class A2CAgent():
"""
init function
input: env, which is the CartPole-v0
gamma, 0.99 in this case
lr, learning rate is 1e-4
define: env = env, which is the CartPole-v0
obs_dim: 4 obervations
Observation:
Type: Box(4)
Num Observation Min Max
0 Cart Position -4.8 4.8
1 Cart Velocity -Inf Inf
2 Pole Angle -24 deg 24 deg
3 Pole Velocity At Tip -Inf Inf
action_dim: 2 actions
Actions:
Type: Discrete(2)
Num Action
0 Push cart to the left
1 Push cart to the right
value_network: two layer network with input 4 (observation dim) and output 1 (reward?)
policy_network: two layer network with input 4 (observation dim) and output 2 (action dim)
value and policy optimizer using default Adam and learning rate
"""
def __init__(self, env, gamma, lr):
self.env = env
self.obs_dim = env.observation_space.shape[0]
self.action_dim = env.action_space.n
self.gamma = gamma
self.lr = lr
self.value_network = ValueNetwork(self.obs_dim, 1)
self.policy_network = PolicyNetwork(self.obs_dim, self.action_dim)
self.value_optimizer = optim.Adam(self.value_network.parameters(),
lr=self.lr)
self.policy_optimizer = optim.Adam(self.policy_network.parameters(),
lr=self.lr)
"""
input state to get the next action
using policy network to get the next state by using softmax
"""
def get_action(self, state):
state = torch.FloatTensor(state)
logits = self.policy_network.forward(state)
dist = F.softmax(logits, dim=0)
probs = Categorical(dist)
return probs.sample().cpu().detach().item()
"""
form trajectory get all of the information, and calculated the discounted_rewards
use value network to train the states with new values and compute the loss between the value and target value by using MSE
same logic for policy network
FloatTensor = FLOAT TYPE ARRAY
t
tensor([[1, 2, 3],
[4, 5, 6]])
t.view(-1,1)
tensor([[1],
[2],
[3],
[4],
[5],
[6]])
"""
def compute_loss(self, trajectory):
states = torch.FloatTensor([sars[0] for sars in trajectory])
actions = torch.LongTensor([sars[1]
for sars in trajectory]).view(-1, 1)
rewards = torch.FloatTensor([sars[2] for sars in trajectory])
next_states = torch.FloatTensor([sars[3] for sars in trajectory])
dones = torch.FloatTensor([sars[4] for sars in trajectory]).view(-1, 1)
# compute value target
## Two for loop to calculate the discounted reward for each one
discounted_rewards = [torch.sum(torch.FloatTensor([self.gamma ** i for i in range(rewards[j:].size(0))]) \
* rewards[j:]) for j in
range(rewards.size(0))] # sorry, not the most readable code.
value_targets = rewards.view(
-1, 1) + torch.FloatTensor(discounted_rewards).view(-1, 1)
# compute value loss
values = self.value_network.forward(states)
value_loss = F.mse_loss(values, value_targets.detach())
# compute policy loss with entropy bonus
logits = self.policy_network.forward(states)
dists = F.softmax(logits, dim=1)
probs = Categorical(dists)
# compute entropy bonus
entropy = []
for dist in dists:
entropy.append(-torch.sum(dist.mean() * torch.log(dist)))
entropy = torch.stack(entropy).sum()
advantage = value_targets - values
policy_loss = -probs.log_prob(actions.view(actions.size(0))).view(
-1, 1) * advantage.detach()
policy_loss = policy_loss.mean() - 0.001 * entropy
return value_loss, policy_loss
"""
zero_grad clears old gradients from the last step (otherwise you’d just accumulate the gradients from all loss.backward() calls).
loss.backward() computes the derivative of the loss w.r.t. the parameters (or anything requiring gradients) using backpropagation.
opt.step() causes the optimizer to take a step based on the gradients of the parameters.
