class QNet:
# Hyper-Parameters
gamma = 0.99 # Discount factor
initial_exploration = 10**4 # Initial exploratoin. original: 5x10^4
replay_size = 32 # Replay (batch) size
target_model_update_freq = 10**4 # Target update frequancy. original: 10^4
data_size = 10**5 # Data size of history. original: 10^6
hist_size = 1 #original: 4
def __init__(self, use_gpu, enable_controller, dim):
self.use_gpu = use_gpu
self.num_of_actions = len(enable_controller)
self.enable_controller = enable_controller
self.dim = dim
print("Initializing Q-Network...")
hidden_dim = 256
self.model = FunctionSet(l4=F.Linear(self.dim * self.hist_size,
hidden_dim,
wscale=np.sqrt(2)),
q_value=F.Linear(
hidden_dim,
self.num_of_actions,
initialW=np.zeros(
(self.num_of_actions, hidden_dim),
dtype=np.float32)))
if self.use_gpu >= 0:
self.model.to_gpu()
self.model_target = copy.deepcopy(self.model)
self.optimizer = optimizers.RMSpropGraves(lr=0.00025,
alpha=0.95,
momentum=0.95,
eps=0.0001)
self.optimizer.setup(self.model.collect_parameters())
# History Data : D=[s, a, r, s_dash, end_episode_flag]
self.d = [
np.zeros((self.data_size, self.hist_size, self.dim),
dtype=np.uint8),
np.zeros(self.data_size, dtype=np.uint8),
np.zeros((self.data_size, 1), dtype=np.int8),
np.zeros((self.data_size, self.hist_size, self.dim),
dtype=np.uint8),
np.zeros((self.data_size, 1), dtype=np.bool)
]
def forward(self, state, action, reward, state_dash, episode_end):
num_of_batch = state.shape[0]
s = Variable(state)
s_dash = Variable(state_dash)
q = self.q_func(s) # Get Q-value
# Generate Target Signals
tmp = self.q_func_target(s_dash) # Q(s',*)
if self.use_gpu >= 0:
tmp = list(map(np.max, tmp.data.get())) # max_a Q(s',a)
else:
tmp = list(map(np.max, tmp.data)) # max_a Q(s',a)
max_q_dash = np.asanyarray(tmp, dtype=np.float32)
if self.use_gpu >= 0:
target = np.asanyarray(q.data.get(), dtype=np.float32)
else:
# make new array
target = np.array(q.data, dtype=np.float32)
for i in xrange(num_of_batch):
if not episode_end[i][0]:
tmp_ = np.sign(reward[i]) + self.gamma * max_q_dash[i]
else:
tmp_ = np.sign(reward[i])
action_index = self.action_to_index(action[i])
target[i, action_index] = tmp_
# TD-error clipping
if self.use_gpu >= 0:
target = cuda.to_gpu(target)
td = Variable(target) - q # TD error
td_tmp = td.data + 1000.0 * (abs(td.data) <= 1) # Avoid zero division
td_clip = td * (abs(td.data) <= 1) + td / abs(td_tmp) * (abs(td.data) >
1)
zero_val = np.zeros((self.replay_size, self.num_of_actions),
dtype=np.float32)
if self.use_gpu >= 0:
zero_val = cuda.to_gpu(zero_val)
zero_val = Variable(zero_val)
loss = F.mean_squared_error(td_clip, zero_val)
return loss, q
def stock_experience(self, time, state, action, reward, state_dash,
episode_end_flag):
data_index = time % self.data_size
if episode_end_flag is True:
self.d[0][data_index] = state
self.d[1][data_index] = action
self.d[2][data_index] = reward
else:
self.d[0][data_index] = state
self.d[1][data_index] = action
self.d[2][data_index] = reward
self.d[3][data_index] = state_dash
self.d[4][data_index] = episode_end_flag
def experience_replay(self, time):
if self.initial_exploration < time:
# Pick up replay_size number of samples from the Data
if time < self.data_size: # during the first sweep of the History Data
replay_index = np.random.randint(0, time,
(self.replay_size, 1))
else:
replay_index = np.random.randint(0, self.data_size,
(self.replay_size, 1))
s_replay = np.ndarray(shape=(self.replay_size, self.hist_size,
self.dim),
dtype=np.float32)
a_replay = np.ndarray(shape=(self.replay_size, 1), dtype=np.uint8)
r_replay = np.ndarray(shape=(self.replay_size, 1),
dtype=np.float32)
s_dash_replay = np.ndarray(shape=(self.replay_size, self.hist_size,
self.dim),
dtype=np.float32)
episode_end_replay = np.ndarray(shape=(self.replay_size, 1),
dtype=np.bool)
for i in xrange(self.replay_size):
s_replay[i] = np.asarray(self.d[0][replay_index[i]],
dtype=np.float32)
a_replay[i] = self.d[1][replay_index[i]]
r_replay[i] = self.d[2][replay_index[i]]
s_dash_replay[i] = np.array(self.d[3][replay_index[i]],
dtype=np.float32)
episode_end_replay[i] = self.d[4][replay_index[i]]
if self.use_gpu >= 0:
s_replay = cuda.to_gpu(s_replay)
s_dash_replay = cuda.to_gpu(s_dash_replay)
# Gradient-based update
self.optimizer.zero_grads()
loss, _ = self.forward(s_replay, a_replay, r_replay, s_dash_replay,
episode_end_replay)
loss.backward()
self.optimizer.update()
def q_func(self, state):
h4 = F.relu(self.model.l4(state))
q = self.model.q_value(h4 / 255.0)
return q
def q_func_target(self, state):
h4 = F.relu(self.model_target.l4(state / 255.0))
q = self.model_target.q_value(h4)
return q
def e_greedy(self, state, epsilon):
s = Variable(state)
q = self.q_func(s)
q = q.data
if np.random.rand() < epsilon:
index_action = np.random.randint(0, self.num_of_actions)
print(" Random"),
else:
if self.use_gpu >= 0:
index_action = np.argmax(q.get())
else:
index_action = np.argmax(q)
print("#Greedy"),
return self.index_to_action(index_action), q
def target_model_update(self):
self.model_target = copy.deepcopy(self.model)
def index_to_action(self, index_of_action):
return self.enable_controller[index_of_action]
def action_to_index(self, action):
return self.enable_controller.index(action)
class DN_class:
# Hyper-Parameters
gamma = 0.99 # Discount factor
initial_exploration = 100#10**4 # Initial exploratoin. original: 5x10^4
replay_size = 32 # Replay (batch) size
target_model_update_freq = 10**4 # Target update frequancy. original: 10^4
data_size = 10**5 # Data size of history. original: 10^6
def __init__(self, enable_controller=[0, 1, 3, 4]):
self.num_of_actions = len(enable_controller)
self.enable_controller = enable_controller # Default setting : "Breakout"
print "Initializing DN..."
# Initialization of Chainer 1.1.0 or older.
# print "CUDA init"
# cuda.init()
print "Model Building"
self.model = FunctionSet(
l1=F.Convolution2D(4, 32, ksize=8, stride=4, nobias=False, wscale=np.sqrt(2)),
l2=F.Convolution2D(32, 64, ksize=4, stride=2, nobias=False, wscale=np.sqrt(2)),
l3=F.Convolution2D(64, 64, ksize=3, stride=1, nobias=False, wscale=np.sqrt(2)),
l4=F.Linear(3136, 256, wscale=np.sqrt(2)),
l5=F.Linear(3136, 256, wscale=np.sqrt(2)),
l6=F.Linear(256, 1, initialW=np.zeros((1, 256), dtype=np.float32)),
l7=F.Linear(256, self.num_of_actions, initialW=np.zeros((self.num_of_actions, 256),
dtype=np.float32)),
q_value=DN_out.DN_out(1, self.num_of_actions, self.num_of_actions, nobias = True)
).to_gpu()
if args.resumemodel:
# load saved model
serializers.load_npz(args.resumemodel, self.model)
print "load model from resume.model"
self.model_target = copy.deepcopy(self.model)
print "Initizlizing Optimizer"
self.optimizer = optimizers.RMSpropGraves(lr=0.00025, alpha=0.95, momentum=0.95, eps=0.0001)
self.optimizer.setup(self.model.collect_parameters())
# History Data : D=[s, a, r, s_dash, end_episode_flag]
if args.resumeD1 and args.resumeD2:
# load saved D1 and D2
npz_tmp1 = np.load(args.resumeD1)
print "finished loading half of D data"
npz_tmp2 = np.load(args.resumeD2)
self.D = [npz_tmp1['D0'],
npz_tmp1['D1'],
npz_tmp1['D2'],
npz_tmp2['D3'],
npz_tmp2['D4']]
npz_tmp1.close()
npz_tmp2.close()
print "loaded stored all D data"
else:
self.D = [np.zeros((self.data_size, 4, 84, 84), dtype=np.uint8),
np.zeros(self.data_size, dtype=np.uint8),
np.zeros((self.data_size, 1), dtype=np.int8),
np.zeros((self.data_size, 4, 84, 84), dtype=np.uint8),
np.zeros((self.data_size, 1), dtype=np.bool)]
print "initialized D data"
def forward(self, state, action, Reward, state_dash, episode_end):
num_of_batch = state.shape[0]
s = Variable(state)
s_dash = Variable(state_dash)
Q = self.Q_func(s) # Get Q-value
# Generate Target Signals
tmp2 = self.Q_func(s_dash)
tmp2 = list(map(np.argmax, tmp2.data.get())) # argmaxQ(s',a)
tmp = self.Q_func_target(s_dash) # Q'(s',*)
tmp = list(tmp.data.get())
# select Q'(s',*) due to argmaxQ(s',a)
res1 = []
for i in range(num_of_batch):
res1.append(tmp[i][tmp2[i]])
#max_Q_dash = np.asanyarray(tmp, dtype=np.float32)
max_Q_dash = np.asanyarray(res1, dtype=np.float32)
target = np.asanyarray(Q.data.get(), dtype=np.float32)
for i in xrange(num_of_batch):
if not episode_end[i][0]:
tmp_ = np.sign(Reward[i]) + self.gamma * max_Q_dash[i]
else:
tmp_ = np.sign(Reward[i])
action_index = self.action_to_index(action[i])
target[i, action_index] = tmp_
# TD-error clipping
td = Variable(cuda.to_gpu(target)) - Q # TD error
td_tmp = td.data + 1000.0 * (abs(td.data) <= 1) # Avoid zero division
td_clip = td * (abs(td.data) <= 1) + td/abs(td_tmp) * (abs(td.data) > 1)
zero_val = Variable(cuda.to_gpu(np.zeros((self.replay_size, self.num_of_actions), dtype=np.float32)))
loss = F.mean_squared_error(td_clip, zero_val)
return loss, Q
def stockExperience(self, time,
state, action, reward, state_dash,
episode_end_flag):
data_index = time % self.data_size
if episode_end_flag is True:
self.D[0][data_index] = state
self.D[1][data_index] = action
self.D[2][data_index] = reward
else:
self.D[0][data_index] = state
self.D[1][data_index] = action
self.D[2][data_index] = reward
self.D[3][data_index] = state_dash
self.D[4][data_index] = episode_end_flag
def experienceReplay(self, time):
if self.initial_exploration < time:
# Pick up replay_size number of samples from the Data
if time < self.data_size: # during the first sweep of the History Data
replay_index = np.random.randint(0, time, (self.replay_size, 1))
else:
replay_index = np.random.randint(0, self.data_size, (self.replay_size, 1))
s_replay = np.ndarray(shape=(self.replay_size, 4, 84, 84), dtype=np.float32)
a_replay = np.ndarray(shape=(self.replay_size, 1), dtype=np.uint8)
r_replay = np.ndarray(shape=(self.replay_size, 1), dtype=np.float32)
s_dash_replay = np.ndarray(shape=(self.replay_size, 4, 84, 84), dtype=np.float32)
episode_end_replay = np.ndarray(shape=(self.replay_size, 1), dtype=np.bool)
for i in xrange(self.replay_size):
s_replay[i] = np.asarray(self.D[0][replay_index[i]], dtype=np.float32)
a_replay[i] = self.D[1][replay_index[i]]
r_replay[i] = self.D[2][replay_index[i]]
s_dash_replay[i] = np.array(self.D[3][replay_index[i]], dtype=np.float32)
episode_end_replay[i] = self.D[4][replay_index[i]]
s_replay = cuda.to_gpu(s_replay)
s_dash_replay = cuda.to_gpu(s_dash_replay)
# Gradient-based update
self.optimizer.zero_grads()
loss, _ = self.forward(s_replay, a_replay, r_replay, s_dash_replay, episode_end_replay)
loss.backward()
self.optimizer.update()
def Q_func(self, state):
print 'now Q_func is implemented'
h1 = F.relu(self.model.l1(state / 254.0)) # scale inputs in [0.0 1.0]
h2 = F.relu(self.model.l2(h1))
h3 = F.relu(self.model.l3(h2))
h4 = F.relu(self.model.l4(h3)) # left side connected with s value
h5 = F.relu(self.model.l5(h3)) # right side connected with A value
h6 = self.model.l6(h4) # s value
h7 = self.model.l7(h5) # A value
Q = self.model.q_value(h6, h7) # Q value
return Q
def Q_func_target(self, state):
print 'now Q_func_target is implemented'
h1 = F.relu(self.model_target.l1(state / 254.0)) # scale inputs in [0.0 1.0]
h2 = F.relu(self.model_target.l2(h1))
h3 = F.relu(self.model_target.l3(h2))
h4 = F.relu(self.model_target.l4(h3)) # left side connected with s value
h5 = F.relu(self.model_target.l5(h3)) # right side connected with A value
h6 = self.model_target.l6(h4) # s value
h7 = self.model_target.l7(h5) # A value
Q = self.model_target.q_value(h6, h7) # Q value
return Q
def e_greedy(self, state, epsilon):
s = Variable(state)
Q = self.Q_func(s)
Q = Q.data
if np.random.rand() < epsilon:
index_action = np.random.randint(0, self.num_of_actions)
print "RANDOM"
else:
index_action = np.argmax(Q.get())
print "GREEDY"
return self.index_to_action(index_action), Q
def target_model_update(self):
self.model_target = copy.deepcopy(self.model)
def index_to_action(self, index_of_action):
return self.enable_controller[index_of_action]
def action_to_index(self, action):
return self.enable_controller.index(action)
class QNet:
# Hyper-Parameters
gamma = 0.99 # Discount factor
initial_exploration = 10**3 # Initial exploratoin. original: 5x10^4
replay_size = 32 # Replay (batch) size
target_model_update_freq = 10**4 # Target update frequancy. original: 10^4
data_size = 10**5 # Data size of history. original: 10^6
hist_size = 1 #original: 4
def __init__(self, use_gpu, enable_controller, dim):
self.use_gpu = use_gpu
self.num_of_actions = len(enable_controller)
self.enable_controller = enable_controller
self.dim = dim
print("Initializing Q-Network...")
