class ApproxQAgent(Agent):
'''使用近似的价值函数实现的Q学习个体
'''
def __init__(self, env: Env = None,
trans_capacity = 20000,
hidden_dim: int = 16):
if env is None:
raise "agent should have an environment"
super(ApproxQAgent, self).__init__(env, trans_capacity)
self.input_dim, self.output_dim = 1, 1
if isinstance(env.observation_space, spaces.Discrete):
self.input_dim = 1
elif isinstance(env.observation_space, spaces.Box):
self.input_dim = env.observation_space.shape[0]
if isinstance(env.action_space, spaces.Discrete):
self.output_dim = env.action_space.n
elif isinstance(env.action_space, spaces.Box):
self.output_dim = env.action_space.shape[0]
# print("{},{}".format(self.input_dim, self.output_dim))
self.hidden_dim = hidden_dim
self.Q = Approximator(dim_input = self.input_dim,
dim_output = self.output_dim,
dim_hidden = self.hidden_dim)
self.PQ = self.Q.clone() # 更新参数的网络
return
def _decayed_epsilon(self,cur_episode: int,
min_epsilon: float,
max_epsilon: float,
target_episode: int) -> float:
'''获得一个在一定范围内的epsilon
'''
slope = (min_epsilon - max_epsilon) / (target_episode)
intercept = max_epsilon
return max(min_epsilon, slope * cur_episode + intercept)
def _curPolicy(self, s, epsilon = None):
'''依据更新策略的价值函数(网络)产生一个行为
'''
Q_s = self.PQ(s)
rand_value = random()
if epsilon is not None and rand_value < epsilon:
return self.env.action_space.sample()
else:
return int(np.argmax(Q_s))
def performPolicy(self, s, epsilon = None):
return self._curPolicy(s, epsilon)
def _update_Q_net(self):
'''将更新策略的Q网络(连带其参数)复制给输出目标Q值的网络
'''
self.Q = self.PQ.clone()
def _learn_from_memory(self, gamma, batch_size, learning_rate, epochs):
trans_pieces = self.sample(batch_size) # 随机获取记忆里的Transmition
states_0 = np.vstack([x.s0 for x in trans_pieces])
actions_0 = np.array([x.a0 for x in trans_pieces])
reward_1 = np.array([x.reward for x in trans_pieces])
is_done = np.array([x.is_done for x in trans_pieces])
states_1 = np.vstack([x.s1 for x in trans_pieces])
X_batch = states_0
y_batch = self.Q(states_0) # 得到numpy格式的结果
Q_target = reward_1 + gamma * np.max(self.Q(states_1), axis=1)*\
(~ is_done) # is_done则Q_target==reward_1
y_batch[np.arange(len(X_batch)), actions_0] = Q_target
# loss is a torch Variable with size of 1
loss = self.PQ.fit(x = X_batch,
y = y_batch,
learning_rate = learning_rate,
epochs = epochs)
mean_loss = loss.sum().data[0] / batch_size
self._update_Q_net()
return mean_loss
def learning(self, gamma = 0.99,
learning_rate=1e-5,
max_episodes=1000,
batch_size = 64,
min_epsilon = 0.2,
epsilon_factor = 0.1,
epochs = 1):
total_steps, step_in_episode, num_episode = 0, 0, 0
target_episode = max_episodes * epsilon_factor
while num_episode < max_episodes:
epsilon = self._decayed_epsilon(cur_episode = num_episode,
min_epsilon = min_epsilon,
max_epsilon = 1,
target_episode = target_episode)
self.state = self.env.reset()
# self.env.render()
step_in_episode = 0
loss, mean_loss = 0.00, 0.00
is_done = False
while not is_done:
s0 = self.state
a0 = self.performPolicy(s0, epsilon)
s1, r1, is_done, info, total_reward = self.act(a0)
# self.env.render()
step_in_episode += 1
if self.total_trans > batch_size:
loss += self._learn_from_memory(gamma,
batch_size,
learning_rate,
epochs)
mean_loss = loss / step_in_episode
print("{0} epsilon:{1:3.2f}, loss:{2:.3f}".
