'), 'a') as f:
f.write(report_str + '\n')
actor.set_actor_weights(weights)
action_noise = np.random.rand() < 1 - param_noise_prob
if not action_noise:
set_params_noise(actor, states_np, random_process.current_sigma)
# clear buffers
del states[:]
del actions[:]
del rewards[:]
del terminals[:]
if total_episodes % 100 == 0:
env = RunEnv2(state_transform, max_obstacles=config.num_obstacles, skip_frame=config.skip_frames)
class RDPG_v2:
def __init__(self, conf, device):
self.conf = conf
self.state_dim = conf['state_dim']
self.action_dim = conf['action_dim']
self.device = device
# create actor and critic network
self.actor = Actor_RDPG(self.state_dim,
self.action_dim).to(self.device)
self.actor_target = Actor_RDPG(self.state_dim,
self.action_dim).to(self.device)
self.critic = Critic_RDPG(self.state_dim,
self.action_dim).to(self.device)
self.critic_target = Critic_RDPG(self.state_dim,
self.action_dim).to(self.device)
hard_update(self.actor_target,
self.actor) # Make sure target is with the same weight
hard_update(self.critic_target, self.critic)
self.critic_optim = optim.Adam(self.critic.parameters(), lr=q_lr)
self.actor_optim = optim.Adam(self.actor.parameters(), lr=policy_lr)
#Create replay buffer
self.random_process = OrnsteinUhlenbeckProcess(size=self.action_dim,
theta=0.15,
mu=0.0,
sigma=0.2)
# args.ou_theta:0.15 (noise theta), args.ou_sigma:0.2 (noise sigma), args.out_mu:0.0 (noise mu)
self.epsilon = 1.0
self.depsilon = 1.0 / 50000
self.is_training = True
self.tau = 0.001 # moving average for target network
def random_action(self):
action = np.random.uniform(
0., 1., self.action_dim) # [-1,1] select as a number of action_dim
return action
def select_action(self, state, noise_enable=True, decay_epsilon=True):
action, _ = self.actor(
to_tensor(state).reshape(-1).unsqueeze(0)
) # input shape = [batch(=1) X state_dim], action : type (tuple), shape [batch X action_dim]
action = action.cpu().detach().numpy().squeeze(
0) # action shape [action_dim,]
if noise_enable == True:
action += self.is_training * max(self.epsilon,
0) * self.random_process.sample()
action = np.clip(action, 0.,
1.) # input 중 -1~1 을 벗어나는 값에 대해 -1 or 1 로 대체
if decay_epsilon:
self.epsilon -= self.depsilon
return action
def update_policy(self, memory, gamma=0.99):
print("updating...")
# Sample batch
experiences = memory.sample(
self.conf['batch_size']
) # type: list | shape: (max_epi_length(2000)-1 X batch(32) X 5(??))
if len(experiences) == 0: # not enough samples
return
dtype = torch.cuda.FloatTensor
policy_loss_total = 0
value_loss_total = 0
for t in range(len(experiences) - 1): # iterate over episodes
# print("t:", t)
target_cx = Variable(torch.zeros(self.conf['batch_size'],
50)).type(dtype)
target_hx = Variable(torch.zeros(self.conf['batch_size'],
50)).type(dtype)
cx = Variable(torch.zeros(self.conf['batch_size'], 50)).type(dtype)
hx = Variable(torch.zeros(self.conf['batch_size'], 50)).type(dtype)
# we first get the data out of the sampled experience
# shape of state0, action, reward: [batch X state_dim], [batch X 1], [batch X 1]
state0 = np.stack([
trajectory.state0 for trajectory in experiences[t]
]) # batch 개수만큼 각 epi 중 t 시점에서 상태만 추출
# action = np.expand_dims(np.stack((trajectory.action for trajectory in experiences[t])), axis=1)
action = np.stack(
[trajectory.action for trajectory in experiences[t]])
reward = np.expand_dims(np.stack(
[trajectory.reward for trajectory in experiences[t]]),
axis=1)
# reward = np.stack((trajectory.reward for trajectory in experiences[t]))
state1 = np.stack(
[trajectory.state0 for trajectory in experiences[t + 1]])
target_action, (target_hx, target_cx) = self.actor_target(
to_tensor(state1).reshape(self.conf['batch_size'], -1),
(target_hx, target_cx))
next_q_value = self.critic_target([
to_tensor(state1).reshape(self.conf['batch_size'], -1),
target_action
])
target_q = to_tensor(reward) + gamma * next_q_value
# Critic update
current_q = self.critic([
to_tensor(state0).reshape(self.conf['batch_size'], -1),
to_tensor(action)
])
value_loss = F.smooth_l1_loss(current_q, target_q)
value_loss /= len(experiences) # divide by trajectory length
value_loss_total += value_loss
# update per trajectory
self.critic.zero_grad()
value_loss.backward()
# Actor update
action, (hx, cx) = self.actor(
to_tensor(state0).reshape(self.conf['batch_size'], -1),
(hx, cx))
policy_loss = -self.critic([
to_tensor(state0).reshape(self.conf['batch_size'], -1), action
])
policy_loss /= len(experiences) # divide by trajectory length
policy_loss_total += policy_loss.mean()
policy_loss = policy_loss.mean()
self.actor.zero_grad()
policy_loss.backward()
self.critic_optim.step()
self.actor_optim.step()
# Target update
soft_update(self.actor_target, self.actor, self.tau)
soft_update(self.critic_target, self.critic, self.tau)
print("update finish!")