hidden_dim = 256
self.model = FunctionSet(
l4=F.Linear(self.dim*self.hist_size, hidden_dim, wscale=np.sqrt(2)),
q_value=F.Linear(hidden_dim, self.num_of_actions,
initialW=np.zeros((self.num_of_actions, hidden_dim),
dtype=np.float32))
)
if self.use_gpu >= 0:
self.model.to_gpu()
self.model_target = copy.deepcopy(self.model)
self.optimizer = optimizers.RMSpropGraves(lr=0.00025, alpha=0.95, momentum=0.95, eps=0.0001)
self.optimizer.setup(self.model.collect_parameters())
# History Data : D=[s, a, r, s_dash, end_episode_flag]
self.d = [np.zeros((self.data_size, self.hist_size, self.dim), dtype=np.uint8),
np.zeros(self.data_size, dtype=np.uint8),
np.zeros((self.data_size, 1), dtype=np.int8),
np.zeros((self.data_size, self.hist_size, self.dim), dtype=np.uint8),
np.zeros((self.data_size, 1), dtype=np.bool)]
def forward(self, state, action, reward, state_dash, episode_end):
num_of_batch = state.shape[0]
s = Variable(state)
s_dash = Variable(state_dash)
q = self.q_func(s) # Get Q-value
# Generate Target Signals
tmp = self.q_func_target(s_dash) # Q(s',*)
if self.use_gpu >= 0:
tmp = list(map(np.max, tmp.data.get())) # max_a Q(s',a)
else:
tmp = list(map(np.max, tmp.data)) # max_a Q(s',a)
max_q_dash = np.asanyarray(tmp, dtype=np.float32)
if self.use_gpu >= 0:
target = np.asanyarray(q.data.get(), dtype=np.float32)
else:
# make new array
target = np.array(q.data, dtype=np.float32)
for i in xrange(num_of_batch):
if not episode_end[i][0]:
tmp_ = reward[i] + self.gamma * max_q_dash[i]
else:
tmp_ = reward[i]
action_index = self.action_to_index(action[i])
target[i, action_index] = tmp_
# TD-error clipping
if self.use_gpu >= 0:
target = cuda.to_gpu(target)
td = Variable(target) - q # TD error
td_tmp = td.data + 1000.0 * (abs(td.data) <= 1) # Avoid zero division
td_clip = td * (abs(td.data) <= 1) + td/abs(td_tmp) * (abs(td.data) > 1)
zero_val = np.zeros((self.replay_size, self.num_of_actions), dtype=np.float32)
if self.use_gpu >= 0:
zero_val = cuda.to_gpu(zero_val)
zero_val = Variable(zero_val)
loss = F.mean_squared_error(td_clip, zero_val)
return loss, q
def stock_experience(self, time,
state, action, reward, state_dash,
episode_end_flag):
data_index = time % self.data_size
if episode_end_flag is True:
self.d[0][data_index] = state
self.d[1][data_index] = action
self.d[2][data_index] = reward
else:
self.d[0][data_index] = state
self.d[1][data_index] = action
self.d[2][data_index] = reward
self.d[3][data_index] = state_dash
self.d[4][data_index] = episode_end_flag
def experience_replay(self, time):
if self.initial_exploration < time:
# Pick up replay_size number of samples from the Data
if time < self.data_size: # during the first sweep of the History Data
replay_index = np.random.randint(0, time, (self.replay_size, 1))
else:
replay_index = np.random.randint(0, self.data_size, (self.replay_size, 1))
s_replay = np.ndarray(shape=(self.replay_size, self.hist_size, self.dim), dtype=np.float32)
a_replay = np.ndarray(shape=(self.replay_size, 1), dtype=np.uint8)
r_replay = np.ndarray(shape=(self.replay_size, 1), dtype=np.float32)
s_dash_replay = np.ndarray(shape=(self.replay_size, self.hist_size, self.dim), dtype=np.float32)
episode_end_replay = np.ndarray(shape=(self.replay_size, 1), dtype=np.bool)
for i in xrange(self.replay_size):
s_replay[i] = np.asarray(self.d[0][replay_index[i]], dtype=np.float32)
a_replay[i] = self.d[1][replay_index[i]]
r_replay[i] = self.d[2][replay_index[i]]
s_dash_replay[i] = np.array(self.d[3][replay_index[i]], dtype=np.float32)
episode_end_replay[i] = self.d[4][replay_index[i]]
if self.use_gpu >= 0:
s_replay = cuda.to_gpu(s_replay)
s_dash_replay = cuda.to_gpu(s_dash_replay)
# Gradient-based update
self.optimizer.zero_grads()
loss, _ = self.forward(s_replay, a_replay, r_replay, s_dash_replay, episode_end_replay)
loss.backward()
self.optimizer.update()
def q_func(self, state):
h4 = F.relu(self.model.l4(state / 255.0))
q = self.model.q_value(h4)
return q
def q_func_target(self, state):
h4 = F.relu(self.model_target.l4(state / 255.0))
q = self.model_target.q_value(h4)
return q
def e_greedy(self, state, epsilon):
s = Variable(state)
q = self.q_func(s)
q = q.data
if np.random.rand() < epsilon:
index_action = np.random.randint(0, self.num_of_actions)
print(" Random"),
else:
if self.use_gpu >= 0:
index_action = np.argmax(q.get())
else:
index_action = np.argmax(q)
print("#Greedy"),
return self.index_to_action(index_action), q
def target_model_update(self):
self.model_target = copy.deepcopy(self.model)
def index_to_action(self, index_of_action):
return self.enable_controller[index_of_action]
def action_to_index(self, action):
return self.enable_controller.index(action)
class DQN_class:
# Hyper-Parameters
gamma = 0.99 # Discount factor
initial_exploration = 10**4 # Initial exploratoin. original: 5x10^4
replay_size = 32 # Replay (batch) size
target_model_update_freq = 10**4 # Target update frequancy. original: 10^4
data_size = 10**5 # Data size of history. original: 10^6
def __init__(self, enable_controller=[0, 3, 4]):
self.num_of_actions = len(enable_controller)
self.enable_controller = enable_controller # Default setting : "Pong"
print "Initializing DQN..."
print "CUDA init"
cuda.init()
print "Model Building"
self.model = FunctionSet(
l1=F.Convolution2D(4, 16, ksize=8, stride=4, wscale=np.sqrt(2)),
l2=F.Convolution2D(16, 32, ksize=4, stride=2, wscale=np.sqrt(2)),
l3=F.Linear(2592, 256),
q_value=F.Linear(256,
self.num_of_actions,
initialW=np.zeros((self.num_of_actions, 256),
dtype=np.float32))).to_gpu()
print "Initizlizing Optimizer"
self.optimizer = optimizers.RMSpropGraves(lr=0.0002,
alpha=0.3,
momentum=0.2)
self.optimizer.setup(self.model.collect_parameters())
# History Data : D=[s, a, r, s_dash, end_episode_flag]
self.D = [
np.zeros((self.data_size, 4, 84, 84), dtype=np.uint8),
np.zeros(self.data_size, dtype=np.uint8),
np.zeros((self.data_size, 1), dtype=np.int8),
np.zeros((self.data_size, 4, 84, 84), dtype=np.uint8),
np.zeros((self.data_size, 1), dtype=np.bool)
]
def forward(self, state, action, Reward, state_dash, episode_end):
num_of_batch = state.shape[0]
s = Variable(state)
s_dash = Variable(state_dash)
Q = self.Q_func(s) # Get Q-value
# Generate Target Signals
max_Q_dash_ = self.Q_func(s_dash)
tmp = list(map(np.max, max_Q_dash_.data.get()))
max_Q_dash = np.asanyarray(tmp, dtype=np.float32)
target = np.asanyarray(Q.data.get(), dtype=np.float32)
for i in xrange(num_of_batch):
if not episode_end[i][0]:
tmp_ = np.sign(Reward[i]) + self.gamma * max_Q_dash[i]
else:
tmp_ = np.sign(Reward[i])
target[i, self.action_to_index(action[i])] = tmp_
loss = F.mean_squared_error(Variable(cuda.to_gpu(target)), Q)
return loss, Q
def stockExperience(self, time, state, action, reward, state_dash,
episode_end_flag):
data_index = time % self.data_size
if episode_end_flag is True:
self.D[0][data_index] = state
self.D[1][data_index] = action
self.D[2][data_index] = reward
else:
self.D[0][data_index] = state
self.D[1][data_index] = action
self.D[2][data_index] = reward
self.D[3][data_index] = state_dash
self.D[4][data_index] = episode_end_flag
def experienceReplay(self, time):
if self.initial_exploration < time:
# Pick up replay_size number of samples from the Data
if time < self.data_size: # during the first sweep of the History Data
replay_index = np.random.randint(0, time,
(self.replay_size, 1))
else:
replay_index = np.random.randint(0, self.data_size,
(self.replay_size, 1))
s_replay = np.ndarray(shape=(self.replay_size, 4, 84, 84),
dtype=np.float32)
a_replay = np.ndarray(shape=(self.replay_size, 1), dtype=np.uint8)
r_replay = np.ndarray(shape=(self.replay_size, 1),
dtype=np.float32)
s_dash_replay = np.ndarray(shape=(self.replay_size, 4, 84, 84),
dtype=np.float32)
episode_end_replay = np.ndarray(shape=(self.replay_size, 1),
dtype=np.bool)
for i in xrange(self.replay_size):
s_replay[i] = np.asarray(self.D[0][replay_index[i]],
dtype=np.float32)
a_replay[i] = self.D[1][replay_index[i]]
r_replay[i] = self.D[2][replay_index[i]]
s_dash_replay[i] = np.array(self.D[3][replay_index[i]],
dtype=np.float32)
episode_end_replay[i] = self.D[4][replay_index[i]]
s_replay = cuda.to_gpu(s_replay)
s_dash_replay = cuda.to_gpu(s_dash_replay)
# Gradient-based update
self.optimizer.zero_grads()
loss, _ = self.forward(s_replay, a_replay, r_replay, s_dash_replay,
episode_end_replay)
loss.backward()
self.optimizer.update()
def Q_func(self, state):
h1 = F.relu(self.model.l1(state / 254.0)) # scale inputs in [0.0, 1.0]
h2 = F.relu(self.model.l2(h1))
h3 = F.relu(self.model.l3(h2))
Q = self.model.q_value(h3)
return Q
def e_greedy(self, state, epsilon):
s = Variable(state)
Q = self.Q_func(s)
Q = Q.data
if np.random.rand() < epsilon:
index_action = np.random.randint(0, self.num_of_actions)
print "RANDOM"
else:
index_action = np.argmax(Q.get())
print "GREEDY"
return self.index_to_action(index_action), Q
def index_to_action(self, index_of_action):
return self.enable_controller[index_of_action]
def action_to_index(self, action):
return self.enable_controller.index(action)
class DQN_class:
# Hyper-Parameters
gamma = 0.99 # Discount factor
def __init__(self, enable_controller=[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17]):
self.num_of_actions = len(enable_controller)
self.enable_controller = enable_controller # Default setting : "Pong"
print "Initializing DQN..."
