class Agent():
"""Interacts with and learns from the environment.
Agent is made of several functions:
init : create an agent
step : save experience in memory and decide if it is time to learn
act : decide the most relevant action in function of the environment state and current knowledge
learn : learn for past experiences
Use the soft_update function to update the target Qnetwork
"""
def __init__(self, state_size, action_size, seed):
"""Initialize an Agent object.
Params
======
state_size (int): dimension of each state
action_size (int): dimension of each action
seed (int): random seed
"""
self.state_size = state_size
self.action_size = action_size
self.seed = random.seed(seed)
# Q-Network
self.qnetwork_local = QNetwork(state_size, action_size, seed).to(DEVICE)
self.qnetwork_target = QNetwork(state_size, action_size, seed).to(DEVICE)
self.optimizer = optim.Adam(self.qnetwork_local.parameters(), lr=LR)
# Replay memory
self.memory = ReplayBuffer(action_size, BUFFER_SIZE, BATCH_SIZE, seed)
# Initialize time step (for updating every UPDATE_EVERY steps)
self.t_step = 0
def step(self, state, action, reward, next_state, done):
"""Function triggering the learning phase if enough exploration has been done
At each step, record the experience in the memory
After collecting experiences, learn from them
Params
======
experience = state, action, reward, next_state, done
"""
# Save experience in replay memory
self.memory.add(state, action, reward, next_state, done)
# Learn every UPDATE_EVERY time steps.
self.t_step = (self.t_step + 1) % UPDATE_EVERY
if self.t_step == 0:
# If enough samples are available in memory, get random subset and learn
if len(self.memory) > BATCH_SIZE:
experiences = self.memory.sample()
self.learn(experiences, GAMMA)
def act(self, state, eps=0.):
"""Returns actions for given state as per current policy.
Params
======
state (array_like): current state
eps (float): epsilon, for epsilon-greedy action selection
"""
state = torch.from_numpy(state).float().unsqueeze(0).to(DEVICE)
# switch network into evaluation mode when not training
self.qnetwork_local.eval()
# deactivate gradient tracking as we do not to back propagate
with torch.no_grad():
action_values = self.qnetwork_local(state)
# switch network into training mode
self.qnetwork_local.train()
# Epsilon-greedy action selection
if random.random() > eps:
return np.argmax(action_values.cpu().data.numpy())
else:
return random.choice(np.arange(self.action_size))
def learn(self, experiences, gamma):
"""Update value parameters using given batch of experience tuples.
Params
======
experiences (Tuple[torch.Tensor]): tuple of (s, a, r, s', done) tuples
gamma (float): discount factor
"""
states, actions, rewards, next_states, dones = experiences
# Get max predicted Q values (for next states) from target model
Q_targets_next = self.qnetwork_target(next_states).detach().max(1)[0].unsqueeze(1)
# Compute Q targets for current states
Q_targets = rewards + (gamma * Q_targets_next * (1 - dones))
# Get expected Q values from local model
Q_expected = self.qnetwork_local(states).gather(1, actions)
# Compute loss using mean square error
loss = F.mse_loss(Q_expected, Q_targets)
# Minimize the loss
self.optimizer.zero_grad()
# update network weights
loss.backward()
self.optimizer.step()
# ------------------- update target network ------------------- #
self.soft_update(self.qnetwork_local, self.qnetwork_target, TAU)
def soft_update(self, local_model, target_model, tau):
"""Soft update model parameters.
θ_target = τ*θ_local + (1 - τ)*θ_target
Params
======
local_model (PyTorch model): weights will be copied from
target_model (PyTorch model): weights will be copied to
tau (float): interpolation parameter
"""
for target_param, local_param in zip(target_model.parameters(), local_model.parameters()):
target_param.data.copy_(tau*local_param.data + (1.0-tau)*target_param.data)
class Agent():
"""Interacts with and learns from the environment."""
def __init__(self, state_size, action_size, seed):
"""Initialize an Agent object.