format(self.experience.last, epsilon, mean_loss))
# print(self.experience)
total_steps += step_in_episode
num_episode += 1
return
class ApproxQAgent(Agent):
def __init__(self,
env: Env = None,
trans_capacity=20000,
hidden_dim: int = 16):
if env is None:
raise "agent should have an environment"
super(ApproxQAgent, self).__init__(env, trans_capacity)
self.input_dim, self.output_dim = 1, 1
if isinstance(env.observation_space, spaces.Discrete):
self.input_dim = 1
elif isinstance(env.observation_space, spaces.Box):
self.input_dim = env.observation_space.shape[0]
if isinstance(env.action_space, spaces.Discrete):
self.output_dim = env.action_space.n
elif isinstance(env.action_space, spaces.Box):
self.output_dim = env.action_space.shape[0]
# print("{},{}".format(self.input_dim, self.output_dim))
self.hidden_dim = hidden_dim
self.Q = Approximator(dim_input=self.input_dim,
dim_output=self.output_dim,
dim_hidden=self.hidden_dim)
self.PQ = self.Q.clone()
return
def _decayed_epsilon(self, cur_episode: int, min_epsilon: float,
max_epsilon: float, target_episode: int) -> float:
slope = (min_epsilon - max_epsilon) / (target_episode)
intercept = max_epsilon
return max(min_epsilon, slope * cur_episode + intercept)
def _curPolicy(self, s, epsilon=None):
Q_s = self.PQ(s)
rand_value = random()
if epsilon is not None and rand_value < epsilon:
return self.env.action_space.sample()
else:
return int(np.argmax(Q_s))
def performPolicy(self, s, epsilon=None):
return self._curPolicy(s, epsilon)
def _update_Q_net(self):
self.Q = self.PQ.clone()
def _learn_from_memory(self, gamma, batch_size, learning_rate, epochs, r,
s):
trans_pieces = self.sample(batch_size)
states_0 = np.vstack([x.s0 for x in trans_pieces])
actions_0 = np.array([x.a0 for x in trans_pieces])
reward_1 = np.array([x.reward for x in trans_pieces])
is_done = np.array([x.is_done for x in trans_pieces])
states_1 = np.vstack([x.s1 for x in trans_pieces])
X_batch = states_0
y_batch = self.Q(states_0)
Q_target = reward_1 + gamma * np.max(self.Q(states_1), axis=1) * \
(~ is_done)
y_batch[np.arange(len(X_batch)), actions_0] = Q_target
# loss is a torch Variable with size of 1
loss = self.PQ.fit(x=X_batch,
y=y_batch,
learning_rate=learning_rate,
epochs=epochs)
mean_loss = loss.sum().data[0] / batch_size
self._update_Q_net()
return mean_loss
def learning(self,
gamma=0.99,
learning_rate=1e-5,
max_episodes=1000,
batch_size=64,
min_epsilon=0.2,
epsilon_factor=0.1,
epochs=1):
total_steps, step_in_episode, num_episode = 0, 0, 0
target_episode = max_episodes * epsilon_factor
file = open('dqn.csv', 'w')
file.write("Episode" + "," + "Distance" + "\n")
tot_dis = 0
file = open('reward.csv', 'w')
file.write("Steps in Episode" + "," + "reward" + "\n")
while num_episode < max_episodes:
epsilon = self._decayed_epsilon(cur_episode=num_episode,
min_epsilon=min_epsilon,
max_epsilon=1,
target_episode=target_episode)
self.state = self.env._reset()
self.env._render()
step_in_episode = 0
loss, mean_loss = 0.00, 0.00
is_done = False
while not is_done:
s0 = self.state
a0 = self.performPolicy(s0, epsilon)
s1, r1, is_done, dis_info = self.env._step_b(a0)
self.env._render()
step_in_episode += 1
tot_dis += r1
print("Step in Episode :: ", step_in_episode)
print("Distance of agent from goal :: ", dis_info)
file.write(str(step_in_episode) + "," + str(tot_dis) + "\n")
if self.total_trans > batch_size:
loss += self._learn_from_memory(gamma, batch_size,
learning_rate, epochs, r1,
s1)
file.close()
mean_loss = loss / step_in_episode
print("{0} epsilon:{1:3.2f}, loss:{2:.3f}".format(
self.experience.last, epsilon, mean_loss))
# print(self.experience)
total_steps += step_in_episode
num_episode += 1
#print("Episode :: ", num_episode)
# print("Distance of agent from goal :: ", dis_info)
return
class ApproxQAgent(Agent):
'''使用近似的价值函数实现的Q学习个体
#Function
1 value function approximation
2 base on Experience Relay, which is good for eliminating relationship of transition in a single episode,
in order to get a better approximation
3 DQN
'''
def __init__(self, env: Env = None,
trans_capacity = 20000,
hidden_dim: int = 16):
'''set input_dim(w.r.t. obs.space) and output_dim(w.r.t. action_space)...