def reset_lstm_hidden_state(self, done=True):
self.actor.reset_lstm_hidden_state(done)
def save_model(self, path):
torch.save(self.critic.state_dict(), path + '_q')
torch.save(self.critic_target.state_dict(), path + '_target_q')
torch.save(self.actor.state_dict(), path + '_policy')
def load_model(self, path):
self.critic.load_state_dict(torch.load(path + '_q'))
self.critic_target.load_state_dict(torch.load(path + '_target_q'))
self.actor.load_state_dict(torch.load(path + '_policy'))
self.critic.eval()
self.critic_target.eval()
self.actor.eval()
class DDPG(object):
def __init__(self, args, nb_states, nb_actions):
USE_CUDA = torch.cuda.is_available()
if args.seed > 0:
self.seed(args.seed)
self.nb_states = nb_states
self.nb_actions= nb_actions
self.gpu_ids = [i for i in range(args.gpu_nums)] if USE_CUDA and args.gpu_nums > 0 else [-1]
self.gpu_used = True if self.gpu_ids[0] >= 0 else False
net_cfg = {
'hidden1':args.hidden1,
'hidden2':args.hidden2,
'init_w':args.init_w
}
self.actor = Actor(self.nb_states, self.nb_actions, **net_cfg).double()
self.actor_target = Actor(self.nb_states, self.nb_actions, **net_cfg).double()
self.actor_optim = Adam(self.actor.parameters(), lr=args.p_lr, weight_decay=args.weight_decay)
self.critic = Critic(self.nb_states, self.nb_actions, **net_cfg).double()
self.critic_target = Critic(self.nb_states, self.nb_actions, **net_cfg).double()
self.critic_optim = Adam(self.critic.parameters(), lr=args.c_lr, weight_decay=args.weight_decay)
hard_update(self.actor_target, self.actor) # Make sure target is with the same weight
hard_update(self.critic_target, self.critic)
#Create replay buffer
self.memory = SequentialMemory(limit=args.rmsize, window_length=args.window_length)
self.random_process = OrnsteinUhlenbeckProcess(size=self.nb_actions,
theta=args.ou_theta, mu=args.ou_mu, sigma=args.ou_sigma)
# Hyper-parameters
self.batch_size = args.bsize
self.tau_update = args.tau_update
self.gamma = args.gamma
# Linear decay rate of exploration policy
self.depsilon = 1.0 / args.epsilon
# initial exploration rate
self.epsilon = 1.0
self.s_t = None # Most recent state
self.a_t = None # Most recent action
self.is_training = True
self.continious_action_space = False
def update_policy(self):
pass
def cuda_convert(self):
if len(self.gpu_ids) == 1:
if self.gpu_ids[0] >= 0:
with torch.cuda.device(self.gpu_ids[0]):
print('model cuda converted')
self.cuda()
if len(self.gpu_ids) > 1:
self.data_parallel()
self.cuda()
self.to_device()
print('model cuda converted and paralleled')
def eval(self):
self.actor.eval()
self.actor_target.eval()
self.critic.eval()
self.critic_target.eval()
def cuda(self):
self.actor.cuda()
self.actor_target.cuda()
self.critic.cuda()
self.critic_target.cuda()
def data_parallel(self):
self.actor = nn.DataParallel(self.actor, device_ids=self.gpu_ids)
self.actor_target = nn.DataParallel(self.actor_target, device_ids=self.gpu_ids)
self.critic = nn.DataParallel(self.critic, device_ids=self.gpu_ids)
self.critic_target = nn.DataParallel(self.critic_target, device_ids=self.gpu_ids)
def to_device(self):
self.actor.to(torch.device('cuda:{}'.format(self.gpu_ids[0])))
self.actor_target.to(torch.device('cuda:{}'.format(self.gpu_ids[0])))
self.critic.to(torch.device('cuda:{}'.format(self.gpu_ids[0])))
self.critic_target.to(torch.device('cuda:{}'.format(self.gpu_ids[0])))
def observe(self, r_t, s_t1, done):
if self.is_training:
self.memory.append(self.s_t, self.a_t, r_t, done)
self.s_t = s_t1
def random_action(self):
action = np.random.uniform(-1.,1.,self.nb_actions)
# self.a_t = action
return action
def select_action(self, s_t, decay_epsilon=True):
# proto action
action = to_numpy(
self.actor(to_tensor(np.array([s_t]), gpu_used=self.gpu_used, gpu_0=self.gpu_ids[0])),
gpu_used=self.gpu_used
).squeeze(0)
action += self.is_training * max(self.epsilon, 0) * self.random_process.sample()
action = np.clip(action, -1., 1.)
