q-learning: ValueError: 'a' 不能為空，除非沒有采樣

Question

我嘗試開發用於強化學習的 q-learning 算法，這是我的代碼：

import numpy as np
R = np.matrix ([[-1, 0, -1, -1, 0, -1, -1, -1, -1], 
            [-1, -1, 100, 0, -1, -1, -1, -1, -1], 
            [-1, -1, 100, -1, -1, -1, -1, -1, -1],
            [-1, -1, -1, -1, -1, -1, -1, -1, -1], 
            [-1, -1, -1, -1, -1, 100, 0, -1, -1], 
            [-1, -1, -1, -1, -1, 100, -1, -1, -1], 
            [-1, -1, -1, -1, -1, -1, -1, 100, 0], 
            [-1, -1, -1, -1, -1, -1, -1, 100, -1],
            [-1, -1, -1, -1, -1, -1, -1, -1, -1]])
    # Q matrix
Q = np.matrix(np.zeros([9,9]))

# Gamma (learning parameter)
gamma = 0.4

# Initial state. (Usually to be chosen at random)
initial_state = 1

# This function returns all available actions in the state given as an argument
def available_actions(state):
    current_state_row = R[state,]
    av_act = np.where(current_state_row >= 0) [1]
    return av_act

# Get available actions in the current state
available_act = available_actions(initial_state)

# This function chooses at random which action to be performed within the range of all the available actions.
def sample_next_action(available_actions_range):
    next_action = int(np.random.choice(available_act, 1))
    return next_action

#sample next action to be performed
action = sample_next_action(available_act)

# This function updates the Q matrix according to the path selected and the Q learning algorithm
def update(current_state, action, gamma):
    max_index = np.where(Q[action,] == np.max(Q[action,]))[1]

    if max_index.shape[0] > 1:
        max_index = int(np.random.choice(max_index, size=1))
    else:
        max_index = int(max_index)
    max_value = Q[action, max_index]

    # Q learning formula
    Q[current_state, action] = R[current_state, action] + gamma * max_value

# Update Q matrix
update(initial_state, action, gamma)

# Training
# Train over 10000 iterations. (Re-iterate the process above)
for i in range(10000):
    current_state = np.random.randint(0, int(Q.shape[0]))
    available_act = available_actions(current_state)
    action = sample_next_action(available_act)
    update(current_state, action, gamma)

# Normalize the trained Q matrix
print ("Trained Q matrix:")
print (Q / np.max(Q) * 100)

# Testing 

# Goal state = 2

current_state = 1
steps = [current_state]

while current_state != 2:
    next_step_index = np.where(Q[current_state,] == np.max(Q[current_state,]))[1]

    if next_step_index.shape[0] > 1:
        next_step_index = int(np.random.choice(next_step_index, size=1))
    else:
        next_step_index = int(next_step_index)

    steps.append(next_step_index)
    current_state = next_step_index

# Print selected sequence of steps
print("Selected path:")
print(steps)

但我總是有這個我不明白的錯誤：

46 current_state = np.random.randint(0, int(Q.shape[0])) 47 available_act = available_actions(current_state) ---> 48 action = sample_next_action(available_act) 49 update (current_state, action, gamma) 50

in sample_next_action(available_actions_range) 19 # 這個 function 在所有可用動作的范圍內隨機選擇要執行的動作。 20 def sample_next_action(available_actions_range): ---> 21 next_action = int(np.random.choice(available_act, 1)) 22 return next_action 23

mtrand.pyx 在 mtrand.RandomState.choice()

ValueError: 'a' 不能為空，除非沒有采樣

請有任何幫助！

Answer 1

代碼有很多缺陷：

1. 將 R 和 Q 的數據結構更改為：

   R = np.array ... Q = np.zeros([9, 9])

2. 更改 state 3 和 state 8 的 R 矩陣，以便至少有一個操作可用。 因此，只需在這些行中添加一個大於零的值。

3. 將 available_actions 定義更改為：

    def available_actions(state): current_state_row = R[state, :] av_act = np.where(current_state_row >= 0)[0]

4. 更改第 39 行以進行正確索引

   max_index = np.where(Q[action,] == np.max(Q[action, :]))[0]

5. 更改第 73 行以進行正確索引

   next_step_index = np.where(Q[current_state,:] == np.max(Q[current_state,:]))[0]

通過這些更改，您應該能夠獲得價值。
最終結果將是：

所選路徑：[1, 2]

import numpy as np
R = np.array([[-1, 0, -1, -1, 0, -1, -1, -1, -1],
            [-1, -1, 100, 0, -1, -1, -1, -1, -1],
            [-1, -1, 100, -1, -1, -1, -1, -1, -1],
            [-1, -1, -1, -1, -1, 0, -1, -1, -1],
            [-1, -1, -1, -1, -1, 100, 0, -1, -1],
            [-1, -1, -1, -1, -1, 100, -1, -1, -1],
            [-1, -1, -1, -1, -1, -1, -1, 100, 0],
            [-1, -1, -1, -1, -1, -1, -1, 100, -1],
            [-1, -1, -1, -1, 0, -1, -1, -1, -1]])
    # Q matrix
Q = np.zeros([9,9])

# Gamma (learning parameter)
gamma = 0.4

# Initial state. (Usually to be chosen at random)
initial_state = 1

# This function returns all available actions in the state given as an argument
def available_actions(state):
    current_state_row = R[state, :]
    av_act = np.where(current_state_row >= 0)[0]
    return av_act

# Get available actions in the current state
available_act = available_actions(initial_state)

# This function chooses at random which action to be performed within the range of all the available actions.
def sample_next_action(available_actions_range):
    next_action = int(np.random.choice(available_act, 1))
    return next_action

#sample next action to be performed
action = sample_next_action(available_act)

# This function updates the Q matrix according to the path selected and the Q learning algorithm
def update(current_state, action, gamma):
    max_index = np.where(Q[action,] == np.max(Q[action, :]))[0]

    if max_index.shape[0] > 1:
        max_index = int(np.random.choice(max_index, size=1))
    else:
        max_index = int(max_index)
    max_value = Q[action, max_index]

    # Q learning formula
    Q[current_state, action] = R[current_state, action] + gamma * max_value

# Update Q matrix
update(initial_state, action, gamma)

# Training
# Train over 10000 iterations. (Re-iterate the process above)
for i in range(10000):
    current_state = np.random.randint(0, int(Q.shape[0]))
    available_act = available_actions(current_state)
    action = sample_next_action(available_act)
    update(current_state, action, gamma)

# Normalize the trained Q matrix
print ("Trained Q matrix:")
print (Q / np.max(Q) * 100)

# Testing

# Goal state = 2

current_state = 1
steps = [current_state]

while current_state != 2:
    next_step_index = np.where(Q[current_state,:] == np.max(Q[current_state,:]))[0]

    if next_step_index.shape[0] > 1:
        next_step_index = int(np.random.choice(next_step_index, size=1))
    else:
        next_step_index = int(next_step_index)

    steps.append(next_step_index)
    current_state = next_step_index

# Print selected sequence of steps
print("Selected path:")
print(steps)