Params
======
state_size (int): dimension of each state
action_size (int): dimension of each action
seed (int): random seed
"""
self.state_size = state_size
self.action_size = action_size
self.seed = seed
# Q-Network
self.qnetwork_local = QNetwork(state_size, action_size,
seed).to(device)
self.qnetwork_target = QNetwork(state_size, action_size,
seed).to(device)
self.optimizer = optim.Adam(self.qnetwork_local.parameters(), lr=LR)
# Replay memory
self.memory = ReplayBuffer(action_size, BUFFER_SIZE, BATCH_SIZE, seed)
# Initialize time step (for updating every UPDATE_EVERY steps)
self.t_step = 0
def step(self, state, action, reward, next_state, done):
# Save experience in replay memory
self.memory.add(state, action, reward, next_state, done)
# Learn every UPDATE_EVERY time steps.
self.t_step = (self.t_step + 1) % UPDATE_EVERY
if self.t_step == 0:
# If enough samples are available in memory, get random subset and learn
if len(self.memory) > BATCH_SIZE:
experiences = self.memory.sample()
self.learn(experiences, GAMMA)
def act(self, state, eps=0.):
"""Returns actions for given state as per current policy.
Params
======
state (array_like): current state
eps (float): epsilon, for epsilon-greedy action selection
"""
state = torch.from_numpy(state).float().unsqueeze(0).to(device)
self.qnetwork_local.eval()
with torch.no_grad():
action_values = self.qnetwork_local(state)
self.qnetwork_local.train()
# Epsilon-greedy action selection
if random.random() > eps:
return np.argmax(action_values.cpu().data.numpy())
else:
return random.choice(np.arange(self.action_size))
def learn(self, experiences, gamma):
"""Update value parameters using given batch of experience tuples.
Params
======
experiences (Tuple[torch.Variable]): tuple of (s, a, r, s', done) tuples
gamma (float): discount factor
"""
states, actions, rewards, next_states, dones = experiences
# get max predicted Q values (for next states) from target model
Q_targets_next = self.qnetwork_target(next_states).detach().max(
1)[0].unsqueeze(1)
# compute Q targets for current states
Q_targets = rewards + (gamma * Q_targets_next * (1 - dones))
# Get expected Q values from local model
Q_expected = self.qnetwork_local(states).gather(1, actions)
# Compute loss
loss = F.mse_loss(Q_expected, Q_targets)
# Minimize the loss
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
# ------------------- update target network ------------------- #
self.soft_update(self.qnetwork_local, self.qnetwork_target, TAU)
def soft_update(self, local_model, target_model, tau):
"""Soft update model parameters.
θ_target = τ*θ_local + (1 - τ)*θ_target
Params
======
local_model (PyTorch model): weights will be copied from
target_model (PyTorch model): weights will be copied to
tau (float): interpolation parameter
"""
for target_param, local_param in zip(target_model.parameters(),
local_model.parameters()):
target_param.data.copy_(tau * local_param.data +
(1.0 - tau) * target_param.data)
class Agent():
"""Interacts with and learns from the environment."""
def __init__(self, state_size, action_size, random_seed):
"""Initialize an Agent object.
Params
======
state_size (int): dimension of each state
action_size (int): dimension of each action
random_seed (int): random seed
"""
self.state_size = state_size
self.action_size = action_size
self.seed = random.seed(random_seed)
# Q-Network
self.qnetwork_local = QNetwork(state_size, action_size,
random_seed).to(device)
self.qnetwork_target = QNetwork(state_size, action_size,
random_seed).to(device)
self.optimizer = optim.Adam(self.qnetwork_local.parameters(), lr=LR)
# Replay memory
self.memory = ReplayBuffer(action_size, BUFFER_SIZE, BATCH_SIZE,
random_seed)
# Initialize time step (for updating every UPDATE_EVERY steps)
self.t_step = 0
def step(self, state, action, reward, next_state, done):
# Save experience in replay memory
self.memory.add(state, action, reward, next_state, done)
# Learn every UPDATE_EVERY time steps.
self.t_step = (self.t_step + 1) % UPDATE_EVERY
if self.t_step == 0:
# If enough samples are available in memory, get random subset and learn
if len(self.memory) > BATCH_SIZE:
experiences = self.memory.sample()
self.learn(experiences, GAMMA)
def act(self, state, eps=0.):
"""Returns actions for given state as per current policy.
Params
======
state (array_like): current state
eps (float): epsilon, for epsilon-greedy action selection
"""
state = torch.from_numpy(state).float().unsqueeze(0).to(device)
self.qnetwork_local.eval()
with torch.no_grad():
action_values = self.qnetwork_local(state)
self.qnetwork_local.train()
# Epsilon-greedy action selection
if random.random() > eps:
return np.argmax(action_values.cpu().data.numpy())
else:
return random.choice(np.arange(self.action_size))
def learn(self, experiences, gamma):
"""Update value parameters using given batch of experience tuples.