if decay_epsilon:
self.epsilon -= self.depsilon
# self.a_t = action
return action
def reset(self, s_t):
self.s_t = s_t
self.random_process.reset_states()
def load_weights(self, dir):
if dir is None: return
if self.gpu_used:
# load all tensors to GPU (gpu_id)
ml = lambda storage, loc: storage.cuda(self.gpu_ids)
else:
# load all tensors to CPU
ml = lambda storage, loc: storage
self.actor.load_state_dict(
torch.load('output/{}/actor.pkl'.format(dir), map_location=ml)
)
self.critic.load_state_dict(
torch.load('output/{}/critic.pkl'.format(dir), map_location=ml)
)
print('model weights loaded')
def save_model(self,output):
if len(self.gpu_ids) == 1 and self.gpu_ids[0] > 0:
with torch.cuda.device(self.gpu_ids[0]):
torch.save(
self.actor.state_dict(),
'{}/actor.pt'.format(output)
)
torch.save(
self.critic.state_dict(),
'{}/critic.pt'.format(output)
)
elif len(self.gpu_ids) > 1:
torch.save(self.actor.module.state_dict(),
'{}/actor.pt'.format(output)
)
torch.save(self.actor.module.state_dict(),
'{}/critic.pt'.format(output)
)
else:
torch.save(
self.actor.state_dict(),
'{}/actor.pt'.format(output)
)
torch.save(
self.critic.state_dict(),
'{}/critic.pt'.format(output)
)
def seed(self,seed):
torch.manual_seed(seed)
if len(self.gpu_ids) > 0:
torch.cuda.manual_seed_all(seed)
class BiCNet():
def __init__(self, s_dim, a_dim, n_agents, **kwargs):
self.s_dim = s_dim
self.a_dim = a_dim
self.config = kwargs['config']
self.n_agents = n_agents
self.device = 'cuda' if self.config.use_cuda else 'cpu'
# Networks
self.policy = Actor(s_dim, a_dim, n_agents)
self.policy_target = Actor(s_dim, a_dim, n_agents)
self.critic = Critic(s_dim, a_dim, n_agents)
self.critic_target = Critic(s_dim, a_dim, n_agents)
if self.config.use_cuda:
self.policy.cuda()
self.policy_target.cuda()
self.critic.cuda()
self.critic_target.cuda()
self.policy_optimizer = torch.optim.Adam(self.policy.parameters(),
lr=self.config.a_lr)
self.critic_optimizer = torch.optim.Adam(self.critic.parameters(),
lr=self.config.c_lr)
hard_update(self.policy, self.policy_target)
hard_update(self.critic, self.critic_target)
self.random_process = OrnsteinUhlenbeckProcess(
size=self.a_dim,
theta=self.config.ou_theta,
mu=self.config.ou_mu,
sigma=self.config.ou_sigma)
self.replay_buffer = list()
self.epsilon = 1.
self.depsilon = self.epsilon / self.config.epsilon_decay
self.c_loss = None
self.a_loss = None
self.action_log = list()
def choose_action(self, obs, noisy=True):
obs = torch.Tensor([obs]).to(self.device)
action = self.policy(obs).cpu().detach().numpy()[0]
self.action_log.append(action)
if noisy:
for agent_idx in range(self.n_agents):
pass
# action[agent_idx] += self.epsilon * self.random_process.sample()
self.epsilon -= self.depsilon
self.epsilon = max(self.epsilon, 0.001)
np.clip(action, -1., 1.)