# Initialization of Chainer 1.1.0 or older.
# print "CUDA init"
# cuda.init()
print "Model Building"
w = math.sqrt(2) # MSRA scaling
self.model = FunctionSet(
conv1=F.Convolution2D(3, 64, 7, wscale=w, stride=2, pad=3),
conv2_1_1=F.Convolution2D(64, 64, 1, wscale=w, stride=1),
conv2_1_2=F.Convolution2D(64, 64, 3, wscale=w, stride=1, pad=1),
conv2_1_3=F.Convolution2D(64, 256, 1, wscale=w, stride=1),
conv2_1_ex=F.Convolution2D(64, 256, 1, wscale=w, stride=1),
conv2_2_1=F.Convolution2D(256, 64, 1, wscale=w, stride=1),
conv2_2_2=F.Convolution2D(64, 64, 3, wscale=w, stride=1, pad=1),
conv2_2_3=F.Convolution2D(64, 256, 1, wscale=w, stride=1),
conv2_3_1=F.Convolution2D(256, 64, 1, wscale=w, stride=1),
conv2_3_2=F.Convolution2D(64, 64, 3, wscale=w, stride=1, pad=1),
conv2_3_3=F.Convolution2D(64, 256, 1, wscale=w, stride=1),
conv3_1_1=F.Convolution2D(256, 128, 1, wscale=w, stride=2),
conv3_1_2=F.Convolution2D(128, 128, 3, wscale=w, stride=1, pad=1),
conv3_1_3=F.Convolution2D(128, 512, 1, wscale=w, stride=1),
conv3_1_ex=F.Convolution2D(256, 512, 1, wscale=w, stride=2),
conv3_2_1=F.Convolution2D(512, 128, 1, wscale=w, stride=1),
conv3_2_2=F.Convolution2D(128, 128, 3, wscale=w, stride=1, pad=1),
conv3_2_3=F.Convolution2D(128, 512, 1, wscale=w, stride=1),
conv3_3_1=F.Convolution2D(512, 128, 1, wscale=w, stride=1),
conv3_3_2=F.Convolution2D(128, 128, 3, wscale=w, stride=1, pad=1),
conv3_3_3=F.Convolution2D(128, 512, 1, wscale=w, stride=1),
conv3_4_1=F.Convolution2D(512, 128, 1, wscale=w, stride=1),
conv3_4_2=F.Convolution2D(128, 128, 3, wscale=w, stride=1, pad=1),
conv3_4_3=F.Convolution2D(128, 512, 1, wscale=w, stride=1),
conv3_5_1=F.Convolution2D(512, 128, 1, wscale=w, stride=1),
conv3_5_2=F.Convolution2D(128, 128, 3, wscale=w, stride=1, pad=1),
conv3_5_3=F.Convolution2D(128, 512, 1, wscale=w, stride=1),
conv3_6_1=F.Convolution2D(512, 128, 1, wscale=w, stride=1),
conv3_6_2=F.Convolution2D(128, 128, 3, wscale=w, stride=1, pad=1),
conv3_6_3=F.Convolution2D(128, 512, 1, wscale=w, stride=1),
conv3_7_1=F.Convolution2D(512, 128, 1, wscale=w, stride=1),
conv3_7_2=F.Convolution2D(128, 128, 3, wscale=w, stride=1, pad=1),
conv3_7_3=F.Convolution2D(128, 512, 1, wscale=w, stride=1),
conv3_8_1=F.Convolution2D(512, 128, 1, wscale=w, stride=1),
conv3_8_2=F.Convolution2D(128, 128, 3, wscale=w, stride=1, pad=1),
conv3_8_3=F.Convolution2D(128, 512, 1, wscale=w, stride=1),
conv4_1_1=F.Convolution2D(512, 256, 1, wscale=w, stride=2),
conv4_1_2=F.Convolution2D(256, 256, 3, wscale=w, stride=1, pad=1),
conv4_1_3=F.Convolution2D(256, 1024, 1, wscale=w, stride=1),
conv4_1_ex=F.Convolution2D(512, 1024, 1, wscale=w, stride=2),
conv4_2_1=F.Convolution2D(1024, 256, 1, wscale=w, stride=1),
conv4_2_2=F.Convolution2D(256, 256, 3, wscale=w, stride=1, pad=1),
conv4_2_3=F.Convolution2D(256, 1024, 1, wscale=w, stride=1),
conv4_3_1=F.Convolution2D(1024, 256, 1, wscale=w, stride=1),
conv4_3_2=F.Convolution2D(256, 256, 3, wscale=w, stride=1, pad=1),
conv4_3_3=F.Convolution2D(256, 1024, 1, wscale=w, stride=1),
conv4_4_1=F.Convolution2D(1024, 256, 1, wscale=w, stride=1),
conv4_4_2=F.Convolution2D(256, 256, 3, wscale=w, stride=1, pad=1),
conv4_4_3=F.Convolution2D(256, 1024, 1, wscale=w, stride=1),
conv4_5_1=F.Convolution2D(1024, 256, 1, wscale=w, stride=1),
conv4_5_2=F.Convolution2D(256, 256, 3, wscale=w, stride=1, pad=1),
conv4_5_3=F.Convolution2D(256, 1024, 1, wscale=w, stride=1),
conv4_6_1=F.Convolution2D(1024, 256, 1, wscale=w, stride=1),
conv4_6_2=F.Convolution2D(256, 256, 3, wscale=w, stride=1, pad=1),
conv4_6_3=F.Convolution2D(256, 1024, 1, wscale=w, stride=1),
conv4_7_1=F.Convolution2D(1024, 256, 1, wscale=w, stride=1),
conv4_7_2=F.Convolution2D(256, 256, 3, wscale=w, stride=1, pad=1),
conv4_7_3=F.Convolution2D(256, 1024, 1, wscale=w, stride=1),
conv4_8_1=F.Convolution2D(1024, 256, 1, wscale=w, stride=1),
conv4_8_2=F.Convolution2D(256, 256, 3, wscale=w, stride=1, pad=1),
conv4_8_3=F.Convolution2D(256, 1024, 1, wscale=w, stride=1),
conv4_9_1=F.Convolution2D(1024, 256, 1, wscale=w, stride=1),
conv4_9_2=F.Convolution2D(256, 256, 3, wscale=w, stride=1, pad=1),
conv4_9_3=F.Convolution2D(256, 1024, 1, wscale=w, stride=1),
conv4_10_1=F.Convolution2D(1024, 256, 1, wscale=w, stride=1),
conv4_10_2=F.Convolution2D(256, 256, 3, wscale=w, stride=1, pad=1),
conv4_10_3=F.Convolution2D(256, 1024, 1, wscale=w, stride=1),
conv4_11_1=F.Convolution2D(1024, 256, 1, wscale=w, stride=1),
conv4_11_2=F.Convolution2D(256, 256, 3, wscale=w, stride=1, pad=1),
conv4_11_3=F.Convolution2D(256, 1024, 1, wscale=w, stride=1),
conv4_12_1=F.Convolution2D(1024, 256, 1, wscale=w, stride=1),
conv4_12_2=F.Convolution2D(256, 256, 3, wscale=w, stride=1, pad=1),
conv4_12_3=F.Convolution2D(256, 1024, 1, wscale=w, stride=1),
conv4_13_1=F.Convolution2D(1024, 256, 1, wscale=w, stride=1),
conv4_13_2=F.Convolution2D(256, 256, 3, wscale=w, stride=1, pad=1),
conv4_13_3=F.Convolution2D(256, 1024, 1, wscale=w, stride=1),
conv4_14_1=F.Convolution2D(1024, 256, 1, wscale=w, stride=1),
conv4_14_2=F.Convolution2D(256, 256, 3, wscale=w, stride=1, pad=1),
conv4_14_3=F.Convolution2D(256, 1024, 1, wscale=w, stride=1),
conv4_15_1=F.Convolution2D(1024, 256, 1, wscale=w, stride=1),
conv4_15_2=F.Convolution2D(256, 256, 3, wscale=w, stride=1, pad=1),
conv4_15_3=F.Convolution2D(256, 1024, 1, wscale=w, stride=1),
conv4_16_1=F.Convolution2D(1024, 256, 1, wscale=w, stride=1),
conv4_16_2=F.Convolution2D(256, 256, 3, wscale=w, stride=1, pad=1),
conv4_16_3=F.Convolution2D(256, 1024, 1, wscale=w, stride=1),
conv4_17_1=F.Convolution2D(1024, 256, 1, wscale=w, stride=1),
conv4_17_2=F.Convolution2D(256, 256, 3, wscale=w, stride=1, pad=1),
conv4_17_3=F.Convolution2D(256, 1024, 1, wscale=w, stride=1),
conv4_18_1=F.Convolution2D(1024, 256, 1, wscale=w, stride=1),
conv4_18_2=F.Convolution2D(256, 256, 3, wscale=w, stride=1, pad=1),
conv4_18_3=F.Convolution2D(256, 1024, 1, wscale=w, stride=1),
conv4_19_1=F.Convolution2D(1024, 256, 1, wscale=w, stride=1),
conv4_19_2=F.Convolution2D(256, 256, 3, wscale=w, stride=1, pad=1),
conv4_19_3=F.Convolution2D(256, 1024, 1, wscale=w, stride=1),
conv4_20_1=F.Convolution2D(1024, 256, 1, wscale=w, stride=1),
conv4_20_2=F.Convolution2D(256, 256, 3, wscale=w, stride=1, pad=1),
conv4_20_3=F.Convolution2D(256, 1024, 1, wscale=w, stride=1),
conv4_21_1=F.Convolution2D(1024, 256, 1, wscale=w, stride=1),
conv4_21_2=F.Convolution2D(256, 256, 3, wscale=w, stride=1, pad=1),
conv4_21_3=F.Convolution2D(256, 1024, 1, wscale=w, stride=1),
conv4_22_1=F.Convolution2D(1024, 256, 1, wscale=w, stride=1),
conv4_22_2=F.Convolution2D(256, 256, 3, wscale=w, stride=1, pad=1),
conv4_22_3=F.Convolution2D(256, 1024, 1, wscale=w, stride=1),
conv4_23_1=F.Convolution2D(1024, 256, 1, wscale=w, stride=1),
conv4_23_2=F.Convolution2D(256, 256, 3, wscale=w, stride=1, pad=1),
conv4_23_3=F.Convolution2D(256, 1024, 1, wscale=w, stride=1),
conv4_24_1=F.Convolution2D(1024, 256, 1, wscale=w, stride=1),
conv4_24_2=F.Convolution2D(256, 256, 3, wscale=w, stride=1, pad=1),
conv4_24_3=F.Convolution2D(256, 1024, 1, wscale=w, stride=1),
conv4_25_1=F.Convolution2D(1024, 256, 1, wscale=w, stride=1),
conv4_25_2=F.Convolution2D(256, 256, 3, wscale=w, stride=1, pad=1),
conv4_25_3=F.Convolution2D(256, 1024, 1, wscale=w, stride=1),
conv4_26_1=F.Convolution2D(1024, 256, 1, wscale=w, stride=1),
conv4_26_2=F.Convolution2D(256, 256, 3, wscale=w, stride=1, pad=1),
conv4_26_3=F.Convolution2D(256, 1024, 1, wscale=w, stride=1),
conv4_27_1=F.Convolution2D(1024, 256, 1, wscale=w, stride=1),
conv4_27_2=F.Convolution2D(256, 256, 3, wscale=w, stride=1, pad=1),
conv4_27_3=F.Convolution2D(256, 1024, 1, wscale=w, stride=1),
conv4_28_1=F.Convolution2D(1024, 256, 1, wscale=w, stride=1),
conv4_28_2=F.Convolution2D(256, 256, 3, wscale=w, stride=1, pad=1),
conv4_28_3=F.Convolution2D(256, 1024, 1, wscale=w, stride=1),
conv4_29_1=F.Convolution2D(1024, 256, 1, wscale=w, stride=1),
conv4_29_2=F.Convolution2D(256, 256, 3, wscale=w, stride=1, pad=1),
conv4_29_3=F.Convolution2D(256, 1024, 1, wscale=w, stride=1),
conv4_30_1=F.Convolution2D(1024, 256, 1, wscale=w, stride=1),
conv4_30_2=F.Convolution2D(256, 256, 3, wscale=w, stride=1, pad=1),
conv4_30_3=F.Convolution2D(256, 1024, 1, wscale=w, stride=1),
conv4_31_1=F.Convolution2D(1024, 256, 1, wscale=w, stride=1),
conv4_31_2=F.Convolution2D(256, 256, 3, wscale=w, stride=1, pad=1),
conv4_31_3=F.Convolution2D(256, 1024, 1, wscale=w, stride=1),
conv4_32_1=F.Convolution2D(1024, 256, 1, wscale=w, stride=1),
conv4_32_2=F.Convolution2D(256, 256, 3, wscale=w, stride=1, pad=1),
conv4_32_3=F.Convolution2D(256, 1024, 1, wscale=w, stride=1),
conv4_33_1=F.Convolution2D(1024, 256, 1, wscale=w, stride=1),
conv4_33_2=F.Convolution2D(256, 256, 3, wscale=w, stride=1, pad=1),
conv4_33_3=F.Convolution2D(256, 1024, 1, wscale=w, stride=1),
conv4_34_1=F.Convolution2D(1024, 256, 1, wscale=w, stride=1),
conv4_34_2=F.Convolution2D(256, 256, 3, wscale=w, stride=1, pad=1),
conv4_34_3=F.Convolution2D(256, 1024, 1, wscale=w, stride=1),
conv4_35_1=F.Convolution2D(1024, 256, 1, wscale=w, stride=1),
conv4_35_2=F.Convolution2D(256, 256, 3, wscale=w, stride=1, pad=1),
conv4_35_3=F.Convolution2D(256, 1024, 1, wscale=w, stride=1),
conv4_36_1=F.Convolution2D(1024, 256, 1, wscale=w, stride=1),
conv4_36_2=F.Convolution2D(256, 256, 3, wscale=w, stride=1, pad=1),
conv4_36_3=F.Convolution2D(256, 1024, 1, wscale=w, stride=1),
conv5_1_1=F.Convolution2D(1024, 512, 1, wscale=w, stride=2),
conv5_1_2=F.Convolution2D(512, 512, 3, wscale=w, stride=1, pad=1),
conv5_1_3=F.Convolution2D(512, 2048, 1, wscale=w, stride=1),
conv5_1_ex=F.Convolution2D(1024, 2048, 1, wscale=w, stride=2),
conv5_2_1=F.Convolution2D(2048, 512, 1, wscale=w, stride=1),
conv5_2_2=F.Convolution2D(512, 512, 3, wscale=w, stride=1, pad=1),
conv5_2_3=F.Convolution2D(512, 2048, 1, wscale=w, stride=1),
conv5_3_1=F.Convolution2D(2048, 512, 1, wscale=w, stride=1),
conv5_3_2=F.Convolution2D(512, 512, 3, wscale=w, stride=1, pad=1),
conv5_3_3=F.Convolution2D(512, 2048, 1, wscale=w, stride=1),
q_value=F.Linear(2048, self.num_of_actions,
initialW=np.zeros((self.num_of_actions, 2048),
dtype=np.float32))
)
self.model_target = copy.deepcopy(self.model)
print "Initizlizing Optimizer"
self.optimizer = optimizers.Adam()
self.optimizer.setup(self.model.collect_parameters())
def forward(self, state, action, Reward, state_dash, episode_end):
num_of_batch = state.shape[0]
s = Variable(state)
s_dash = Variable(state_dash)
Q = self.Q_func(s) # Get Q-value
# Generate Target Signals
tmp = self.Q_func_target(s_dash) # Q(s',*)
tmp = list(map(np.max, tmp.data.get())) # max_a Q(s',a)
max_Q_dash = np.asanyarray(tmp, dtype=np.float32)
target = np.asanyarray(Q.data.get(), dtype=np.float32)
for i in xrange(num_of_batch):
if not episode_end[i][0]:
tmp_ = np.sign(Reward[i]) + self.gamma * max_Q_dash[i]
else:
tmp_ = np.sign(Reward[i])
action_index = self.action_to_index(action[i])
target[i, action_index] = tmp_
# TD-error clipping
td = Variable(cuda.to_gpu(target)) - Q # TD error
td_tmp = td.data + 1000.0 * (abs(td.data) <= 1) # Avoid zero division
td_clip = td * (abs(td.data) <= 1) + td/abs(td_tmp) * (abs(td.data) > 1)
zero_val = Variable(cuda.to_gpu(np.zeros((num_of_batch, self.num_of_actions), dtype=np.float32)))
loss = F.mean_squared_error(td_clip, zero_val)
return loss, Q
def Q_func(self, state):
h = F.relu(self.model.conv1(state))
h = F.max_pooling_2d(h, 3, stride=2)
h_rem = self.model.conv2_1_ex(h)
h = F.relu(self.model.conv2_1_1(h))
h = F.relu(self.model.conv2_1_2(h))
h = self.model.conv2_1_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model.conv2_2_1(h))
h = F.relu(self.model.conv2_2_2(h))
h = self.model.conv2_2_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model.conv2_3_1(h))
h = F.relu(self.model.conv2_3_2(h))
h = self.model.conv2_3_3(h)
h = F.relu(h + h_rem)
h_rem = self.model.conv3_1_ex(h)
h = F.relu(self.model.conv3_1_1(h))
h = F.relu(self.model.conv3_1_2(h))
h = self.model.conv3_1_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model.conv3_2_1(h))
h = F.relu(self.model.conv3_2_2(h))
h = self.model.conv3_2_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model.conv3_3_1(h))
h = F.relu(self.model.conv3_3_2(h))
h = self.model.conv3_3_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model.conv3_4_1(h))
h = F.relu(self.model.conv3_4_2(h))
h = self.model.conv3_4_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model.conv3_5_1(h))
h = F.relu(self.model.conv3_5_2(h))
h = self.model.conv3_5_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model.conv3_6_1(h))
h = F.relu(self.model.conv3_6_2(h))
h = self.model.conv3_6_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model.conv3_7_1(h))
h = F.relu(self.model.conv3_7_2(h))
h = self.model.conv3_7_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model.conv3_8_1(h))
h = F.relu(self.model.conv3_8_2(h))
h = self.model.conv3_8_3(h)
h = F.relu(h + h_rem)
h_rem = self.model.conv4_1_ex(h)
h = F.relu(self.model.conv4_1_1(h))
h = F.relu(self.model.conv4_1_2(h))
h = self.model.conv4_1_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model.conv4_2_1(h))
h = F.relu(self.model.conv4_2_2(h))
h = self.model.conv4_2_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model.conv4_3_1(h))
h = F.relu(self.model.conv4_3_2(h))
h = self.model.conv4_3_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model.conv4_4_1(h))
h = F.relu(self.model.conv4_4_2(h))
h = self.model.conv4_4_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model.conv4_5_1(h))
h = F.relu(self.model.conv4_5_2(h))
h = self.model.conv4_5_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model.conv4_6_1(h))
h = F.relu(self.model.conv4_6_2(h))