Params
======
experiences (Tuple[torch.Tensor]): tuple of (s, a, r, s', done) tuples
gamma (float): discount factor
"""
states, actions, rewards, next_states, dones = experiences
# Get max predicted Q values (for next states) from target model
Q_targets_next = self.qnetwork_target(next_states).detach().max(
1)[0].unsqueeze(1)
# Compute Q targets for current states
Q_targets = rewards + (gamma * Q_targets_next * (1 - dones))
# Get expected Q values from local model
Q_expected = self.qnetwork_local(states).gather(1, actions)
# Compute loss
loss = F.mse_loss(Q_expected, Q_targets)
# Minimize the loss
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
# ------------------- update target network ------------------- #
self.soft_update(self.qnetwork_local, self.qnetwork_target, TAU)
def soft_update(self, local_model, target_model, tau):
"""Soft update model parameters.
θ_target = τ*θ_local + (1 - τ)*θ_target
Params
======
local_model (PyTorch model): weights will be copied from
target_model (PyTorch model): weights will be copied to
tau (float): interpolation parameter
"""
for target_param, local_param in zip(target_model.parameters(),
local_model.parameters()):
target_param.data.copy_(tau * local_param.data +
(1.0 - tau) * target_param.data)
def train(self, env, n_episodes=1000, max_t=1000):
scores_deque = deque(maxlen=100)
scores = []
eps = 1.0
for i_episode in range(1, n_episodes + 1):
state = env.reset()
score = 0
for t in range(max_t):
action = self.act(state, eps)
next_state, reward, done, _ = env.step(action)
self.step(state, action, reward, next_state, done)
state = next_state
score += reward
if eps > EPS_MIN:
eps *= EPS_DECAY
if done:
break
scores_deque.append(score)
scores.append(score)
print('\rEpisode {}\tAverage Score: {:.2f}\tScore: {:.2f}'.format(
i_episode, np.mean(scores_deque), score),
end="")
if i_episode % 100 == 0:
torch.save(self.qnetwork_local.state_dict(), 'checkpoint.pth')
print('\rEpisode {}\tAverage Score: {:.2f}\tEps: {:.2f}'.format(
i_episode, np.mean(scores_deque), eps))
return scores
def load(self):
self.qnetwork_local.load_state_dict(torch.load('checkpoint.pth'))
self.qnetwork_local.eval()
self.qnetwork_target.load_state_dict(torch.load('checkpoint.pth'))
self.qnetwork_target.eval()
class DQN_Agent():
"""Interacts with and learns from the environment."""
def __init__(self,
state_size,
action_size,
brain_name,
seed,
params=default_params,
verbose=False,
device=None):
"""Initialize an Agent object.
Params
======
state_size (int): dimension of each state
action_size (int): dimension of each action
seed (int): random seed
"""
params = self._fill_params(params)
# implementation details and identity
self.device = device if device is not None else torch.device(
"cuda:0" if torch.cuda.is_available() else "cpu")
self.name = params['name']
self.brain_name = brain_name
# set environment information
self.state_size = state_size
self.action_size = action_size
self.seed = random.seed(seed)
# Q-Network
self.qnetwork_local = QNetwork(state_size,
action_size,
seed,
layers=params['layers']).to(self.device)
self.qnetwork_target = QNetwork(state_size,
action_size,
seed,
layers=params['layers']).to(
self.device)
self.optimizer = optim.Adam(self.qnetwork_local.parameters(),
lr=params['learning_rate'])
# Replay memory
self.memory = ReplayBuffer(action_size, params['buffer_size'],
params['batch_size'], seed)
# Initialize time steps, t_step for updates, c_step for copying weights
self.t_step = 0
self.c_step = 0
# store params for later
self.params = params
def _fill_params(self, src_params):
params = {
'name':
self._get_param_or_default('name', src_params, default_params),
'layers':
self._get_param_or_default('layers', src_params, default_params),
'buffer_size':
self._get_param_or_default('buffer_size', src_params,
default_params),
'batch_size':
self._get_param_or_default('batch_size', src_params,
default_params),
'update_every':
self._get_param_or_default('update_every', src_params,
default_params),
'copy_every':
self._get_param_or_default('copy_every', src_params,
default_params),
'learning_rate':
self._get_param_or_default('learning_rate', src_params,
default_params),
'gamma':
self._get_param_or_default('gamma', src_params, default_params),
'tau':
self._get_param_or_default('tau', src_params, default_params)
}
return params
def _get_param_or_default(self, key, src_params, default_params):
if key in src_params:
return src_params[key]
else:
return default_params[key]
def display_params(self, force_print=False):
p = '{}: h{}, exp[{}, {}], u,c[{}, {}], g,t,lr[{}, {}, {}]'.format(
self.params['name'], self.params['layers'],
self.params['buffer_size'], self.params['batch_size'],
self.params['update_every'], self.params['copy_every'],
self.params['gamma'], self.params['tau'],
self.params['learning_rate'])
if force_print:
print(p)
return p
def step(self, state, action, reward, next_state, done):
# Save experience in replay memory
self.memory.add(state, action, reward, next_state, done)
# update timesteps
self.t_step = (self.t_step + 1) % self.params['update_every']
self.c_step = (self.c_step + 1) % self.params['copy_every']
# only update every params.UpdateEvery timesteps
if self.t_step == 0:
if len(self.memory) >= self.params['batch_size']:
# get random subset and learn
experiences = self.memory.sample()
self.learn(experiences, self.params['gamma'])
def act(self, state, eps=0.):
"""Returns actions for given state as per current policy.