"""
def update(self, trajectory):
value_loss, policy_loss = self.compute_loss(trajectory)
self.value_optimizer.zero_grad()
value_loss.backward()
self.value_optimizer.step()
self.policy_optimizer.zero_grad()
policy_loss.backward()
self.policy_optimizer.step()
class SACAgent:
def __init__(self, env, gamma, tau, v_lr, q_lr, policy_lr, buffer_maxlen):
self.device = torch.device(
"cuda" if torch.cuda.is_available() else "cpu")
self.env = env
# self.action_range = [env.action_space.low, env.action_space.high]
# TODO: as a simple demo, I changed here; for the implementation, we should pass this as parameters
self.action_range = [[-1, 1], [-1, 1]]
self.obs_dim = env.observation_space.shape[0]
self.action_dim = 2
# self.action_dim = 1
# hyperparameters
self.gamma = gamma
self.tau = tau
self.update_step = 0
self.delay_step = 2
# initialize networks
self.value_net = ValueNetwork(self.obs_dim, 1).to(self.device)
self.target_value_net = ValueNetwork(self.obs_dim, 1).to(self.device)
self.q_net1 = SoftQNetwork(self.obs_dim,
self.action_dim).to(self.device)
self.q_net2 = SoftQNetwork(self.obs_dim,
self.action_dim).to(self.device)
self.policy_net = PolicyNetwork(self.obs_dim,
self.action_dim).to(self.device)
# copy params to target param
for target_param, param in zip(self.target_value_net.parameters(),
self.value_net.parameters()):
target_param.data.copy_(param)
# initialize optimizers
self.value_optimizer = optim.Adam(self.value_net.parameters(), lr=v_lr)
self.q1_optimizer = optim.Adam(self.q_net1.parameters(), lr=q_lr)
self.q2_optimizer = optim.Adam(self.q_net2.parameters(), lr=q_lr)
self.policy_optimizer = optim.Adam(self.policy_net.parameters(),
lr=policy_lr)
self.replay_buffer = BasicBuffer(buffer_maxlen)
# pi: state -> acton
def get_action(self, state):
state = torch.FloatTensor(state).unsqueeze(0).to(self.device)
mean, log_std = self.policy_net.forward(state)
std = log_std.exp()
normal = Normal(mean, std)
z = normal.sample()
action = torch.tanh(z)
action = action.cpu().detach().squeeze(0).numpy()
return self.rescale_action(action)
def rescale_action(self, action):
'''if action < 0.5:
return 0
else:
return 1'''
scaled_action = []
for idx, a in enumerate(action):
action_range = self.action_range[idx]
a = (action_range[1] - action_range[0]) / 2.0 + (
action_range[1] + action_range[0]) / 2.0
scaled_action.append(a)
return scaled_action
def update(self, batch_size):
states, actions, rewards, next_states, dones = self.replay_buffer.sample(
batch_size)
states = torch.FloatTensor(states).to(self.device)
actions = torch.FloatTensor(actions).to(self.device)
rewards = torch.FloatTensor(rewards).to(self.device)
next_states = torch.FloatTensor(next_states).to(self.device)
dones = torch.FloatTensor(dones).to(self.device)
dones = dones.view(dones.size(0), -1)
next_actions, next_log_pi = self.policy_net.sample(next_states)
next_q1 = self.q_net1(next_states, next_actions)
next_q2 = self.q_net2(next_states, next_actions)
next_v = self.target_value_net(next_states)
# value Loss
next_v_target = torch.min(next_q1, next_q2) - next_log_pi
curr_v = self.value_net.forward(states)
v_loss = F.mse_loss(curr_v, next_v_target.detach())
#TODO: Question: why using 2 Q-networks?
# To reduce bias in training.
# q loss
curr_q1 = self.q_net1.forward(states, actions)
curr_q2 = self.q_net2.forward(states, actions)
expected_q = rewards + (1 - dones) * self.gamma * next_v
q1_loss = F.mse_loss(curr_q1, expected_q.detach())
q2_loss = F.mse_loss(curr_q2, expected_q.detach())
# update value network and q networks
self.value_optimizer.zero_grad()
v_loss.backward()
self.value_optimizer.step()
self.q1_optimizer.zero_grad()
q1_loss.backward()
self.q1_optimizer.step()
self.q2_optimizer.zero_grad()
q2_loss.backward()
self.q2_optimizer.step()
# delayed update for policy net and target value nets
# TODO: Question: what does this part do?
# The original paper mentioned 2 methods for approximating the value function
# 1. the EMA of policy weights to update the Q network
# 2. periodical update of the policy network, which is used in this code
if self.update_step % self.delay_step == 0:
new_actions, log_pi = self.policy_net.sample(states)
min_q = torch.min(self.q_net1.forward(states, new_actions),
self.q_net2.forward(states, new_actions))
policy_loss = (log_pi - min_q).mean()
self.policy_optimizer.zero_grad()
policy_loss.backward()
self.policy_optimizer.step()
# target networks
for target_param, param in zip(self.target_value_net.parameters(),
self.value_net.parameters()):
target_param.data.copy_(self.tau * param +
(1 - self.tau) * target_param)
self.update_step += 1