h = self.model.conv4_6_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model.conv4_7_1(h))
h = F.relu(self.model.conv4_7_2(h))
h = self.model.conv4_7_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model.conv4_8_1(h))
h = F.relu(self.model.conv4_8_2(h))
h = self.model.conv4_8_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model.conv4_9_1(h))
h = F.relu(self.model.conv4_9_2(h))
h = self.model.conv4_9_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model.conv4_10_1(h))
h = F.relu(self.model.conv4_10_2(h))
h = self.model.conv4_10_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model.conv4_11_1(h))
h = F.relu(self.model.conv4_11_2(h))
h = self.model.conv4_11_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model.conv4_12_1(h))
h = F.relu(self.model.conv4_12_2(h))
h = self.model.conv4_12_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model.conv4_13_1(h))
h = F.relu(self.model.conv4_13_2(h))
h = self.model.conv4_13_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model.conv4_14_1(h))
h = F.relu(self.model.conv4_14_2(h))
h = self.model.conv4_14_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model.conv4_15_1(h))
h = F.relu(self.model.conv4_15_2(h))
h = self.model.conv4_15_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model.conv4_16_1(h))
h = F.relu(self.model.conv4_16_2(h))
h = self.model.conv4_16_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model.conv4_17_1(h))
h = F.relu(self.model.conv4_17_2(h))
h = self.model.conv4_17_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model.conv4_18_1(h))
h = F.relu(self.model.conv4_18_2(h))
h = self.model.conv4_18_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model.conv4_19_1(h))
h = F.relu(self.model.conv4_19_2(h))
h = self.model.conv4_19_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model.conv4_20_1(h))
h = F.relu(self.model.conv4_20_2(h))
h = self.model.conv4_20_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model.conv4_21_1(h))
h = F.relu(self.model.conv4_21_2(h))
h = self.model.conv4_21_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model.conv4_22_1(h))
h = F.relu(self.model.conv4_22_2(h))
h = self.model.conv4_22_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model.conv4_23_1(h))
h = F.relu(self.model.conv4_23_2(h))
h = self.model.conv4_23_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model.conv4_24_1(h))
h = F.relu(self.model.conv4_24_2(h))
h = self.model.conv4_24_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model.conv4_25_1(h))
h = F.relu(self.model.conv4_25_2(h))
h = self.model.conv4_25_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model.conv4_26_1(h))
h = F.relu(self.model.conv4_26_2(h))
h = self.model.conv4_26_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model.conv4_27_1(h))
h = F.relu(self.model.conv4_27_2(h))
h = self.model.conv4_27_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model.conv4_28_1(h))
h = F.relu(self.model.conv4_28_2(h))
h = self.model.conv4_28_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model.conv4_29_1(h))
h = F.relu(self.model.conv4_29_2(h))
h = self.model.conv4_29_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model.conv4_30_1(h))
h = F.relu(self.model.conv4_30_2(h))
h = self.model.conv4_30_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model.conv4_31_1(h))
h = F.relu(self.model.conv4_31_2(h))
h = self.model.conv4_31_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model.conv4_32_1(h))
h = F.relu(self.model.conv4_32_2(h))
h = self.model.conv4_32_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model.conv4_33_1(h))
h = F.relu(self.model.conv4_33_2(h))
h = self.model.conv4_33_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model.conv4_34_1(h))
h = F.relu(self.model.conv4_34_2(h))
h = self.model.conv4_34_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model.conv4_35_1(h))
h = F.relu(self.model.conv4_35_2(h))
h = self.model.conv4_35_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model.conv4_36_1(h))
h = F.relu(self.model.conv4_36_2(h))
h = self.model.conv4_36_3(h)
h = F.relu(h + h_rem)
h_rem = self.model.conv5_1_ex(h)
h = F.relu(self.model.conv5_1_1(h))
h = F.relu(self.model.conv5_1_2(h))
h = self.model.conv5_1_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model.conv5_2_1(h))
h = F.relu(self.model.conv5_2_2(h))
h = self.model.conv5_2_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model.conv5_3_1(h))
h = F.relu(self.model.conv5_3_2(h))
h = self.model.conv5_3_3(h)
h = F.relu(h + h_rem)
h = F.average_pooling_2d(h, 7)
Q = self.model.q_value(h)
return Q
def Q_func_target(self, state):
h = F.relu(self.model_target.conv1(state))
h = F.max_pooling_2d(h, 3, stride=2)
h_rem = self.model_target.conv2_1_ex(h)
h = F.relu(self.model_target.conv2_1_1(h))
h = F.relu(self.model_target.conv2_1_2(h))
h = self.model_target.conv2_1_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model_target.conv2_2_1(h))
h = F.relu(self.model_target.conv2_2_2(h))
h = self.model_target.conv2_2_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model_target.conv2_3_1(h))
h = F.relu(self.model_target.conv2_3_2(h))
h = self.model_target.conv2_3_3(h)
h = F.relu(h + h_rem)
h_rem = self.model_target.conv3_1_ex(h)
h = F.relu(self.model_target.conv3_1_1(h))
h = F.relu(self.model_target.conv3_1_2(h))
h = self.model_target.conv3_1_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model_target.conv3_2_1(h))
h = F.relu(self.model_target.conv3_2_2(h))
h = self.model_target.conv3_2_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model_target.conv3_3_1(h))
h = F.relu(self.model_target.conv3_3_2(h))
h = self.model_target.conv3_3_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model_target.conv3_4_1(h))
h = F.relu(self.model_target.conv3_4_2(h))
h = self.model_target.conv3_4_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model_target.conv3_5_1(h))
h = F.relu(self.model_target.conv3_5_2(h))
h = self.model_target.conv3_5_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model_target.conv3_6_1(h))
h = F.relu(self.model_target.conv3_6_2(h))
h = self.model_target.conv3_6_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model_target.conv3_7_1(h))
h = F.relu(self.model_target.conv3_7_2(h))
h = self.model_target.conv3_7_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model_target.conv3_8_1(h))
h = F.relu(self.model_target.conv3_8_2(h))
h = self.model_target.conv3_8_3(h)
h = F.relu(h + h_rem)
h_rem = self.model_target.conv4_1_ex(h)
h = F.relu(self.model_target.conv4_1_1(h))
h = F.relu(self.model_target.conv4_1_2(h))
h = self.model_target.conv4_1_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model_target.conv4_2_1(h))
h = F.relu(self.model_target.conv4_2_2(h))
h = self.model_target.conv4_2_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model_target.conv4_3_1(h))
h = F.relu(self.model_target.conv4_3_2(h))
h = self.model_target.conv4_3_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model_target.conv4_4_1(h))
h = F.relu(self.model_target.conv4_4_2(h))
h = self.model_target.conv4_4_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model_target.conv4_5_1(h))
h = F.relu(self.model_target.conv4_5_2(h))
h = self.model_target.conv4_5_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model_target.conv4_6_1(h))
h = F.relu(self.model_target.conv4_6_2(h))
h = self.model_target.conv4_6_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model_target.conv4_7_1(h))
h = F.relu(self.model_target.conv4_7_2(h))
h = self.model_target.conv4_7_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model_target.conv4_8_1(h))
h = F.relu(self.model_target.conv4_8_2(h))
h = self.model_target.conv4_8_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model_target.conv4_9_1(h))
h = F.relu(self.model_target.conv4_9_2(h))
h = self.model_target.conv4_9_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model_target.conv4_10_1(h))
h = F.relu(self.model_target.conv4_10_2(h))
h = self.model_target.conv4_10_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model_target.conv4_11_1(h))
h = F.relu(self.model_target.conv4_11_2(h))
h = self.model_target.conv4_11_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model_target.conv4_12_1(h))
h = F.relu(self.model_target.conv4_12_2(h))
h = self.model_target.conv4_12_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model_target.conv4_13_1(h))
h = F.relu(self.model_target.conv4_13_2(h))
h = self.model_target.conv4_13_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model_target.conv4_14_1(h))
h = F.relu(self.model_target.conv4_14_2(h))
h = self.model_target.conv4_14_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model_target.conv4_15_1(h))
h = F.relu(self.model_target.conv4_15_2(h))
h = self.model_target.conv4_15_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model_target.conv4_16_1(h))
h = F.relu(self.model_target.conv4_16_2(h))
h = self.model_target.conv4_16_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model_target.conv4_17_1(h))
h = F.relu(self.model_target.conv4_17_2(h))
h = self.model_target.conv4_17_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model_target.conv4_18_1(h))
h = F.relu(self.model_target.conv4_18_2(h))
h = self.model_target.conv4_18_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model_target.conv4_19_1(h))
h = F.relu(self.model_target.conv4_19_2(h))
h = self.model_target.conv4_19_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model_target.conv4_20_1(h))
h = F.relu(self.model_target.conv4_20_2(h))
h = self.model_target.conv4_20_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model_target.conv4_21_1(h))
h = F.relu(self.model_target.conv4_21_2(h))
h = self.model_target.conv4_21_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model_target.conv4_22_1(h))
h = F.relu(self.model_target.conv4_22_2(h))
h = self.model_target.conv4_22_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model_target.conv4_23_1(h))
h = F.relu(self.model_target.conv4_23_2(h))
h = self.model_target.conv4_23_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model_target.conv4_24_1(h))
h = F.relu(self.model_target.conv4_24_2(h))
h = self.model_target.conv4_24_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model_target.conv4_25_1(h))
h = F.relu(self.model_target.conv4_25_2(h))
h = self.model_target.conv4_25_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model_target.conv4_26_1(h))
h = F.relu(self.model_target.conv4_26_2(h))
h = self.model_target.conv4_26_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model_target.conv4_27_1(h))
h = F.relu(self.model_target.conv4_27_2(h))
h = self.model_target.conv4_27_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model_target.conv4_28_1(h))
h = F.relu(self.model_target.conv4_28_2(h))
h = self.model_target.conv4_28_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model_target.conv4_29_1(h))
h = F.relu(self.model_target.conv4_29_2(h))
h = self.model_target.conv4_29_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model_target.conv4_30_1(h))
h = F.relu(self.model_target.conv4_30_2(h))
h = self.model_target.conv4_30_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model_target.conv4_31_1(h))
h = F.relu(self.model_target.conv4_31_2(h))
h = self.model_target.conv4_31_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model_target.conv4_32_1(h))
h = F.relu(self.model_target.conv4_32_2(h))
h = self.model_target.conv4_32_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model_target.conv4_33_1(h))
h = F.relu(self.model_target.conv4_33_2(h))
h = self.model_target.conv4_33_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model_target.conv4_34_1(h))
h = F.relu(self.model_target.conv4_34_2(h))
h = self.model_target.conv4_34_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model_target.conv4_35_1(h))
h = F.relu(self.model_target.conv4_35_2(h))
h = self.model_target.conv4_35_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model_target.conv4_36_1(h))
h = F.relu(self.model_target.conv4_36_2(h))
h = self.model_target.conv4_36_3(h)
h = F.relu(h + h_rem)
h_rem = self.model_target.conv5_1_ex(h)
h = F.relu(self.model_target.conv5_1_1(h))
h = F.relu(self.model_target.conv5_1_2(h))
h = self.model_target.conv5_1_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model_target.conv5_2_1(h))
h = F.relu(self.model_target.conv5_2_2(h))
h = self.model_target.conv5_2_3(h)
h = F.relu(h + h_rem)
h_rem = h
h = F.relu(self.model_target.conv5_3_1(h))
h = F.relu(self.model_target.conv5_3_2(h))
h = self.model_target.conv5_3_3(h)
h = F.relu(h + h_rem)
h = F.average_pooling_2d(h, 7)
Q = self.model_target.q_value(h)
return Q
def e_greedy(self, state, epsilon):
s = Variable(state)
Q = self.Q_func(s)
Q = Q.data
if np.random.rand() < epsilon:
index_action = np.random.randint(0, self.num_of_actions)
print "RANDOM"
else:
index_action = np.argmax(Q.get())
print "GREEDY"
return self.index_to_action(index_action)
def target_model_update(self):
self.model_target = copy.deepcopy(self.model)
def index_to_action(self, index_of_action):
return self.enable_controller[index_of_action]
def action_to_index(self, action):
return self.enable_controller.index(action)
class DQN_class:
# Hyper-Parameters
gamma = 0.99 # Discount factor
initial_exploration = 100 #10**4 # Initial exploratoin. original: 5x10^4
replay_size = 32 # Replay (batch) size
target_model_update_freq = 10**4 # Target update frequancy. original: 10^4
data_size = 10**5 # Data size of history. original: 10^6
def __init__(self, enable_controller=[0, 3, 4]):
self.num_of_actions = len(enable_controller)
self.enable_controller = enable_controller # Default setting : "Pong"
print "Initializing DQN..."