Params
======
state (array_like): current state
eps (float): epsilon, for epsilon-greedy action selection
"""
# get action values for state
state = torch.from_numpy(state).float().unsqueeze(0).to(self.device)
self.qnetwork_local.eval()
with torch.no_grad():
action_values = self.qnetwork_local(state)
self.qnetwork_local.train()
# Epsilon-greedy action selection
if random.random() > eps:
return np.argmax(action_values.cpu().data.numpy())
else:
return random.choice(np.arange(self.action_size))
def learn(self, experiences, gamma):
"""Update value parameters using given batch of experience tuples.
Params
======
experiences (Tuple[torch.Tensor]): tuple of (s, a, r, s', done) tuples
gamma (float): discount factor
"""
states, actions, rewards, next_states, dones = experiences
# Get max predicted Q values (for next states) from target model
Q_targets_next = self.qnetwork_target(next_states).detach().max(
1)[0].unsqueeze(1)
# Compute Q targets for current states
Q_targets = rewards + (gamma * Q_targets_next * (1 - dones))
# Get expected Q values from local model for evaluation
Q_expected = self.qnetwork_local(states).gather(1, actions)
# Minimize the loss
loss = F.mse_loss(Q_expected, Q_targets)
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
# update the weights in the target network every c_steps
if self.c_step == 0:
self.soft_update(self.qnetwork_local, self.qnetwork_target,
self.params['tau'])
def soft_update(self, local_model, target_model, tau):
"""Soft update model parameters.
θ_target = τ*θ_local + (1 - τ)*θ_target
Params
======
local_model (PyTorch model): weights will be copied from
target_model (PyTorch model): weights will be copied to
tau (float): interpolation parameter (0 = all target, 1 = all local)
"""
for target_param, local_param in zip(target_model.parameters(),
local_model.parameters()):
target_param.data.copy_(tau * local_param.data +
(1.0 - tau) * target_param.data)
class DQNAgent():
"""Interacts with and learns from the environment."""
def __init__(self,
state_size,
action_size,
seed,
hidden_layer1=64,
hidden_layer2=108):
"""Initialize an Agent object.
Params
======
state_size (int): dimension of each state
action_size (int): dimension of each action
seed (int): random seed
"""
self.state_size = state_size
self.action_size = action_size
self.seed = random.seed(seed)
# Q-Network
self.qnetwork_local = QNetwork(state_size, action_size, seed,
hidden_layer1, hidden_layer2).to(device)
self.qnetwork_target = QNetwork(state_size, action_size, seed,
hidden_layer1,
hidden_layer2).to(device)
self.optimizer = optim.Adam(self.qnetwork_local.parameters(), lr=LR)
# Replay memory
self.memory = ReplayBuffer(action_size, BUFFER_SIZE, BATCH_SIZE, seed)
# Initialize time step (for updating every UPDATE_EVERY steps)
self.t_step = 0
def step(self, state, action, reward, next_state, done):
# Save experience in replay memory
self.memory.add(state, action, reward, next_state, done)
# Learn every UPDATE_EVERY time steps.
self.t_step = (self.t_step + 1) % UPDATE_EVERY
if self.t_step == 0:
# If enough samples are available in memory, get random subset and learn
if len(self.memory) > BATCH_SIZE:
experiences = self.memory.sample()
self.learn(experiences, GAMMA)
def act(self, state, eps=0.):
"""Returns actions for given state as per current policy.