return action
def reset(self):
self.random_process.reset_states()
self.action_log.clear()
def prep_train(self):
self.policy.train()
self.critic.train()
self.policy_target.train()
self.critic_target.train()
def prep_eval(self):
self.policy.eval()
self.critic.eval()
self.policy_target.eval()
self.critic_target.eval()
def random_action(self):
return np.random.uniform(low=-1, high=1, size=(self.n_agents, 2))
def memory(self, s, a, r, s_, done):
self.replay_buffer.append((s, a, r, s_, done))
if len(self.replay_buffer) >= self.config.memory_length:
self.replay_buffer.pop(0)
def get_batches(self):
experiences = random.sample(self.replay_buffer, self.config.batch_size)
state_batches = np.array([_[0] for _ in experiences])
action_batches = np.array([_[1] for _ in experiences])
reward_batches = np.array([_[2] for _ in experiences])
next_state_batches = np.array([_[3] for _ in experiences])
done_batches = np.array([_[4] for _ in experiences])
return state_batches, action_batches, reward_batches, next_state_batches, done_batches
def train(self):
state_batches, action_batches, reward_batches, next_state_batches, done_batches = self.get_batches(
)
state_batches = torch.Tensor(state_batches).to(self.device)
action_batches = torch.Tensor(action_batches).to(self.device)
reward_batches = torch.Tensor(reward_batches).reshape(
self.config.batch_size, self.n_agents, 1).to(self.device)
next_state_batches = torch.Tensor(next_state_batches).to(self.device)
done_batches = torch.Tensor(
(done_batches == False) * 1).reshape(self.config.batch_size,
self.n_agents,
1).to(self.device)
target_next_actions = self.policy_target.forward(next_state_batches)
target_next_q = self.critic_target.forward(next_state_batches,
target_next_actions)
main_q = self.critic(state_batches, action_batches)
'''
How to concat each agent's Q value?
'''
#target_next_q = target_next_q
#main_q = main_q.mean(dim=1)
'''
Reward Norm
'''
# reward_batches = (reward_batches - reward_batches.mean(dim=0)) / reward_batches.std(dim=0) / 1024
# Critic Loss
self.critic.zero_grad()
baselines = reward_batches + done_batches * self.config.gamma * target_next_q
loss_critic = torch.nn.MSELoss()(main_q, baselines.detach())
loss_critic.backward()
torch.nn.utils.clip_grad_norm_(self.critic.parameters(), 0.5)
self.critic_optimizer.step()
# Actor Loss
self.policy.zero_grad()
clear_action_batches = self.policy.forward(state_batches)
loss_actor = -self.critic.forward(state_batches,
clear_action_batches).mean()
loss_actor += (clear_action_batches**2).mean() * 1e-3
loss_actor.backward()
torch.nn.utils.clip_grad_norm_(self.policy.parameters(), 0.5)
self.policy_optimizer.step()
# This is for logging
self.c_loss = loss_critic.item()
self.a_loss = loss_actor.item()
soft_update(self.policy, self.policy_target, self.config.tau)
soft_update(self.critic, self.critic_target, self.config.tau)
def get_loss(self):
return self.c_loss, self.a_loss
def get_action_std(self):
return np.array(self.action_log).std(axis=-1).mean()
class DDPG(object):
def __init__(self, nb_states, nb_actions, args):
if args.seed > 0:
self.seed(args.seed)
self.nb_states = nb_states
self.nb_actions = nb_actions
# Create Actor and Critic Network
net_cfg = {
"hidden1": args.hidden1,
"hidden2": args.hidden2,
"init_w": args.init_w,
}
self.actor = Actor(self.nb_states, self.nb_actions, **net_cfg)
self.actor_target = Actor(self.nb_states, self.nb_actions, **net_cfg)
self.actor_optim = Adam(self.actor.parameters(), lr=args.prate)
self.critic = Critic(self.nb_states, self.nb_actions, **net_cfg)
self.critic_target = Critic(self.nb_states, self.nb_actions, **net_cfg)
self.critic_optim = Adam(self.critic.parameters(), lr=args.rate)
hard_update(
self.actor_target, self.actor
) # Make sure target is with the same weight
hard_update(self.critic_target, self.critic)
# Create replay buffer
self.memory = SequentialMemory(
limit=args.rmsize, window_length=args.window_length
)
self.random_process = OrnsteinUhlenbeckProcess(
size=nb_actions, theta=args.ou_theta, mu=args.ou_mu, sigma=args.ou_sigma