super(...).__init__(...),
self.Q = Approximator(...)
self.PQ = self.Q.clone() #PQ for updating parameters
#args
env: environment of this agent
trans_capacity:<int>max num. of transitions in memory
hiddden_dim:<int>num. of nodes in hidden layer
'''
if env is None:
raise "agent should have an environment"
super(ApproxQAgent, self).__init__(env, trans_capacity)
self.input_dim, self.output_dim = 1, 1
if isinstance(env.observation_space, spaces.Discrete):
self.input_dim = 1
elif isinstance(env.observation_space, spaces.Box):
self.input_dim = env.observation_space.shape[0] #e.g. observation_space>>Box(6,), .shape>>(6,)
if isinstance(env.action_space, spaces.Discrete):
self.output_dim = env.action_space.n #
elif isinstance(env.action_space, spaces.Box):
self.output_dim = env.action_space.shape[0]
# print("{},{}".format(self.input_dim, self.output_dim))
self.hidden_dim = hidden_dim
self.Q = Approximator(dim_input = self.input_dim,
dim_output = self.output_dim,
dim_hidden = self.hidden_dim)
self.PQ = self.Q.clone() # 更新参数的网络
return
def _decayed_epsilon(self,cur_episode: int,
min_epsilon: float,
max_epsilon: float,
target_episode: int) -> float:
'''获得一个在一定范围内的epsilon
#return
epsilon<float>changing from max_epsilon(when cur_episode=0) to min_epsilon w.r.t. cur_episode
'''
slope = (min_epsilon - max_epsilon) / (target_episode)
intercept = max_epsilon
return max(min_epsilon, slope * cur_episode + intercept) #slope*cur_episode is negative
def _curPolicy(self, s, epsilon = None):
'''依据更新策略的价值函数(网络)产生一个行为
#args
s: state s0<6x1 ndarray>
epsilon: =None means greedy, otherwise epsilon greedy
#return
an action a0<int> w.r.t. PQ(policy evaluation) using decayed epsilon-greedy(policy improvement)
'''
Q_s = self.PQ(s) #
rand_value = random()
if epsilon is not None and rand_value < epsilon:
return self.env.action_space.sample()
else:
return int(np.argmax(Q_s))
def performPolicy(self, s, epsilon = None):
#若只有一个Policy,则可略
return self._curPolicy(s, epsilon)
def _update_Q_net(self):
'''将更新策略的Q网络(连带其参数)复制给输出目标Q值的网络
'''
self.Q = self.PQ.clone()
def _learn_from_memory(self, gamma, batch_size, learning_rate, epochs):
# get Transmition randomly from experience, return a <list>, consists of batch_size * Transition object(consists of data,s0,a0,reward,s1,is_done)
trans_pieces = self.sample(batch_size)
states_0 = np.vstack([x.s0 for x in trans_pieces]) #ndarray
actions_0 = np.array([x.a0 for x in trans_pieces])
reward_1 = np.array([x.reward for x in trans_pieces])