Params
======
state (array_like): current state
eps (float): epsilon, for epsilon-greedy action selection
"""
state = torch.from_numpy(state).float().unsqueeze(0).to(device)
self.qnetwork_local.eval()
with torch.no_grad():
action_values = self.qnetwork_local(state)
self.qnetwork_local.train()
# Epsilon-greedy action selection
if random.random() > eps:
return np.argmax(action_values.cpu().data.numpy())
else:
return random.choice(np.arange(self.action_size))
def save(self, file_name):
torch.save(self.qnetwork_local.state_dict(), file_name)
def load(self, file_name):
self.qnetwork_local.load_state_dict(torch.load(file_name))
def save_bin(self, file_name):
torch.save(self.qnetwork_local.state_dict(), file_name)
keymap = self.qnetwork_local.state_dict()
outputfloats = []
for key, value in keymap.items():
# ignore keys order is all that matteers
if isinstance(value, torch.Tensor):
tup = value.size()
lv = value.tolist()
if isinstance(lv, float):
outputfloats.append(lv)
else:
for row in lv:
if isinstance(row, float):
outputfloats.append(row)
else:
for item in row:
outputfloats.append(item)
while (len(outputfloats) < 8192):
outputfloats.append(0.0)
output_file = open(file_name, 'wb')
float_array = array('f', outputfloats)
float_array.tofile(output_file)
output_file.close()
def load_bin(self, file_name):
keymap = self.qnetwork_local.state_dict()
new_keymap = OrderedDict()
sz = os.path.getsize(file_name)
input_file = open(file_name, 'rb')
n = sz / 4
buff = input_file.read(sz)
fmtstr = '{:d}f'.format(trunc(n))
inputfloats = struct.unpack(fmtstr, buff)
input_file.close()
index = 0
for key, value in keymap.items():
# ignore keys order is all that matteers
if isinstance(value, torch.Tensor):
tup = value.size()
if len(tup) == 2:
dtensor = []
for row in range(tup[0]):
trow = []
for col in range(tup[1]):
trow.append(inputfloats[index])
index += 1
dtensor.append(trow)
tensor_from_list = torch.FloatTensor(dtensor)
new_keymap[key] = tensor_from_list
else:
dtensor = []
for row in range(tup[0]):
dtensor.append(inputfloats[index])
index += 1
tensor_from_list = torch.FloatTensor(dtensor)
new_keymap[key] = tensor_from_list
self.qnetwork_local.load_state_dict(new_keymap)
def weights(self):
return self.qnetwork_local.state_dict()
def fitness(self, episode_reward):
# How we calculate the % fitness
# Episide reward is a value from about -500 to +300
# < 0 is very bad crash
# < 100 is bad crash 0%
# < 150 is failure
# < 200 is poor
# > 200 is okay, pass mark 50%
# > 220 is good
# > 240 is very good
# > 270 is excellent
# > 300 is perfect, 100%
# Assuming 100% = 300 and 100 is 0
fitness = episode_reward - 100
if fitness < 0:
fitness = 0 # Bad lowest floor of 0%, anything below is consisdered 0%
# Now divide score by 2 to get %
fitness = fitness / 2
if fitness > 100:
fitness = 100 # Highest cap to 100%, anything above is considered 100%
# Note that 50% is now considered 'okay' its a successful landing, passing mark
return fitness
def learn(self, experiences, gamma):
"""Update value parameters using given batch of experience tuples.
Params
======
experiences (Tuple[torch.Tensor]): tuple of (s, a, r, s', done) tuples
gamma (float): discount factor
"""
states, actions, rewards, next_states, dones = experiences
# Get max predicted Q values (for next states) from target model
Q_targets_next = self.qnetwork_target(next_states).detach().max(
1)[0].unsqueeze(1)
# Compute Q targets for current states
Q_targets = rewards + (gamma * Q_targets_next * (1 - dones))
# Get expected Q values from local model
Q_expected = self.qnetwork_local(states).gather(1, actions)
# Compute loss
loss = F.mse_loss(Q_expected, Q_targets)
# Minimize the loss
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
# ------------------- update target network ------------------- #
self.soft_update(self.qnetwork_local, self.qnetwork_target, TAU)
def soft_update(self, local_model, target_model, tau):
"""Soft update model parameters.
θ_target = τ*θ_local + (1 - τ)*θ_target
Params
======
local_model (PyTorch model): weights will be copied from
target_model (PyTorch model): weights will be copied to
tau (float): interpolation parameter
"""
for target_param, local_param in zip(target_model.parameters(),
local_model.parameters()):
target_param.data.copy_(tau * local_param.data +
(1.0 - tau) * target_param.data)