)
# Hyper-parameters
self.batch_size = args.bsize
self.tau = args.tau
self.discount = args.discount
self.depsilon = 1.0 / args.epsilon
#
self.epsilon = 1.0
self.s_t = None # Most recent state
self.a_t = None # Most recent action
self.is_training = True
#
if USE_CUDA:
self.cuda()
def update_policy(self):
# Sample batch
(
state_batch,
action_batch,
reward_batch,
next_state_batch,
terminal_batch,
) = self.memory.sample_and_split(self.batch_size)
# Prepare for the target q batch
next_q_values = self.critic_target(
[
to_tensor(next_state_batch, volatile=True),
self.actor_target(to_tensor(next_state_batch, volatile=True)),
]
)
# next_q_values.volatile = False
target_q_batch = (
to_tensor(reward_batch)
+ self.discount * to_tensor(terminal_batch.astype(np.float)) * next_q_values
)
# Critic update
self.critic.zero_grad()
q_batch = self.critic([to_tensor(state_batch), to_tensor(action_batch)])
value_loss = criterion(q_batch, target_q_batch)
value_loss.backward()
self.critic_optim.step()
# Actor update
self.actor.zero_grad()
policy_loss = -self.critic(
[to_tensor(state_batch), self.actor(to_tensor(state_batch))]
)
policy_loss = policy_loss.mean()
policy_loss.backward()
self.actor_optim.step()
# Target update
soft_update(self.actor_target, self.actor, self.tau)
soft_update(self.critic_target, self.critic, self.tau)
def eval(self):
self.actor.eval()
self.actor_target.eval()
self.critic.eval()
self.critic_target.eval()
def cuda(self):
self.actor.cuda()
self.actor_target.cuda()
self.critic.cuda()
self.critic_target.cuda()
def observe(self, r_t, s_t1, done):
if self.is_training:
self.memory.append(self.s_t, self.a_t, r_t, done)
self.s_t = s_t1
def random_action(self):
action = np.random.uniform(-1.0, 1.0, self.nb_actions)
self.a_t = action
return action
def select_action(self, s_t, decay_epsilon=True):
action = to_numpy(self.actor(to_tensor(np.array([s_t])))).squeeze(0)
action += self.is_training * max(self.epsilon, 0) * self.random_process.sample()
action = np.clip(action, -1.0, 1.0)
if decay_epsilon:
self.epsilon -= self.depsilon
self.a_t = action
return action
def reset(self, obs):
self.s_t = obs
self.random_process.reset_states()
def load_weights(self, output):
if output is None:
return
self.actor.load_state_dict(torch.load("{}/actor.pkl".format(output)))
self.critic.load_state_dict(torch.load("{}/critic.pkl".format(output)))
def save_model(self, output):
torch.save(self.actor.state_dict(), "{}/actor.pkl".format(output))
torch.save(self.critic.state_dict(), "{}/critic.pkl".format(output))
def seed(self, s):
torch.manual_seed(s)
if USE_CUDA:
torch.cuda.manual_seed(s)
def run_agent(args,
model_params,
weights,
data_queue,
weights_queue,
process,
global_step,
updates,
best_reward,
param_noise_prob,
save_dir,
max_steps=10000000):
train_fn, actor_fn, target_update_fn, params_actor, params_crit, actor_lr, critic_lr = build_model(
**model_params)
actor = Agent(actor_fn, params_actor, params_crit)
actor.set_actor_weights(weights)
env = RunEnv2(model=args.modeldim,
prosthetic=args.prosthetic,
difficulty=args.difficulty,
skip_frame=config.skip_frames)
env.spec.timestep_limit = 3000 # ndrw
# random_process = OrnsteinUhlenbeckProcess(theta=.1, mu=0., sigma=.3, size=env.noutput, sigma_min=0.05, n_steps_annealing=1e6)
sigma_rand = random.uniform(0.05, 0.5)
dt_rand = random.uniform(0.002, 0.02)
param_noise_prob = random.uniform(param_noise_prob * 0.25,
min(param_noise_prob * 1.5, 1.))
random_process = OrnsteinUhlenbeckProcess(theta=.1,
mu=0.,
sigma=sigma_rand,
dt=dt_rand,
size=env.noutput,
sigma_min=0.05,
n_steps_annealing=1e6)
print('OUProcess_sigma = ' + str(sigma_rand) + ' OUProcess_dt = ' +
str(dt_rand) + ' param_noise_prob = ' + str(param_noise_prob))
# prepare buffers for data
states = []
actions = []
rewards = []
terminals = []
total_episodes = 0
start = time()
action_noise = True
while global_step.value < max_steps:
seed = random.randrange(2**32 - 2)
state = env.reset(seed=seed, difficulty=args.difficulty)
random_process.reset_states()
total_reward = 0.