# Initialization of Chainer 1.1.0 or older.
# print "CUDA init"
# cuda.init()
print "Model Building"
self.model = FunctionSet(
l1=convlstm_link.CONVLSTM(7056, 7056),
l4=F.Linear(7056, 512, wscale=np.sqrt(2)),
q_value=F.Linear(512,
self.num_of_actions,
initialW=np.zeros((self.num_of_actions, 512),
dtype=np.float32))).to_gpu()
self.model_target = copy.deepcopy(self.model)
print "Initizlizing Optimizer"
self.optimizer = optimizers.RMSpropGraves(lr=0.00025,
alpha=0.95,
momentum=0.95,
eps=0.0001)
self.optimizer.setup(self.model.collect_parameters())
# History Data : D=[s, a, r, s_dash, end_episode_flag]
self.D = [
np.zeros((self.data_size, 4, 84, 84), dtype=np.uint8),
np.zeros(self.data_size, dtype=np.uint8),
np.zeros((self.data_size, 1), dtype=np.int8),
np.zeros((self.data_size, 4, 84, 84), dtype=np.uint8),
np.zeros((self.data_size, 1), dtype=np.bool)
]
def reset_state(self):
self.model.l2.reset_state()
def forward(self, state, action, Reward, state_dash, episode_end):
num_of_batch = state.shape[0]
s = Variable(state)
s_dash = Variable(state_dash)
Q = self.Q_func(s) # Get Q-value
# Generate Target Signals
tmp = self.Q_func_target(s_dash) # Q(s',*)
tmp = list(map(np.max, tmp.data.get())) # max_a Q(s',a)
max_Q_dash = np.asanyarray(tmp, dtype=np.float32)
target = np.asanyarray(Q.data.get(), dtype=np.float32)
for i in xrange(num_of_batch):
if not episode_end[i][0]:
tmp_ = np.sign(Reward[i]) + self.gamma * max_Q_dash[i]
else:
tmp_ = np.sign(Reward[i])
action_index = self.action_to_index(action[i])
target[i, action_index] = tmp_
# TD-error clipping
td = Variable(cuda.to_gpu(target)) - Q # TD error
td_tmp = td.data + 1000.0 * (abs(td.data) <= 1) # Avoid zero division
td_clip = td * (abs(td.data) <= 1) + td / abs(td_tmp) * (abs(td.data) >
1)
zero_val = Variable(
cuda.to_gpu(
np.zeros((self.replay_size, self.num_of_actions),
dtype=np.float32)))
loss = F.mean_squared_error(td_clip, zero_val)
return loss, Q
def stockExperience(self, time, state, action, reward, state_dash,
episode_end_flag):
data_index = time % self.data_size
if episode_end_flag is True:
self.D[0][data_index] = state
self.D[1][data_index] = action
self.D[2][data_index] = reward
else:
self.D[0][data_index] = state
self.D[1][data_index] = action
self.D[2][data_index] = reward
self.D[3][data_index] = state_dash
self.D[4][data_index] = episode_end_flag
def experienceReplay(self, time):
if self.initial_exploration < time:
# Pick up replay_size number of samples from the Data
if time < self.data_size: # during the first sweep of the History Data
replay_index = np.random.randint(0, time,
(self.replay_size, 1))
else:
replay_index = np.random.randint(0, self.data_size,
(self.replay_size, 1))
s_replay = np.ndarray(shape=(self.replay_size, 4, 84, 84),
dtype=np.float32)
a_replay = np.ndarray(shape=(self.replay_size, 1), dtype=np.uint8)
r_replay = np.ndarray(shape=(self.replay_size, 1),
dtype=np.float32)
s_dash_replay = np.ndarray(shape=(self.replay_size, 4, 84, 84),
dtype=np.float32)
episode_end_replay = np.ndarray(shape=(self.replay_size, 1),
dtype=np.bool)
for i in xrange(self.replay_size):
s_replay[i] = np.asarray(self.D[0][replay_index[i]],
dtype=np.float32)
a_replay[i] = self.D[1][replay_index[i]]
r_replay[i] = self.D[2][replay_index[i]]
s_dash_replay[i] = np.array(self.D[3][replay_index[i]],
dtype=np.float32)
episode_end_replay[i] = self.D[4][replay_index[i]]
s_replay = cuda.to_gpu(s_replay)
s_dash_replay = cuda.to_gpu(s_dash_replay)
# Gradient-based update
self.optimizer.zero_grads()
loss, _ = self.forward(s_replay, a_replay, r_replay, s_dash_replay,
episode_end_replay)
loss.backward()
self.optimizer.update()
'''
def Q_func(self, state):
h1 = F.relu(self.model.l1(state / 254.0)) # scale inputs in [0.0 1.0]
#h2 = F.relu(self.model.l2(h1))
#h3 = F.relu(self.model.l3(h2))
h4 = F.relu(self.model.l4(h1))
Q = self.model.q_value(h4)
return Q
def Q_func_target(self, state):
h1 = F.relu(self.model_target.l1(state / 254.0)) # scale inputs in [0.0 1.0]
#h2 = F.relu(self.model_target.l2(h1))
#h3 = F.relu(self.model_target.l3(h2))
h4 = F.relu(self.model.l4(h1))
Q = self.model_target.q_value(h4)
return Q
'''
def Q_func(self, state):
self.model.l1.reset_state()
for i in range(4):
h1 = F.relu(self.model.l1(state / 254.0))
h4 = F.relu(self.model.l4(h1))
Q = self.model.q_value(h4)
return Q
def Q_func_target(self, state):
self.model_target.l1.reset_state()
for i in range(4):
h1 = F.relu(self.model_target.l1(state / 254.0))
h4 = F.relu(self.model_target.l4(h1))
Q = self.model_target.q_value(h4)
return Q
def e_greedy(self, state, epsilon):
s = Variable(state)
Q = self.Q_func(s)
#Q = self.Q_func(state)
Q = Q.data
if np.random.rand() < epsilon:
index_action = np.random.randint(0, self.num_of_actions)
print "RANDOM"
else:
index_action = np.argmax(Q.get())
print "GREEDY"
return self.index_to_action(index_action), Q
def target_model_update(self):
self.model_target = copy.deepcopy(self.model)
def index_to_action(self, index_of_action):
return self.enable_controller[index_of_action]
def action_to_index(self, action):
return self.enable_controller.index(action)
class DQN_class:
# Hyper-Parameters
gamma = 0.99 # Discount factor
initial_exploration = 100#10**4 # Initial exploratoin. original: 5x10^4
replay_size = 32 # Replay (batch) size
target_model_update_freq = 10**4 # Target update frequancy. original: 10^4
data_size = 10**5 #10**5 # Data size of history. original: 10^6
def __init__(self, enable_controller=[0, 3, 4]):
self.num_of_actions = len(enable_controller)
self.enable_controller = enable_controller # Default setting : "Pong"
print "Initializing DQN..."
print "Model Building"
self.CNN_model = FunctionSet(
l1=F.Convolution2D(4, 32, ksize=8, stride=4, nobias=False, wscale=np.sqrt(2)),
l2=F.Convolution2D(32, 64, ksize=4, stride=2, nobias=False, wscale=np.sqrt(2)),
l3=F.Convolution2D(64, 64, ksize=3, stride=1, nobias=False, wscale=np.sqrt(2)),
)
self.model = FunctionSet(
l4=F.Linear(3136, 512, wscale=np.sqrt(2)),
q_value=F.Linear(512, self.num_of_actions,
initialW=np.zeros((self.num_of_actions, 512),
dtype=np.float32))
).to_gpu()
d = 'elite/'
self.CNN_model.l1.W.data = np.load(d+'l1_W.npy')#.astype(np.float32)
self.CNN_model.l1.b.data = np.load(d+'l1_b.npy')#.astype(np.float32)
self.CNN_model.l2.W.data = np.load(d+'l2_W.npy')#.astype(np.float32)
self.CNN_model.l2.b.data = np.load(d+'l2_b.npy')#.astype(np.float32)
self.CNN_model.l3.W.data = np.load(d+'l3_W.npy')#.astype(np.float32)
self.CNN_model.l3.b.data = np.load(d+'l3_b.npy')#.astype(np.float32)
self.CNN_model = self.CNN_model.to_gpu()
self.CNN_model_target = copy.deepcopy(self.CNN_model)
self.model_target = copy.deepcopy(self.model)
print "Initizlizing Optimizer"
self.optimizer = optimizers.RMSpropGraves(lr=0.00025, alpha=0.95, momentum=0.95, eps=0.0001)
self.optimizer.setup(self.model.collect_parameters())
# History Data : D=[s, a, r, s_dash, end_episode_flag]
self.D = [np.zeros((self.data_size, 4, 84, 84), dtype=np.uint8),
np.zeros(self.data_size, dtype=np.uint8),
np.zeros((self.data_size, 1), dtype=np.int8),
np.zeros((self.data_size, 4, 84, 84), dtype=np.uint8),
np.zeros((self.data_size, 1), dtype=np.bool),
np.zeros((self.data_size, 1), dtype=np.uint8)]
def forward(self, state, action, Reward, state_dash, episode_end):
num_of_batch = state.shape[0]
s = Variable(state)
s_dash = Variable(state_dash)
Q = self.Q_func(s) # Get Q-value
# Generate Target Signals
tmp = self.Q_func_target(s_dash) # Q(s',*)
tmp = list(map(np.max, tmp.data.get())) # max_a Q(s',a)
max_Q_dash = np.asanyarray(tmp, dtype=np.float32)
target = np.asanyarray(Q.data.get(), dtype=np.float32)
for i in xrange(num_of_batch):
if not episode_end[i][0]:
tmp_ = np.sign(Reward[i]) + self.gamma * max_Q_dash[i]
else:
tmp_ = np.sign(Reward[i])
action_index = self.action_to_index(action[i])
target[i, action_index] = tmp_
# TD-error clipping
td = Variable(cuda.to_gpu(target)) - Q # TD error
td_tmp = td.data + 1000.0 * (abs(td.data) <= 1) # Avoid zero division
td_clip = td * (abs(td.data) <= 1) + td/abs(td_tmp) * (abs(td.data) > 1)
zero_val = Variable(cuda.to_gpu(np.zeros((self.replay_size, self.num_of_actions), dtype=np.float32)))
loss = F.mean_squared_error(td_clip, zero_val)
return loss, Q
def stockExperience(self, time,
state, action, lstm_reward, state_dash,
episode_end_flag, ale_reward):
data_index = time % self.data_size
if episode_end_flag is True:
self.D[0][data_index] = state
self.D[1][data_index] = action
self.D[2][data_index] = lstm_reward
self.D[5][data_index] = ale_reward
else:
self.D[0][data_index] = state
self.D[1][data_index] = action
self.D[2][data_index] = lstm_reward
self.D[3][data_index] = state_dash
self.D[5][data_index] = ale_reward
self.D[4][data_index] = episode_end_flag
def experienceReplay(self, time):
if self.initial_exploration < time:
# Pick up replay_size number of samples from the Data
if time < self.data_size: # during the first sweep of the History Data
replay_index = np.random.randint(0, time, (self.replay_size, 1))
else:
replay_index = np.random.randint(0, self.data_size, (self.replay_size, 1))
s_replay = np.ndarray(shape=(self.replay_size, 4, 84, 84), dtype=np.float32)
a_replay = np.ndarray(shape=(self.replay_size, 1), dtype=np.uint8)
r_replay = np.ndarray(shape=(self.replay_size, 1), dtype=np.float32)
s_dash_replay = np.ndarray(shape=(self.replay_size, 4, 84, 84), dtype=np.float32)
episode_end_replay = np.ndarray(shape=(self.replay_size, 1), dtype=np.bool)
for i in xrange(self.replay_size):
s_replay[i] = np.asarray(self.D[0][replay_index[i]], dtype=np.float32)
a_replay[i] = self.D[1][replay_index[i]]
r_replay[i] = self.D[2][replay_index[i]]
s_dash_replay[i] = np.array(self.D[3][replay_index[i]], dtype=np.float32)
episode_end_replay[i] = self.D[4][replay_index[i]]
s_replay = cuda.to_gpu(s_replay)
s_dash_replay = cuda.to_gpu(s_dash_replay)
# Gradient-based update
self.optimizer.zero_grads()
loss, _ = self.forward(s_replay, a_replay, r_replay, s_dash_replay, episode_end_replay)
loss.backward()
self.optimizer.update()
def Q_func(self, state):
h1 = F.relu(self.CNN_model.l1(state / 254.0)) # scale inputs in [0.0 1.0]
h2 = F.relu(self.CNN_model.l2(h1))
h3 = F.relu(self.CNN_model.l3(h2))
h4 = F.relu(self.model.l4(h3))
#test now
#print h3.data.shape
Q = self.model.q_value(h4)
return Q
def Q_func_LSTM(self, state):
h1 = F.relu(self.CNN_model.l1(state / 254.0)) # scale inputs in [0.0 1.0]
h2 = F.relu(self.CNN_model.l2(h1))
h3 = F.relu(self.CNN_model.l3(h2))
return h3.data.get()
def Q_func_target(self, state):
h1 = F.relu(self.CNN_model_target.l1(state / 254.0)) # scale inputs in [0.0 1.0]
h2 = F.relu(self.CNN_model_target.l2(h1))
h3 = F.relu(self.CNN_model_target.l3(h2))
h4 = F.relu(self.model.l4(h3))
Q = self.model_target.q_value(h4)
return Q
def LSTM_reward(self, lstm_out, state_next):
lstm_reward = np.sign((self.lstm_loss - (lstm_out - state_next)**2))
return lstm_reward
def e_greedy(self, state, epsilon):
s = Variable(state)
Q = self.Q_func(s)
Q = Q.data
if np.random.rand() < epsilon:
index_action = np.random.randint(0, self.num_of_actions)
print "RANDOM"
else:
index_action = np.argmax(Q.get())
print "GREEDY"
return self.index_to_action(index_action), Q
def target_model_update(self):
self.model_target = copy.deepcopy(self.model)
def index_to_action(self, index_of_action):
return self.enable_controller[index_of_action]
def action_to_index(self, action):
return self.enable_controller.index(action)
class DQN_class:
# Hyper-Parameters
gamma = 0.99 # Discount factor
initial_exploration = 10**4 # Initial exploratoin. original: 5x10^4
replay_size = 32 # Replay (batch) size
target_model_update_freq = 10**4 # Target update frequancy. original: 10^4
data_size = 10**5 # Data size of history. original: 10^6
def __init__(self, enable_controller=[0, 3, 4]):
self.num_of_actions = len(enable_controller)
self.enable_controller = enable_controller # Default setting : "Pong"
print "Initializing DQN..."