is_done = np.array([x.is_done for x in trans_pieces])
states_1 = np.vstack([x.s1 for x in trans_pieces])
X_batch = states_0
# ndarray, consists of list([Q(s0)(a_0), Q(s0)(a_1),....]), describe all Q of all actions in state s0
#y_batch = self.Q(states_0) #main difference in a0 dimension
y_batch = self.PQ(states_0) #only Q(s,a,w) in a0 dimension different. But always walk around
#matrix-weise calculation
Q_target = reward_1 + gamma * np.max(self.Q(states_1), axis=1)*\
(~ is_done) # is_done则Q_target==reward_1
#Attension:
y_batch[np.arange(len(X_batch)), actions_0] = Q_target
# loss is a torch Variable with size of 1
loss = self.PQ.fit(x = X_batch,
y = y_batch,
learning_rate = learning_rate,
epochs = epochs)
mean_loss = loss.sum().data[0] / batch_size
self._update_Q_net()
return mean_loss
def learning(self, gamma = 0.99,
learning_rate=1e-5,
max_episodes=1000,
batch_size = 64,
min_epsilon = 0.2,
epsilon_factor = 0.1,
epochs = 1):
'''contruct experience, when nums of trans. in experience enough, start learning from experience, compute loss
Methods details see below
#Arguments
gamma = 0.99, # discount factor, range from [0,1]
learning_rate=1e-5, # 集中学习的规模
max_episodes=1000, # 最大训练Episode数量
batch_size = 64,
min_epsilon = 0.2,
epsilon_factor = 0.1, # 开始使用最小Epsilon时Episode的序号占最大
# Episodes序号之比，该比值越小，表示使用
# min_epsilon的episode越多
epochs = 1): # 每个batch_size训练的次数
'''
total_steps, step_in_episode, num_episode = 0, 0, 0
target_episode = max_episodes * epsilon_factor
while num_episode < max_episodes: #for each episode until max_episode, get loss
epsilon = self._decayed_epsilon(cur_episode = num_episode,
min_epsilon = min_epsilon,
max_epsilon = 1,
target_episode = target_episode)
self.state = self.env.reset()
self.env.render()
step_in_episode = 0
loss, mean_loss = 0.00, 0.00 #
is_done = False
while not is_done:#for every transition
s0 = self.state #self.state change inside self.act(a0)
a0 = self.performPolicy(s0, epsilon) #get action w.r.t. PQ using decayed epsilon-greedy
s1, r1, is_done, info, total_reward = self.act(a0) #inside self.act(a0): self.state = s1
#inside act also:sotre trans as episode_list in experience, and as trans_list in episode, and accumulate the total_reward
self.env.render()
step_in_episode += 1
if self.total_trans > batch_size:
loss += self._learn_from_memory(gamma,
batch_size,
learning_rate,
epochs)
mean_loss = loss / step_in_episode
print("{0} epsilon:{1:3.2f}, loss:{2:.3f}".