total_reward_original = 0.
terminal = False
steps = 0
while not terminal:
state = np.asarray(state, dtype='float32')
action = actor.act(state)
if action_noise:
action += random_process.sample()
next_state, reward, next_terminal, info = env._step(action)
total_reward += reward
total_reward_original += info['original_reward']
steps += 1
global_step.value += 1
# add data to buffers
states.append(state)
actions.append(action)
rewards.append(reward)
terminals.append(terminal)
state = next_state
terminal = next_terminal
if terminal:
break
total_episodes += 1
# add data to buffers after episode end
states.append(state)
actions.append(np.zeros(env.noutput))
rewards.append(0)
terminals.append(terminal)
states_np = np.asarray(states).astype(np.float32)
data = (
states_np,
np.asarray(actions).astype(np.float32),
np.asarray(rewards).astype(np.float32),
np.asarray(terminals),
)
weight_send = None
if total_reward > best_reward.value:
weight_send = actor.get_actor_weights()
# send data for training
data_queue.put((process, data, weight_send, total_reward))
# receive weights and set params to weights
weights = weights_queue.get()
# report_str = 'Global step: {}, steps/sec: {:.2f}, updates: {}, episode len: {}, pelvis_X: {:.2f}, reward: {:.2f}, original_reward {:.4f}, best reward: {:.2f}, noise: {}'. \
# format(global_step.value, 1. * global_step.value / (time() - start), updates.value, steps, info['pelvis_X'], total_reward, total_reward_original, best_reward.value, 'actions' if action_noise else 'params')
# report_str = 'Global step: {}, steps/sec: {:.2f}, updates: {}, episode len: {}, pelvis_X: {:.2f}, reward: {:.2f}, best reward: {:.2f}, noise: {}'. \
# format(global_step.value, 1. * global_step.value / (time() - start), updates.value, steps, info['pelvis_X'], total_reward, best_reward.value, 'actions' if action_noise else 'params')
report_str = 'Global step: {}, steps/sec: {:.2f}, updates: {}, episode len: {}, pelvis_X: {:.2f}, pelvis_Z: {:.2f}, reward: {:.2f}, best reward: {:.2f}, noise: {}'. \
format(global_step.value, 1. * global_step.value / (time() - start), updates.value, steps, info['pelvis'][0], info['pelvis'][2], total_reward, best_reward.value, 'actions' if action_noise else 'params')
print(report_str)
try:
with open(os.path.join(save_dir, 'train_report.log'), 'a') as f:
f.write(report_str + '\n')
except:
print('#############################################')
print(
'except » with open(os.path.join(save_dir, train_report.log), a) as f:'
)
print('#############################################')
actor.set_actor_weights(weights)
action_noise = np.random.rand() < 1 - param_noise_prob
if not action_noise:
set_params_noise(actor, states_np, random_process.current_sigma)
# clear buffers
del states[:]
del actions[:]
del rewards[:]
del terminals[:]
if total_episodes % 100 == 0:
env = RunEnv2(model=args.modeldim,
prosthetic=args.prosthetic,
difficulty=args.difficulty,
skip_frame=config.skip_frames)
class DDPG_trainer(object):
def __init__(self, nb_state, nb_action):
self.nb_state = nb_state
self.nb_action = nb_action
self.actor = Actor(self.nb_state, self.nb_action)
self.actor_target = Actor(self.nb_state, self.nb_action)
self.actor_optim = Adam(self.actor.parameters(), lr=LEARNING_RATE)
self.critic = Critic(self.nb_state, self.nb_action)
self.critic_target = Critic(self.nb_state, self.nb_action)
self.critic_optim = Adam(self.critic.parameters(), lr=LEARNING_RATE)
hard_update(self.actor_target,
self.actor) # Make sure target is with the same weight
hard_update(self.critic_target, self.critic)
#Create replay buffer
self.memory = SequentialMemory(limit=MEMORY_SIZE, window_length=1)
self.random_process = OrnsteinUhlenbeckProcess(size=nb_action,
theta=OU_THETA,
mu=OU_MU,
sigma=OU_SIGMA)
self.is_training = True
self.epsilon = 1.0
self.a_t = None
self.s_t = None
if USE_CUDA: self.cuda()
def cuda(self):
self.actor.cuda()
self.actor_target.cuda()
self.critic.cuda()
self.critic_target.cuda()
def select_action(self, s_t, decay_epsilon=True):
action = to_numpy(self.actor(to_tensor(np.array([s_t])))).squeeze(0)
action += self.is_training * max(self.epsilon,
0) * self.random_process.sample()
action = np.clip(action, -1., 1.)