# Initialization for Chainer 1.1.0 or older.
# print "CUDA init"
# cuda.init()
print "Model Building"
self.model = FunctionSet(
l1=F.Convolution2D(4, 16, ksize=8, stride=4, wscale=np.sqrt(2)),
l2=F.Convolution2D(16, 32, ksize=4, stride=2, wscale=np.sqrt(2)),
l3=F.Linear(2592, 256),
q_value=F.Linear(256, self.num_of_actions,
initialW=np.zeros((self.num_of_actions, 256),
dtype=np.float32))
).to_gpu()
print "Initizlizing Optimizer"
self.optimizer = optimizers.RMSpropGraves(lr=0.0002, alpha=0.3, momentum=0.2)
self.optimizer.setup(self.model.collect_parameters())
# History Data : D=[s, a, r, s_dash, end_episode_flag]
self.D = [np.zeros((self.data_size, 4, 84, 84), dtype=np.uint8),
np.zeros(self.data_size, dtype=np.uint8),
np.zeros((self.data_size, 1), dtype=np.int8),
np.zeros((self.data_size, 4, 84, 84), dtype=np.uint8),
np.zeros((self.data_size, 1), dtype=np.bool)]
def forward(self, state, action, Reward, state_dash, episode_end):
num_of_batch = state.shape[0]
s = Variable(state)
s_dash = Variable(state_dash)
Q = self.Q_func(s) # Get Q-value
# Generate Target Signals
max_Q_dash_ = self.Q_func(s_dash)
tmp = list(map(np.max, max_Q_dash_.data.get()))
max_Q_dash = np.asanyarray(tmp, dtype=np.float32)
target = np.asanyarray(Q.data.get(), dtype=np.float32)
for i in xrange(num_of_batch):
if not episode_end[i][0]:
tmp_ = np.sign(Reward[i]) + self.gamma * max_Q_dash[i]
else:
tmp_ = np.sign(Reward[i])
target[i, self.action_to_index(action[i])] = tmp_
loss = F.mean_squared_error(Variable(cuda.to_gpu(target)), Q)
return loss, Q
def stockExperience(self, time,
state, action, reward, state_dash,
episode_end_flag):
data_index = time % self.data_size
if episode_end_flag is True:
self.D[0][data_index] = state
self.D[1][data_index] = action
self.D[2][data_index] = reward
else:
self.D[0][data_index] = state
self.D[1][data_index] = action
self.D[2][data_index] = reward
self.D[3][data_index] = state_dash
self.D[4][data_index] = episode_end_flag
def experienceReplay(self, time):
if self.initial_exploration < time:
# Pick up replay_size number of samples from the Data
if time < self.data_size: # during the first sweep of the History Data
replay_index = np.random.randint(0, time, (self.replay_size, 1))
else:
replay_index = np.random.randint(0, self.data_size, (self.replay_size, 1))
s_replay = np.ndarray(shape=(self.replay_size, 4, 84, 84), dtype=np.float32)
a_replay = np.ndarray(shape=(self.replay_size, 1), dtype=np.uint8)
r_replay = np.ndarray(shape=(self.replay_size, 1), dtype=np.float32)
s_dash_replay = np.ndarray(shape=(self.replay_size, 4, 84, 84), dtype=np.float32)
episode_end_replay = np.ndarray(shape=(self.replay_size, 1), dtype=np.bool)
for i in xrange(self.replay_size):
s_replay[i] = np.asarray(self.D[0][replay_index[i]], dtype=np.float32)
a_replay[i] = self.D[1][replay_index[i]]
r_replay[i] = self.D[2][replay_index[i]]
s_dash_replay[i] = np.array(self.D[3][replay_index[i]], dtype=np.float32)
episode_end_replay[i] = self.D[4][replay_index[i]]
s_replay = cuda.to_gpu(s_replay)
s_dash_replay = cuda.to_gpu(s_dash_replay)
# Gradient-based update
self.optimizer.zero_grads()
loss, _ = self.forward(s_replay, a_replay, r_replay, s_dash_replay, episode_end_replay)
loss.backward()
self.optimizer.update()
def Q_func(self, state):
h1 = F.relu(self.model.l1(state / 254.0)) # scale inputs in [0.0, 1.0]
h2 = F.relu(self.model.l2(h1))
h3 = F.relu(self.model.l3(h2))
Q = self.model.q_value(h3)
return Q
def e_greedy(self, state, epsilon):
s = Variable(state)
Q = self.Q_func(s)
Q = Q.data
if np.random.rand() < epsilon:
index_action = np.random.randint(0, self.num_of_actions)
print "RANDOM"
else:
index_action = np.argmax(Q.get())
print "GREEDY"
return self.index_to_action(index_action), Q
def index_to_action(self, index_of_action):
return self.enable_controller[index_of_action]
def action_to_index(self, action):
return self.enable_controller.index(action)
class DQN_class:
# Hyper-Parameters
gamma = 0.99 # Discount factor
initial_exploration = 50000 # 10**4 # Initial exploratoin. original: 5x10^4
replay_size = 32 # Replay (batch) size
target_model_update_freq = 10 ** 4 # Target update frequancy. original: 10^4
data_size = 5 * (10 ** 5) # Data size of history. original: 10^6
field_num = 7
field_size = 17
def __init__(self, control_size=10, field_num=7, field_size=17):
self.num_of_actions = control_size
self.field_size = field_size
# self.enable_controller = enable_controller # Default setting : "Pong"
print "Initializing DQN..."
# Initialization of Chainer 1.1.0 or older.
# print "CUDA init"
# cuda.init()
self.field_num = field_num
print "Model Building"
self.model = FunctionSet(
l1=F.Convolution2D(self.field_num * 4, 16, ksize=5, stride=1, nobias=False, wscale=np.sqrt(2)),
l2=F.Convolution2D(16, 24, ksize=4, stride=1, nobias=False, wscale=np.sqrt(2)),
l3=F.Linear(2400, 512, wscale=np.sqrt(2)),
q_value=F.Linear(512, self.num_of_actions,
initialW=np.zeros((self.num_of_actions, 512),
dtype=np.float32))
).to_gpu()
self.model_target = copy.deepcopy(self.model)
print "Initizlizing Optimizer"
self.optimizer = optimizers.RMSpropGraves(lr=0.00025, alpha=0.95, momentum=0.95, eps=0.0001)
# self.optimizer.setup(self.model.collect_parameters())
self.optimizer.setup(self.model)
# History Data : D=[s, a, r, s_dash, end_episode_flag]
self.D = [np.zeros((self.data_size, self.field_num * 4, self.field_size, self.field_size), dtype=np.uint8),
np.zeros(self.data_size, dtype=np.uint8),
np.zeros((self.data_size, 1), dtype=np.float32),
np.zeros((self.data_size, self.field_num * 4, self.field_size, self.field_size), dtype=np.uint8),
np.zeros((self.data_size, 1), dtype=np.bool)]
def forward(self, state, action, Reward, state_dash, episode_end):
num_of_batch = state.shape[0]
s = Variable(state)
s_dash = Variable(state_dash)
Q = self.Q_func(s) # Get Q-value
# Generate Target Signals
tmp = self.Q_func_target(s_dash) # Q(s',*)
tmp = list(map(np.max, tmp.data.get())) # max_a Q(s',a)
max_Q_dash = np.asanyarray(tmp, dtype=np.float32)
target = np.asanyarray(Q.data.get(), dtype=np.float32)
for i in xrange(num_of_batch):
if not episode_end[i][0]:
tmp_ = np.sign(Reward[i]) + self.gamma * max_Q_dash[i]
else:
tmp_ = np.sign(Reward[i])
# action_index = self.action_to_index(action[i])
target[i, action[i]] = tmp_
# TD-error clipping
td = Variable(cuda.to_gpu(target)) - Q # TD error
# td = Variable(target) - Q # TD error
td_tmp = td.data + 1000.0 * (abs(td.data) <= 1) # Avoid zero division
td_clip = td * (abs(td.data) <= 1) + td / abs(td_tmp) * (abs(td.data) > 1)
#print "td_data " + str(td_clip.data)
zero_val = Variable(cuda.to_gpu(np.zeros((self.replay_size, self.num_of_actions), dtype=np.float32)))
# zero_val = Variable(np.zeros((self.replay_size, self.num_of_actions), dtype=np.float32))
loss = F.mean_squared_error(td_clip, zero_val)
return loss, Q
def stockExperience(self, time,
state, action, reward, state_dash,
episode_end_flag):
data_index = time % self.data_size
if episode_end_flag is True:
self.D[0][data_index] = state
self.D[1][data_index] = action
self.D[2][data_index] = reward
else:
self.D[0][data_index] = state
self.D[1][data_index] = action
self.D[2][data_index] = reward
self.D[3][data_index] = state_dash
self.D[4][data_index] = episode_end_flag
def experienceReplay(self, time):
if self.initial_exploration < time:
# Pick up replay_size number of samples from the Data
if time < self.data_size: # during the first sweep of the History Data
replay_index = np.random.randint(0, time, (self.replay_size, 1))
else:
replay_index = np.random.randint(0, self.data_size, (self.replay_size, 1))
s_replay = np.ndarray(shape=(self.replay_size, self.field_num * 4, self.field_size, self.field_size),
dtype=np.float32)
a_replay = np.ndarray(shape=(self.replay_size, 1), dtype=np.uint8)
r_replay = np.ndarray(shape=(self.replay_size, 1), dtype=np.float32)
s_dash_replay = np.ndarray(shape=(self.replay_size, self.field_num * 4, self.field_size, self.field_size),
dtype=np.float32)
episode_end_replay = np.ndarray(shape=(self.replay_size, 1), dtype=np.bool)
for i in xrange(self.replay_size):
s_replay[i] = np.asarray(self.D[0][replay_index[i]], dtype=np.float32)
a_replay[i] = self.D[1][replay_index[i]]
r_replay[i] = self.D[2][replay_index[i]]
s_dash_replay[i] = np.array(self.D[3][replay_index[i]], dtype=np.float32)
episode_end_replay[i] = self.D[4][replay_index[i]]
s_replay = cuda.to_gpu(s_replay)
s_dash_replay = cuda.to_gpu(s_dash_replay)
##修正なし
# Gradient-based update
self.optimizer.zero_grads()
loss, _ = self.forward(s_replay, a_replay, r_replay, s_dash_replay, episode_end_replay)
loss.backward()
self.optimizer.update()
def Q_func(self, state):
h1 = F.relu(self.model.l1(state / 254.0)) # scale inputs in [0.0 1.0]
h2 = F.relu(self.model.l2(h1))
h3 = F.relu(self.model.l3(h2))
Q = self.model.q_value(h3)
return Q
def Q_func_target(self, state):
h1 = F.relu(self.model_target.l1(state / 254.0)) # scale inputs in [0.0 1.0]
h2 = F.relu(self.model_target.l2(h1))
h3 = F.relu(self.model_target.l3(h2))
Q = self.model_target.q_value(h3)
return Q
def e_greedy(self, state, epsilon):
s = Variable(state)
Q = self.Q_func(s)
Q = Q.data
if np.random.rand() < epsilon:
index_action = np.random.randint(0, self.num_of_actions)
print "RANDOM"
else:
index_action = np.argmax(Q.get())
print "GREEDY"
return index_action, Q
def target_model_update(self):
self.model_target = copy.deepcopy(self.model)
def save_model(self, model_name, opt_name):
serializers.save_hdf5(model_name, self.model)
serializers.save_hdf5(opt_name, self.optimizer)
def read_model(self, model_name, opt_name):
serializers.load_hdf5(model_name, self.model)
serializers.load_hdf5(opt_name, self.optimizer)
self.model_target = copy.deepcopy(self.model)
class DQN_class:
# Hyper-Parameters
gamma = 0.99 # Discount factor
initial_exploration = 5*10**4 # 10**4 # Initial exploratoin. original: 5x10^4
replay_size = 32 # Replay (batch) size
target_model_update_freq = 10**4 # Target update frequancy. original: 10^4
data_size = 10**6 # Data size of history. original: 10^6
num_of_actions = 2 # Action dimention
num_of_states = 12 # State dimention
def __init__(self):
print "Initializing DQN..."