format(self.experience.last, epsilon, mean_loss))
# print(self.experience)
total_steps += step_in_episode
num_episode += 1
return
class ApproxQAgent(Agent):
'''使用近似的价值函数实现的Q学习的个体
'''
def __init__(self, env: Env = None,
trans_capacity=20000,
hidden_dim: int = 16):
if env is None:
raise Exception("agent should have an environment")
super(ApproxQAgent, self).__init__(env, trans_capacity)
self.input_dim, self.output_dim = 1, 1
# 适应不同的状态和行为空间类型
if isinstance(env.observation_space, spaces.Discrete):
self.input_dim = 1
elif isinstance(env.observation_space, spaces.Box):
self.input_dim = env.observation_space.shape[0]
if isinstance(env.action_space, spaces.Discrete):
self.output_dim = env.action_space.n
elif isinstance(env.action_space, spaces.Box):
self.output_dim = env.action_space.shape[0]
# print("{},{}".format(self.input_dim, self.output_dim))
# 隐藏层神经元数目
self.hidden_dim = hidden_dim
# 关键在下面两句，声明了两个近似价值函数
# 变量Q是一个计算价值，产生loss的近似函数（网络），
# 该网络参数在一定时间段内不更新参数
self.Q = Approximator(dim_input=self.input_dim,
dim_output=self.output_dim,
dim_hidden=self.hidden_dim)
# 变量PQ是一个生成策略的近似函数，该函数（网络）的参数频繁更新
# 更新参数的网络
self.PQ = self.Q.clone()
return
def _learning_from_memory(self, gamma, batch_size, learning_rate, epochs):
# 随机获取记忆里的Transmition
trans_pieces = self.sample(batch_size)
states_0 = np.vstack([x.s0 for x in trans_pieces])
actions_0 = np.array([x.a0 for x in trans_pieces])
reward_1 = np.array([x.reward for x in trans_pieces])
is_done = np.array([x.is_done for x in trans_pieces])
states_1 = np.vstack([x.s1 for x in trans_pieces])
X_batch = states_0
# 调用的时approximator的__call__方法
y_batch = self.Q(states_0)
# 使用了Batch，代码是矩阵运算
# np.max => axis=1时取出最大的一列；axis=0时取出最大的一行
# ～ True = -2; ~ False = -1
Q_target = reward_1 + gamma * np.max(self.Q(states_1), axis=1) * (~ is_done)
y_batch[np.arange(len(X_batch)), actions_0] = Q_target
# loss is a torch Variable with size of 1
loss = self.PQ.fit(x=X_batch,
y=y_batch,
learning_rate=learning_rate,
epochs=epochs)
mean_loss = loss.sum().item() / batch_size
self._update_Q_net()
return mean_loss
def learning(self, gamma=0.99,
learning_rate=1e-5,
max_episodes=1000,
batch_size=64,
min_epsilon=0.2,
epsilon_factor=0.1,
epochs=1):
'''learning的主要工作是构建经历，当构建的经历足够时，同时启动基于经历的学习
'''
total_steps, step_in_episode, num_episode = 0, 0, 0
target_episode = max_episodes * epsilon_factor
while num_episode < max_episodes:
epsilon = self._decayed_epsilon(cur_episode=num_episode,
min_epsilon=min_epsilon,
max_epsilon=1,
target_episode=target_episode)
self.state = self.env.reset()
self.env.render()
step_in_episode = 0
loss, mean_loss = 0.00, 0.00
is_done = False
while not is_done:
s0 = self.state
a0 = self.performPolicy(s0, epsilon)
# act方法封装了将Transition记录至Experience中的过程
s1, r1, is_done, info, total_reward = self.act(a0)
# self.env.render()
step_in_episode += 1
# 当经历里有足够大小的Transition时，开始启用基于经历的学习
if self.total_trans > batch_size:
loss += self._learning_from_memory(gamma,
batch_size,
learning_rate,
epochs)
mean_loss = loss / step_in_episode
print("{0} epsilon:{1:3.2f}, loss:{2:.3f}".
format(self.experience.last, epsilon, mean_loss))
# print(self.experience)
total_steps += step_in_episode
num_episode += 1
return
def _decayed_epsilon(self, cur_episode: int,
min_epsilon: float,
max_epsilon: float,
target_episode: int) -> float:
'''获得一个在一定范围内的epsilon
'''
slope = (min_epsilon - max_epsilon) / (target_episode)
intercept = max_epsilon
return max(min_epsilon, slope * cur_episode + intercept)
def _curPolicy(self, s, epsilon=None):
'''依据更新策略的价值函数(网络)产生一个行为
'''
Q_s = self.PQ(s)
rand_value = random()
if epsilon is not None and rand_value < epsilon:
return self.env.action_space.sample()
else:
return int(np.argmax(Q_s))
def performPolicy(self, s, epsilon=None):
return self._curPolicy(s, epsilon)
def _update_Q_net(self):
'''将更新策略的Q网络(连带其参数)复制给输出目标Q值的网络
'''
self.Q = self.PQ.clone()