if decay_epsilon:
self.epsilon -= DELTA_EPSILON
self.a_t = action
return action
def reset(self, observation):
self.start_state = observation
self.random_process.reset_states()
def observe(self, r_t, s_t1, done):
if self.is_training:
self.memory.append(self.s_t, self.a_t, r_t, done)
self.s_t = s_t1
def update_all(self):
# Help Warm Up
if self.memory.nb_entries < BATCH_SIZE * 2:
return
# Sample batch
state_batch, action_batch, reward_batch, \
next_state_batch, terminal_batch = self.memory.sample_and_split(BATCH_SIZE)
# Prepare for the target q batch
with torch.no_grad():
next_q_values = self.critic_target([
to_tensor(next_state_batch),
self.actor_target(to_tensor(next_state_batch)),
])
target_q_batch = to_tensor(reward_batch) + \
DISCOUNT * to_tensor(terminal_batch.astype(np.float)) * next_q_values
# Critic update
self.critic.zero_grad()
for state in state_batch:
if state.shape[0] <= 2:
# print("Error sampled memory!")
return
q_batch = self.critic(
[to_tensor(state_batch),
to_tensor(action_batch)])
value_loss = CRITERION(q_batch, target_q_batch)
value_loss.backward()
self.critic_optim.step()
# Actor update
self.actor.zero_grad()
policy_loss = -self.critic(
[to_tensor(state_batch),
self.actor(to_tensor(state_batch))])
policy_loss = policy_loss.mean()
policy_loss.backward()
self.actor_optim.step()
# Target update
soft_update(self.actor_target, self.actor, TAU)
soft_update(self.critic_target, self.critic, TAU)
def __init__(self, gamma, tau, actor_hidden_size, critic_hidden_size,
observation_space, action_space, args):
self.num_inputs = observation_space.shape[0]
self.action_space = action_space
self.actor_hidden_size = actor_hidden_size
self.critic_hidden_size = critic_hidden_size
self.comm_hidden_size = actor_hidden_size // 2
self.gamma = gamma
self.tau = tau
self.args = args
# replay for the update of attention unit
self.queue = queue.Queue()
# Define actor part 1
self.actor_p1 = ActorPart1(self.num_inputs,
actor_hidden_size).to(device)
self.actor_target_p1 = ActorPart1(self.num_inputs,
actor_hidden_size).to(device)
# attention unit is not end-to-end trained
self.atten = AttentionUnit(actor_hidden_size,
actor_hidden_size).to(device)
self.atten_optim = Adam(self.atten.parameters(), lr=self.args.actor_lr)
# Define Communication Channel
self.comm = CommunicationChannel(actor_hidden_size,
self.comm_hidden_size).to(device)
self.comm_target = CommunicationChannel(
actor_hidden_size, self.comm_hidden_size).to(device)
self.comm_optim = Adam(self.comm.parameters(), lr=self.args.actor_lr)
# Define actor part 2
# input -- [thoughts, intergrated thoughts]
self.actor_p2 = ActorPart2(
actor_hidden_size + self.comm_hidden_size * 2, self.action_space,
actor_hidden_size).to(device)
self.actor_target_p2 = ActorPart2(
actor_hidden_size + self.comm_hidden_size * 2, self.action_space,
actor_hidden_size).to(device)
self.actor_optim = Adam([{
'params': self.actor_p1.parameters(),
'lr': self.args.actor_lr
}, {
'params': self.actor_p2.parameters(),
'lr': self.args.actor_lr
}])
self.critic = Critic(self.num_inputs, self.action_space,
critic_hidden_size).to(device)
self.critic_target = Critic(self.num_inputs, self.action_space,
critic_hidden_size).to(device)
self.critic_optim = Adam(self.critic.parameters(),
lr=self.args.critic_lr)
# Make sure target is with the same weight
hard_update(self.actor_target_p1, self.actor_p1)
hard_update(self.comm_target, self.comm)
hard_update(self.actor_target_p2, self.actor_p2)
hard_update(self.critic_target, self.critic)
# Create replay buffer
self.memory = ReplayMemory(args.memory_size)
self.random_process = OrnsteinUhlenbeckProcess(size=action_space.n,
theta=args.ou_theta,
mu=args.ou_mu,
sigma=args.ou_sigma)
critic = Critic(state_dim, action_dim, max_action, args)
critic_t = Critic(state_dim, action_dim, max_action, args)