# Initialization of Chainer 1.1.0 or older.
# print "CUDA init"
# cuda.init()
print "Model Building"
# self.model = FunctionSet(
# l1=F.Convolution2D(4, 32, ksize=8, stride=4, nobias=False, wscale=np.sqrt(2)),
# l2=F.Convolution2D(32, 64, ksize=4, stride=2, nobias=False, wscale=np.sqrt(2)),
# l3=F.Convolution2D(64, 64, ksize=3, stride=1, nobias=False, wscale=np.sqrt(2)),
# l4=F.Linear(3136, 512, wscale=np.sqrt(2)),
# q_value=F.Linear(512, self.num_of_actions,
# initialW=np.zeros((self.num_of_actions, 512),
# dtype=np.float32))
# ).to_gpu()
# self.critic = FunctionSet(
# l1=F.Linear(self.num_of_actions+self.num_of_states,512),
# l2=F.Linear(512,256),
# l3=F.Linear(256,128),
# q_value=F.Linear(128,1,initialW=np.zeros((1,128),dtype=np.float32))
# ).to_gpu()
#
# self.actor = FunctionSet(
# l1=F.Linear(self.num_of_states,512),
# l2=F.Linear(512,256),
# l3=F.Linear(256,128),
# a_value=F.Linear(128,self.num_of_actions,initialW=np.zeros((1,128),dtype=np.float32))
# ).to_gpu()
self.critic = FunctionSet(
l1=F.Linear(self.num_of_actions+self.num_of_states,1024),
l2=F.Linear(1024,512),
l3=F.Linear(512,256),
l4=F.Linear(256,128),
q_value=F.Linear(128,1,initialW=np.zeros((1,128),dtype=np.float32))
).to_gpu()
self.actor = FunctionSet(
l1=F.Linear(self.num_of_states,1024),
l2=F.Linear(1024,512),
l3=F.Linear(512,256),
l4=F.Linear(256,128),
a_value=F.Linear(128,self.num_of_actions,initialW=np.zeros((1,128),dtype=np.float32))
).to_gpu()
# self.critic = FunctionSet(
# l1=F.Linear(self.num_of_actions+self.num_of_states,1024,wscale=0.01*math.sqrt(self.num_of_actions+self.num_of_states)),
# l2=F.Linear(1024,512,wscale=0.01*math.sqrt(1024)),
# l3=F.Linear(512,256,wscale=0.01*math.sqrt(512)),
# l4=F.Linear(256,128,wscale=0.01*math.sqrt(256)),
# q_value=F.Linear(128,1,wscale=0.01*math.sqrt(128))
# ).to_gpu()
#
# self.actor = FunctionSet(
# l1=F.Linear(self.num_of_states,1024,wscale=0.01*math.sqrt(self.num_of_states)),
# l2=F.Linear(1024,512,wscale=0.01*math.sqrt(1024)),
# l3=F.Linear(512,256,wscale=0.01*math.sqrt(512)),
# l4=F.Linear(256,128,wscale=0.01*math.sqrt(256)),
# a_value=F.Linear(128,self.num_of_actions,wscale=0.01*math.sqrt(128))
# ).to_gpu()
self.critic_target = copy.deepcopy(self.critic)
self.actor_target = copy.deepcopy(self.actor)
print "Initizlizing Optimizer"
#self.optim_critic = optimizers.RMSpropGraves(lr=0.0001, alpha=0.95, momentum=0.95, eps=0.0001)
#self.optim_actor = optimizers.RMSpropGraves(lr=0.0001, alpha=0.95, momentum=0.95, eps=0.0001)
self.optim_critic = optimizers.Adam(alpha=0.00001)
self.optim_actor = optimizers.Adam(alpha=0.00001)
self.optim_critic.setup(self.critic)
self.optim_actor.setup(self.actor)
# self.optim_critic.add_hook(chainer.optimizer.WeightDecay(0.00001))
# self.optim_critic.add_hook(chainer.optimizer.GradientClipping(10))
# self.optim_actor.add_hook(chainer.optimizer.WeightDecay(0.00001))
# self.optim_actor.add_hook(chainer.optimizer.GradientClipping(10))
# History Data : D=[s, a, r, s_dash, end_episode_flag]
self.D = [np.zeros((self.data_size, self.num_of_states), dtype=np.float32),
np.zeros((self.data_size, self.num_of_actions), dtype=np.float32),
np.zeros((self.data_size, 1), dtype=np.float32),
np.zeros((self.data_size, self.num_of_states), dtype=np.float32),
np.zeros((self.data_size, 1), dtype=np.bool)]
# with open('dqn_dump.json', 'a') as f:
# json.dump(datetime.datetime.now().strftime('%Y-%m-%dT%H:%M:%S'), f)
# f.write('\n')
# json.dump({"alpha": 0.00001, "beta1": 0.7, "beta2": 0.999, "weight_decay": 0.00001}, f)
# f.write('\n')
# f.close()
#self.x_PID = Hover_PID_Controller(12.1, 1.25)
#self.y_PID = Hover_PID_Controller(12.1, 1.25)
def forward(self, state, action, Reward, state_dash, episode_end):
num_of_batch = state.shape[0]
s = Variable(cuda.to_gpu(np.concatenate([state, action],1)))
s_dash = Variable(cuda.to_gpu(state_dash))
Q = self.Q_func(s) # Get Q-value
# Generate Target through target nets
action_dash_tmp = self.A_func_target(s_dash)
action_dash = np.asanyarray(action_dash_tmp.data.get(), dtype=np.float32)
tmp_dash = Variable(cuda.to_gpu(np.concatenate([state_dash, action_dash],1)))
Q_dash_tmp = self.Q_func_target(tmp_dash)
Q_dash = np.asanyarray(Q_dash_tmp.data.get(), dtype=np.float32)
target = np.asanyarray(Q.data.get(), dtype=np.float32)
for i in xrange(num_of_batch):
if not episode_end[i][0]:
tmp_ = Reward[i] + self.gamma * Q_dash[i]
else:
tmp_ = Reward[i]
target[i] = tmp_
# TD-error clipping
td = Variable(cuda.to_gpu(target)) - Q # TD error
td_tmp = td.data + 1000.0 * (abs(td.data) <= 1) # Avoid zero division
td_clip = td * (abs(td.data) <= 1) + td/abs(td_tmp) * (abs(td.data) > 1)
zero_val = Variable(cuda.to_gpu(np.zeros((self.replay_size, 1), dtype=np.float32)))
loss = F.mean_squared_error(td_clip, zero_val)
return loss, Q
def updateActor(self, state):
num_of_batch = state.shape[0]
A_max = 1.0
A_min = -1.0
A = self.A_func(Variable(cuda.to_gpu(state)))
tmp = Variable(cuda.to_gpu(np.concatenate([state, A.data.get()],1)))
Q = self.Q_func(tmp)
# Backward prop towards actor net
#self.critic.zerograds()
#self.actor.zerograds()
Q.grad = cuda.to_gpu(np.ones((num_of_batch, 1), dtype=np.float32)*(-1.0))
# Q.grad = Q.data*(-1.0)
Q.backward()
A.grad = tmp.grad[:,-self.num_of_actions:]
print("sample_A.grad: "+str(A.grad[0]))
for i in xrange(num_of_batch):
for j in xrange(self.num_of_actions):
if A.grad[i][j] < 0:
A.grad[i][j] *= (A_max-A.data[i][j])/(A_max-A_min)
elif A.grad[i][j] > 0:
A.grad[i][j] *= (A.data[i][j]-A_min)/(A_max-A_min)
A.backward()
self.optim_actor.update()
print("sample_A.grad: "+str(A.grad[0]))
def stockExperience(self, time,
state, action, reward, state_dash,
episode_end_flag):
data_index = time % self.data_size
if episode_end_flag is True:
self.D[0][data_index] = state
self.D[1][data_index] = action
self.D[2][data_index] = reward
else:
self.D[0][data_index] = state
self.D[1][data_index] = action
self.D[2][data_index] = reward
self.D[3][data_index] = state_dash
self.D[4][data_index] = episode_end_flag
def experienceReplay(self, time):
if self.initial_exploration < time:
# Pick up replay_size number of samples from the Data
if time < self.data_size: # during the first sweep of the History Data
replay_index = np.random.randint(0, time, (self.replay_size, 1))
else:
replay_index = np.random.randint(0, self.data_size, (self.replay_size, 1))
#reward_list = list(self.D[2])
#replay_index = [i[0] for i in sorted(enumerate(reward_list),key=itemgetter(1),reverse=True)[:32]]
#replay_index = np.asarray(replay_index).reshape(32,1)
s_replay = np.ndarray(shape=(self.replay_size, self.num_of_states), dtype=np.float32)
a_replay = np.ndarray(shape=(self.replay_size, self.num_of_actions), dtype=np.float32)
r_replay = np.ndarray(shape=(self.replay_size, 1), dtype=np.float32)
s_dash_replay = np.ndarray(shape=(self.replay_size, self.num_of_states), dtype=np.float32)
episode_end_replay = np.ndarray(shape=(self.replay_size, 1), dtype=np.bool)
for i in xrange(self.replay_size):
s_replay[i] = np.asarray(self.D[0][replay_index[i]], dtype=np.float32)
a_replay[i] = np.asarray(self.D[1][replay_index[i]], dtype=np.float32)
r_replay[i] = self.D[2][replay_index[i]]
s_dash_replay[i] = np.asarray(self.D[3][replay_index[i]], dtype=np.float32)
episode_end_replay[i] = self.D[4][replay_index[i]]
#s_replay = cuda.to_gpu(s_replay)
#s_dash_replay = cuda.to_gpu(s_dash_replay)
# Gradient-based critic update
self.optim_critic.zero_grads()
loss, q = self.forward(s_replay, a_replay, r_replay, s_dash_replay, episode_end_replay)
loss.backward()
self.optim_critic.update()
# Update the actor
self.optim_critic.zero_grads()
self.optim_actor.zero_grads()
self.updateActor(s_replay)
self.soft_target_model_update()
print "AVG_Q %f" %(np.average(q.data.get()))
print("loss " + str(loss.data))
# with open('dqn_dump.json', 'a') as f:
# json.dump({"time": time, "avg_Q": float(np.average(q.data.get())), "loss": float(loss.data)}, f)
# f.write('\n')
# f.close()
def Q_func(self, state):
# h1 = F.relu(self.critic.l1(state))
# h2 = F.relu(self.critic.l2(h1))
# h3 = F.relu(self.critic.l3(h2))
# Q = self.critic.q_value(h3)
h1 = F.relu(self.critic.l1(state))
h2 = F.relu(self.critic.l2(h1))
h3 = F.relu(self.critic.l3(h2))
h4 = F.relu(self.critic.l4(h3))
Q = self.critic.q_value(h4)
return Q
def Q_func_target(self, state):
# h1 = F.relu(self.critic_target.l1(state))
# h2 = F.relu(self.critic_target.l2(h1))
# h3 = F.relu(self.critic.l3(h2))
# Q = self.critic_target.q_value(h3)
h1 = F.relu(self.critic_target.l1(state))
h2 = F.relu(self.critic_target.l2(h1))
h3 = F.relu(self.critic_target.l3(h2))
h4 = F.relu(self.critic.l4(h3))
Q = self.critic_target.q_value(h4)
return Q
def A_func(self, state):
# h1 = F.relu(self.actor.l1(state))
# h2 = F.relu(self.actor.l2(h1))
# h3 = F.relu(self.actor.l3(h2))
# A = self.actor.a_value(h3)
h1 = F.relu(self.actor.l1(state))
h2 = F.relu(self.actor.l2(h1))
h3 = F.relu(self.actor.l3(h2))
h4 = F.relu(self.actor.l4(h3))
A = self.actor.a_value(h4)
return A
def A_func_target(self, state):
# h1 = F.relu(self.actor_target.l1(state))
# h2 = F.relu(self.actor_target.l2(h1))
# h3 = F.relu(self.actor.l3(h2))
# A = self.actor_target.a_value(h3)
h1 = F.relu(self.actor_target.l1(state))
h2 = F.relu(self.actor_target.l2(h1))
h3 = F.relu(self.actor_target.l3(h2))
h4 = F.relu(self.actor.l4(h3))
A = self.actor_target.a_value(h4)
return A
def e_greedy(self, state, epsilon):
s = Variable(state)
A = self.A_func(s)
A = A.data
if np.random.rand() < epsilon:
action = np.random.uniform(-1.,1.,(1,self.num_of_actions)).astype(np.float32)
# action = np.zeros((1,self.num_of_actions),dtype=np.float32)
# if state[0,0] > 0:
# action[0,0] = np.random.uniform(0.0,0.5)
# elif state[0,0] < 0:
# action[0,0] = np.random.uniform(-0.5,0.0)
# if state[0,1] < 0:
# action[0,1] = np.random.uniform(0.0,0.5)
# elif state[0,1] > 0:
# action[0,1] = np.random.uniform(-0.5,0.0)
#print("teststate"+str(state))
#action[0,0] = -self.x_PID.getCorrection(state[0][0], 0.0)
#action[0,1] = self.y_PID.getCorrection(state[0][1], 0.0)
print "RANDOM"
else:
action = A.get()
print "GREEDY"
#print(str(action))
return action
def hard_target_model_update(self):
self.critic_target = copy.deepcopy(self.critic)
self.actor_target = copy.deepcopy(self.actor)
def soft_target_model_update(self, tau=0.001):
self.critic_target.l1.W.data = tau*self.critic.l1.W.data + (1-tau)*self.critic_target.l1.W.data
self.critic_target.l2.W.data = tau*self.critic.l2.W.data + (1-tau)*self.critic_target.l2.W.data
self.critic_target.l3.W.data = tau*self.critic.l3.W.data + (1-tau)*self.critic_target.l3.W.data
self.critic_target.l4.W.data = tau*self.critic.l4.W.data + (1-tau)*self.critic_target.l4.W.data
self.critic_target.q_value.W.data = tau*self.critic.q_value.W.data + (1-tau)*self.critic_target.q_value.W.data
self.actor_target.l1.W.data = tau*self.actor.l1.W.data + (1-tau)*self.actor_target.l1.W.data
self.actor_target.l2.W.data = tau*self.actor.l2.W.data + (1-tau)*self.actor_target.l2.W.data
self.actor_target.l3.W.data = tau*self.actor.l3.W.data + (1-tau)*self.actor_target.l3.W.data
self.actor_target.l4.W.data = tau*self.actor.l4.W.data + (1-tau)*self.actor_target.l4.W.data
self.actor_target.a_value.W.data = tau*self.actor.a_value.W.data + (1-tau)*self.actor_target.a_value.W.data
class QNet:
# Hyper-Parameters
gamma = 0.99 # Discount factor
initial_exploration = 10**3 # Initial exploratoin. original: 5x10^4
replay_size = 32 # Replay (batch) size
target_model_update_freq = 10**4 # Target update frequancy. original: 10^4
data_size = 10**5 # Data size of history. original: 10^6
hist_size = 1 # original: 4
def __init__(self, use_gpu, enable_controller, dim, epsilon, epsilon_delta, min_eps):
self.use_gpu = use_gpu
self.num_of_actions = len(enable_controller)
self.enable_controller = enable_controller
self.dim = dim
self.epsilon = epsilon
self.epsilon_delta = epsilon_delta
self.min_eps = min_eps
self.time = 0
app_logger.info("Initializing Q-Network...")