critic_t.load_state_dict(critic.state_dict())
print("OK 3")
# actor
actor = Actor(state_dim, action_dim, max_action, args)
actor_t = Actor(state_dim, action_dim, max_action, args)
actor_t.load_state_dict(actor.state_dict())
# action noise
if not args.ou_noise:
a_noise = GaussianNoise(action_dim, sigma=args.gauss_sigma)
else:
a_noise = OrnsteinUhlenbeckProcess(action_dim,
mu=args.ou_mu,
theta=args.ou_theta,
sigma=args.ou_sigma)
if USE_CUDA:
critic.cuda()
critic_t.cuda()
actor.cuda()
actor_t.cuda()
print("OK 4")
# CEM
es = sepCEM(actor.get_size(),
mu_init=actor.get_params(),
sigma_init=args.sigma_init,
damp=args.damp,
damp_limit=args.damp_limit,
class DDPG(object):
def __init__(self,
state_size,
action_size,
memory_size,
batch_size=128,
tan=0.001,
actor_lr=0.001,
critic_lr=0.001,
epsilon=1.):
self.state_size = state_size
self.action_size = action_size
self.batch_size = batch_size
self.tan = tan
self.warmup = WARM_UP
self.epsilon = epsilon
self.epsilon_decay = hyperparameters['D_EPSILON']
self.actor = Actor(state_size, action_size)
self.actor_target = Actor(state_size, action_size)
self.actor_optimizer = optim.Adam(self.actor.parameters(), lr=actor_lr)
self.critic = Critic(state_size, action_size)
self.critic_target = Critic(state_size, action_size)
self.critic_optimizer = optim.Adam(self.critic.parameters(),
lr=critic_lr)
self.memory = Memory(memory_size)
self.criterion = nn.MSELoss()
self.random_process = OrnsteinUhlenbeckProcess(size=action_size,
theta=0.15,
mu=0.,
sigma=0.2)
copy_parameter(self.actor, self.actor_target)
copy_parameter(self.critic, self.critic_target)
def train(self):
# if not warm up
if self.memory.counter < self.warmup:
return
# get batch
state_batch, action_batch, next_state_batch, reward_batch, done_batch = self.memory.sample(
self.batch_size)
action_batch = action_batch.reshape((-1, self.action_size))
reward_batch = reward_batch.reshape((-1, 1))
done_batch = done_batch.reshape((-1, 1))
# update critic
nsb = Variable(torch.from_numpy(next_state_batch).float(),
volatile=True) # next_state_batch
nab = self.actor_target(nsb) # next_action_batch
next_q = self.critic_target(nsb, nab)
next_q.volatile = False
rb = Variable(torch.from_numpy(reward_batch).float()) # reward_batch
db = Variable(torch.from_numpy(done_batch).float(
)) # if next state is None, next_q should be 0, which means q = r
q_target = rb + hyperparameters['GAMMA'] * db * next_q
sb_grad = Variable(torch.from_numpy(state_batch).float()
) # state_batch with grad, mean output need grad
ab = Variable(torch.from_numpy(action_batch).float()) # action_batch
q_eval = self.critic(sb_grad, ab)
value_loss = self.criterion(q_eval, q_target)
self.critic.zero_grad()
value_loss.backward()
# nn.utils.clip_grad_norm(self.critic.parameters(),0.8)
self.critic_optimizer.step()
# update actor
sb_grad = Variable(
torch.from_numpy(state_batch).float()) # state_batch
aab = self.actor(sb_grad) # actor_action_batch
q = self.critic(sb_grad, aab)
policy_loss = torch.mean(-q)
self.actor.zero_grad()
policy_loss.backward()
# nn.utils.clip_grad_norm(self.actor.parameters(),0.8)
self.actor_optimizer.step()
# update parameter between two network
update_parameter(self.critic_target, self.critic, self.tan)
update_parameter(self.actor_target, self.actor, self.tan)
def select_action(self, s, is_train=True, decay_e=True):
if self.memory.counter < self.warmup:
action = env.action_space.sample()[0]
# action = random.uniform(-2.,2.)
return action
state = Variable(torch.FloatTensor([s]).float())
action = self.actor(state).squeeze(1).data.numpy()
action += is_train * max(self.epsilon,
0) * self.random_process.sample()
action = float(np.clip(action, -1., 1.)[0])
if decay_e:
if self.memory.counter > self.warmup:
self.epsilon -= self.epsilon_decay
return action