hidden_dim = 256
self.model = FunctionSet(
l4=F.Linear(self.dim*self.hist_size, hidden_dim, wscale=np.sqrt(2)),
q_value=F.Linear(hidden_dim, self.num_of_actions,
initialW=np.zeros((self.num_of_actions, hidden_dim),
dtype=np.float32))
)
if self.use_gpu >= 0:
self.model.to_gpu()
self.model_target = copy.deepcopy(self.model)
self.optimizer = optimizers.RMSpropGraves(lr=0.00025, alpha=0.95, momentum=0.95, eps=0.0001)
self.optimizer.setup(self.model.collect_parameters())
# History Data : D=[s, a, r, s_dash, end_episode_flag]
self.d = [np.zeros((self.data_size, self.hist_size, self.dim), dtype=np.uint8),
np.zeros(self.data_size, dtype=np.uint8),
np.zeros((self.data_size, 1), dtype=np.int8),
np.zeros((self.data_size, self.hist_size, self.dim), dtype=np.uint8),
np.zeros((self.data_size, 1), dtype=np.bool)]
def forward(self, state, action, reward, state_dash, episode_end):
num_of_batch = state.shape[0]
s = Variable(state)
s_dash = Variable(state_dash)
q = self.q_func(s) # Get Q-value
# Generate Target Signals
tmp = self.q_func_target(s_dash) # Q(s',*)
if self.use_gpu >= 0:
tmp = list(map(np.max, tmp.data.get())) # max_a Q(s',a)
else:
tmp = list(map(np.max, tmp.data)) # max_a Q(s',a)
max_q_dash = np.asanyarray(tmp, dtype=np.float32)
if self.use_gpu >= 0:
target = np.asanyarray(q.data.get(), dtype=np.float32)
else:
# make new array
target = np.array(q.data, dtype=np.float32)
for i in xrange(num_of_batch):
if not episode_end[i][0]:
tmp_ = reward[i] + self.gamma * max_q_dash[i]
else:
tmp_ = reward[i]
action_index = self.action_to_index(action[i])
target[i, action_index] = tmp_
# TD-error clipping
if self.use_gpu >= 0:
target = cuda.to_gpu(target)
td = Variable(target) - q # TD error
td_tmp = td.data + 1000.0 * (abs(td.data) <= 1) # Avoid zero division
td_clip = td * (abs(td.data) <= 1) + td/abs(td_tmp) * (abs(td.data) > 1)
zero_val = np.zeros((self.replay_size, self.num_of_actions), dtype=np.float32)
if self.use_gpu >= 0:
zero_val = cuda.to_gpu(zero_val)
zero_val = Variable(zero_val)
loss = F.mean_squared_error(td_clip, zero_val)
return loss, q
def q_func(self, state):
h4 = F.relu(self.model.l4(state / 255.0))
q = self.model.q_value(h4)
return q
def q_func_target(self, state):
h4 = F.relu(self.model_target.l4(state / 255.0))
q = self.model_target.q_value(h4)
return q
def e_greedy(self, state, epsilon):
s = Variable(state)
q = self.q_func(s)
q = q.data
if np.random.rand() < epsilon:
index_action = np.random.randint(0, self.num_of_actions)
app_logger.info(" Random")
else:
if self.use_gpu >= 0:
index_action = np.argmax(q.get())
else:
index_action = np.argmax(q)
app_logger.info("#Greedy")
return self.index_to_action(index_action), q
def target_model_update(self):
self.model_target = copy.deepcopy(self.model)
def index_to_action(self, index_of_action):
return self.enable_controller[index_of_action]
def action_to_index(self, action):
return self.enable_controller.index(action)
def start(self, feature):
self.state = np.zeros((self.hist_size, self.dim), dtype=np.uint8)
self.state[0] = feature
state_ = np.asanyarray(self.state.reshape(1, self.hist_size, self.dim), dtype=np.float32)
if self.use_gpu >= 0:
state_ = cuda.to_gpu(state_)
# Generate an Action e-greedy
action, q_now = self.e_greedy(state_, self.epsilon)
return_action = action
return return_action
def update_model(self, replayed_experience):
if replayed_experience[0]:
self.optimizer.zero_grads()
loss, _ = self.forward(replayed_experience[1], replayed_experience[2],
replayed_experience[3], replayed_experience[4], replayed_experience[5])
loss.backward()
self.optimizer.update()
# Target model update
if replayed_experience[0] and np.mod(self.time, self.target_model_update_freq) == 0:
app_logger.info("Model Updated")
self.target_model_update()
self.time += 1
app_logger.info("step: {}".format(self.time))
def step(self, features):
if self.hist_size == 4:
self.state = np.asanyarray([self.state[1], self.state[2], self.state[3], features], dtype=np.uint8)
elif self.hist_size == 2:
self.state = np.asanyarray([self.state[1], features], dtype=np.uint8)
elif self.hist_size == 1:
self.state = np.asanyarray([features], dtype=np.uint8)
else:
app_logger.error("self.DQN.hist_size err")
state_ = np.asanyarray(self.state.reshape(1, self.hist_size, self.dim), dtype=np.float32)
if self.use_gpu >= 0:
state_ = cuda.to_gpu(state_)
# Exploration decays along the time sequence
if self.initial_exploration < self.time:
self.epsilon -= self.epsilon_delta
if self.epsilon < self.min_eps:
self.epsilon = self.min_eps
eps = self.epsilon
else: # Initial Exploation Phase
app_logger.info("Initial Exploration : {}/{} steps".format(self.time, self.initial_exploration))
eps = 1.0
# Generate an Action by e-greedy action selection
action, q_now = self.e_greedy(state_, eps)
if self.use_gpu >= 0:
q_max = np.max(q_now.get())
else:
q_max = np.max(q_now)
return action, eps, q_max
class DQN_class:
# Hyper-Parameters
gamma = 0.99 # Discount factor
initial_exploration = 100 #10**4 # Initial exploratoin. original: 5x10^4
replay_size = 32 # Replay (batch) size
target_model_update_freq = 10**4 # Target update frequancy. original: 10^4
data_size = 30000 # Data size of history. original: 10^6
def __init__(self, enable_controller=[0, 3, 4]):
self.num_of_actions = len(enable_controller)
self.enable_controller = enable_controller # Default setting : "Pong"
print "Initializing DQN..."
print "Model Building"
self.model = FunctionSet(
l1=F.Convolution2D(4,
32,
ksize=8,
stride=4,
nobias=False,
wscale=np.sqrt(2)),
l2=F.Convolution2D(32,
64,
ksize=4,
stride=2,
nobias=False,
wscale=np.sqrt(2)),
l3=F.Convolution2D(64,
64,
ksize=3,
stride=1,
nobias=False,
wscale=np.sqrt(2)),
l4=F.Linear(3136, 512, wscale=np.sqrt(2)),
q_value=F.Linear(512,
self.num_of_actions,
initialW=np.zeros((self.num_of_actions, 512),
dtype=np.float32))) #.to_gpu()
self.model.l1.W = np.load('elite/l1_W.npy')
self.model.l1.b = np.load('elite/l1_b.npy')
self.model.l2.W = np.load('elite/l2_W.npy')
self.model.l2.b = np.load('elite/l2_b.npy')
self.model.l3.W = np.load('elite/l3_W.npy')
self.model.l3.b = np.load('elite/l3_b.npy')
self.model.l4.W = np.load('elite/l4_W.npy')
self.model.l4.b = np.load('elite/l4_b.npy')
self.model.q_value.W = np.load('elite/q_value_W.npy')
self.model.q_value.b = np.load('elite/q_value_b.npy')
self.model_target = copy.deepcopy(self.model)
# History Data : D=[s, a, r, s_dash, end_episode_flag]
self.D = [
np.zeros((self.data_size, 4, 84, 84), dtype=np.uint8),
np.zeros(self.data_size, dtype=np.uint8),
np.zeros((self.data_size, 1), dtype=np.int8),
np.zeros((self.data_size, 4, 84, 84), dtype=np.uint8),
np.zeros((self.data_size, 1), dtype=np.bool)
]
def forward(self, state, action, Reward, state_dash, episode_end):
num_of_batch = state.shape[0]
s = Variable(state)
s_dash = Variable(state_dash)
Q = self.Q_func(s) # Get Q-value
# Generate Target Signals
tmp = self.Q_func_target(s_dash) # Q(s',*)
tmp = list(map(np.max, tmp.data)) # max_a Q(s',a)
max_Q_dash = np.asanyarray(tmp, dtype=np.float32)
target = np.asanyarray(Q.data, dtype=np.float32)
for i in xrange(num_of_batch):
if not episode_end[i][0]:
tmp_ = np.sign(Reward[i]) + self.gamma * max_Q_dash[i]
else:
tmp_ = np.sign(Reward[i])
action_index = self.action_to_index(action[i])
target[i, action_index] = tmp_
# TD-error clipping
td = Variable(target) - Q
td_tmp = td.data + 1000.0 * (abs(td.data) <= 1) # Avoid zero division
td_clip = td * (abs(td.data) <= 1) + td / abs(td_tmp) * (abs(td.data) >
1)
zero_val = Variable(
np.zeros((self.replay_size, self.num_of_actions),
dtype=np.float32))
loss = F.mean_squared_error(td_clip, zero_val)
return loss, Q
def stockExperience(self, time, state, action, reward, state_dash,
episode_end_flag):
data_index = time % self.data_size
if episode_end_flag is True:
self.D[0][data_index] = state
self.D[1][data_index] = action
self.D[2][data_index] = reward
else:
self.D[0][data_index] = state
self.D[1][data_index] = action
self.D[2][data_index] = reward
self.D[3][data_index] = state_dash
self.D[4][data_index] = episode_end_flag
def experienceReplay(self, time):
if self.initial_exploration < time:
# Pick up replay_size number of samples from the Data
if time < self.data_size: # during the first sweep of the History Data
replay_index = np.random.randint(0, time,
(self.replay_size, 1))
else:
replay_index = np.random.randint(0, self.data_size,
(self.replay_size, 1))
s_replay = np.ndarray(shape=(self.replay_size, 4, 84, 84),
dtype=np.float32)
a_replay = np.ndarray(shape=(self.replay_size, 1), dtype=np.uint8)
r_replay = np.ndarray(shape=(self.replay_size, 1),
dtype=np.float32)
s_dash_replay = np.ndarray(shape=(self.replay_size, 4, 84, 84),
dtype=np.float32)
episode_end_replay = np.ndarray(shape=(self.replay_size, 1),
dtype=np.bool)
for i in xrange(self.replay_size):
s_replay[i] = np.asarray(self.D[0][replay_index[i]],
dtype=np.float32)
a_replay[i] = self.D[1][replay_index[i]]
r_replay[i] = self.D[2][replay_index[i]]
s_dash_replay[i] = np.array(self.D[3][replay_index[i]],
dtype=np.float32)
episode_end_replay[i] = self.D[4][replay_index[i]]
# Gradient-based update
loss, _ = self.forward(s_replay, a_replay, r_replay, s_dash_replay,
episode_end_replay)
def Q_func(self, state):
h1 = F.relu(self.model.l1(state / 254.0)) # scale inputs in [0.0 1.0]
h2 = F.relu(self.model.l2(h1))
h3 = F.relu(self.model.l3(h2))
h4 = F.relu(self.model.l4(h3))
Q = self.model.q_value(h4)
return Q
def Q_func_target(self, state):
h1 = F.relu(self.model_target.l1(state /
254.0)) # scale inputs in [0.0 1.0]
h2 = F.relu(self.model_target.l2(h1))
h3 = F.relu(self.model_target.l3(h2))
h4 = F.relu(self.model.l4(h3))
Q = self.model_target.q_value(h4)
return Q
def e_greedy(self, state, epsilon):
s = Variable(state)
Q = self.Q_func(s)
Q = Q.data
if np.random.rand() < epsilon:
index_action = np.random.randint(0, self.num_of_actions)
print "RANDOM"
else:
#index_action = np.argmax(Q.get())
index_action = np.argmax(Q)
print "GREEDY"
return self.index_to_action(index_action), Q
def target_model_update(self):
self.model_target = copy.deepcopy(self.model)
def index_to_action(self, index_of_action):
return self.enable_controller[index_of_action]
def action_to_index(self, action):
return self.enable_controller.index(action)
class DQN_class:
# Hyper-Parameters
gamma = 0.99 # Discount factor
initial_exploration = 10**3 # Initial exploratoin. original
replay_size = 32 # Replay (batch) size
target_model_update_freq = 10**2 # Target update frequancy. original
data_size = 10**5 # Data size of history. original
#actions are 0 => do nothing, 1 -> buy, -1 sell
def __init__(self, input_vector_length,enable_controller=[0, 1, 2]):
self.num_of_actions = len(enable_controller)
self.enable_controller = enable_controller # Default setting : "Pong"
self.input_vector_length = input_vector_length
print "Initializing DQN..."
# Initialization for Chainer 1.1.0 or older.
# print "CUDA init"
# cuda.init()
#inputs --> 5 * 14 (with 10 temporality) + 5 (of last one hour) + 5 (of last 24 hour)
print "Model Building"
self.model = FunctionSet(
l1=F.Linear(input_vector_length, 500),
l2=F.Linear(500, 250),
l3=F.Linear(250, 80),
q_value=F.Linear(80, self.num_of_actions,
initialW=np.zeros((self.num_of_actions, 80),
dtype=np.float32))
).to_gpu()
print "Initizlizing Optimizer"
self.optimizer = optimizers.RMSpropGraves(lr=0.0002, alpha=0.3, momentum=0.2)
self.optimizer.setup(self.model.collect_parameters())
# History Data : D=[s, a, r, s_dash, end_episode_flag]
self.D = [np.zeros((self.data_size, self.input_vector_length), dtype=np.uint8),
np.zeros(self.data_size, dtype=np.uint8),
np.zeros((self.data_size, 1), dtype=np.int8),
np.zeros((self.data_size, self.input_vector_length), dtype=np.uint8),
np.zeros((self.data_size, 1), dtype=np.bool)]
def forward(self, state, action, Reward, state_dash, episode_end):
num_of_batch = state.shape[0]
s = Variable(state)
s_dash = Variable(state_dash)
Q = self.Q_func(s) # Get Q-value
# Generate Target Signals
max_Q_dash_ = self.Q_func(s_dash)
tmp = list(map(np.max, max_Q_dash_.data.get()))
max_Q_dash = np.asanyarray(tmp, dtype=np.float32)
target = np.asanyarray(Q.data.get(), dtype=np.float32)
for i in xrange(num_of_batch):
if not episode_end[i][0]:
tmp_ = np.sign(Reward[i]) + self.gamma * max_Q_dash[i]
else:
tmp_ = np.sign(Reward[i])
target[i, self.action_to_index(action[i])] = tmp_
loss = F.mean_squared_error(Variable(cuda.to_gpu(target)), Q)
return loss, Q
def stockExperience(self, time,
state, action, reward, state_dash,
episode_end_flag):
data_index = time % self.data_size
if episode_end_flag is True:
self.D[0][data_index] = state
self.D[1][data_index] = action
self.D[2][data_index] = reward
else:
self.D[0][data_index] = state
self.D[1][data_index] = action
self.D[2][data_index] = reward
self.D[3][data_index] = state_dash
self.D[4][data_index] = episode_end_flag
def experienceReplay(self, time):
if self.initial_exploration < time:
# Pick up replay_size number of samples from the Data
if time < self.data_size: # during the first sweep of the History Data
replay_index = np.random.randint(0, time, (self.replay_size, 1))
else:
replay_index = np.random.randint(0, self.data_size, (self.replay_size, 1))
s_replay = np.ndarray(shape=(self.replay_size, self.input_vector_length), dtype=np.float32)
a_replay = np.ndarray(shape=(self.replay_size, 1), dtype=np.uint8)
r_replay = np.ndarray(shape=(self.replay_size, 1), dtype=np.float32)
s_dash_replay = np.ndarray(shape=(self.replay_size, self.input_vector_length), dtype=np.float32)
episode_end_replay = np.ndarray(shape=(self.replay_size, 1), dtype=np.bool)
for i in xrange(self.replay_size):
s_replay[i] = np.asarray(self.D[0][replay_index[i]], dtype=np.float32)
a_replay[i] = self.D[1][replay_index[i]]
r_replay[i] = self.D[2][replay_index[i]]
s_dash_replay[i] = np.array(self.D[3][replay_index[i]], dtype=np.float32)
episode_end_replay[i] = self.D[4][replay_index[i]]
s_replay = cuda.to_gpu(s_replay)
s_dash_replay = cuda.to_gpu(s_dash_replay)
# Gradient-based update
self.optimizer.zero_grads()
loss, _ = self.forward(s_replay, a_replay, r_replay, s_dash_replay, episode_end_replay)
loss.backward()
self.optimizer.update()
def Q_func(self, state):
#todo might want to normalize input, but for now I will do that outside this class
h1 = F.relu(self.model.l1(state))
h2 = F.relu(self.model.l2(h1))
h3 = F.relu(self.model.l3(h2))
Q = self.model.q_value(h3)
return Q
def e_greedy(self, state, epsilon):
s = Variable(state)
Q = self.Q_func(s)
Q = Q.data
if np.random.rand() < epsilon:
index_action = np.random.randint(0, self.num_of_actions)
print "RANDOM"
else:
index_action = np.argmax(Q.get())
print "GREEDY"
return self.index_to_action(index_action), Q
def target_model_update(self):
self.model_target = copy.deepcopy(self.model)
def index_to_action(self, index_of_action):
return self.enable_controller[index_of_action]
def action_to_index(self, action):
return self.enable_controller.index(action)
