对于学校作业，我必须使用深度 Q 网络解决 201 件物品的有界背包问题，准确地说是来自“或健身房”的“背包-v2”。第一步是创建一个神经网络，在 Deep Q Network 算法中使用两次作为所谓的 Q 网络和目标 Q 网络。
Q 网络必须将我的问题的当前状态作为其输入，即背包信息，如下所示：array([[item weights], [item values], [item limits]])
array([[ 39, 40, 13, 34, 11, 28, 40, 37, 61, 53, 13, 32, 9,
31, 64, 74, 21, 12, 40, 59, 61, 78, 3, 42, 45, 64,
61, 48, 34, 12, 18, 17, 70, 59, 66, 16, 40, 80, 62,
54, 2, 53, 56, 9, 91, 56, 66, 47, 65, 40, 49, 40,
20, 23, 94, 2, 28, 42, 92, 98, 28, 52, 7, 18, 68,
32, 23, 64, 32, 22, 48, 76, 38, 71, 56, 16, 72, 72,
53, 17, 28, 11, 74, 57, 76, 32, 27, 80, 79, 50, 44,
97, 62, 38, 80, 28, 5, 15, 76, 59, 80, 89, 48, 73,
27, 86, 24, 49, 43, 1, 84, 56, 20, 57, 42, 61, 82,
44, 29, 61, 4, 72, 82, 41, 25, 88, 48, 43, 83, 14,
94, 73, 37, 8, 52, 78, 44, 64, 23, 47, 91, 8, 71,
79, 23, 87, 96, 83, 5, 13, 8, 83, 78, 42, 69, 62,
10, 61, 58, 29, 19, 11, 38, 12, 74, 62, 68, 60, 89,
53, 94, 78, 70, 66, 24, 99, 92, 60, 56, 94, 6, 77,
95, 76, 17, 96, 10, 20, 64, 28, 58, 1, 37, 98, 8,
34, 3, 57, 15, 40, 200],
[ 22, 8, 48, 85, 40, 11, 99, 89, 50, 81, 28, 46, 42,
31, 22, 80, 78, 58, 19, 94, 43, 3, 8, 35, 72, 67,
24, 12, 82, 71, 84, 59, 93, 82, 37, 28, 44, 37, 87,
35, 69, 77, 74, 71, 67, 3, 3, 89, 78, 43, 60, 79,
37, 75, 32, 32, 30, 67, 48, 81, 35, 98, 59, 11, 50,
8, 92, 98, 65, 59, 31, 42, 43, 59, 81, 32, 15, 82,
95, 83, 71, 12, 25, 90, 25, 21, 68, 68, 4, 14, 91,
69, 4, 94, 65, 54, 87, 93, 54, 94, 95, 34, 44, 62,
44, 99, 5, 36, 29, 14, 93, 5, 44, 87, 27, 55, 14,
6, 32, 61, 30, 3, 21, 34, 32, 18, 31, 65, 90, 29,
49, 6, 73, 58, 4, 70, 10, 71, 99, 60, 3, 32, 91,
40, 53, 70, 37, 92, 66, 67, 76, 51, 32, 85, 72, 98,
98, 92, 58, 10, 30, 92, 53, 87, 59, 13, 73, 95, 64,
15, 32, 65, 35, 15, 90, 83, 23, 92, 39, 28, 74, 55,
4, 49, 14, 22, 24, 40, 48, 49, 36, 81, 98, 86, 84,
32, 24, 56, 96, 40, 0],
[ 3, 8, 9, 9, 1, 8, 2, 6, 3, 4, 6, 1, 7,
6, 6, 8, 5, 2, 9, 3, 8, 8, 9, 6, 6, 5,
8, 1, 6, 2, 7, 7, 4, 6, 5, 6, 1, 4, 9,
3, 3, 6, 6, 8, 3, 2, 4, 9, 2, 2, 4, 2,
2, 2, 3, 7, 9, 8, 5, 8, 5, 6, 9, 6, 2,
3, 1, 5, 1, 9, 6, 8, 4, 4, 1, 4, 9, 2,
8, 4, 9, 4, 9, 3, 1, 5, 3, 9, 5, 6, 2,
7, 4, 1, 2, 7, 9, 8, 9, 8, 9, 6, 9, 5,
6, 1, 8, 3, 5, 8, 1, 7, 6, 5, 4, 3, 4,
7, 1, 4, 2, 1, 2, 7, 8, 1, 2, 9, 1, 5,
3, 1, 7, 7, 7, 3, 2, 6, 7, 7, 6, 2, 2,
1, 9, 9, 7, 8, 1, 5, 4, 2, 2, 3, 3, 2,
2, 4, 2, 9, 5, 2, 3, 7, 5, 6, 2, 2, 7,
9, 2, 3, 1, 2, 7, 9, 5, 2, 7, 1, 9, 9,
8, 6, 6, 6, 4, 2, 6, 8, 6, 8, 7, 5, 5,
3, 6, 3, 2, 8, 0]])
每次我迈出一步时，数组的第三部分都会更新，所以当我“选择一个项目”时，限制会减少 1。
输入这个数组后，神经网络必须吐出预期的 Q 值以执行某个动作（因此选择一个项目），此后，必须发生许多其他事情，但现在超出了范围。
问题在于我的输入和输出的尺寸规格，我想在给定状态数组的情况下进行预测，并得到 201 个值的输出，即我的 Q 值。
这是我当前的网络（以及一些测试其输出的代码）：
env = or_gym.make("Knapsack-v2")
env.mask = False
state_space = env.reset() #array of the choosable items --> item- weights, values, limits
action_space = env.action_space.n #max 200 weight
def NN_model():
learning_rate = 0.001
init = tf.keras.initializers.HeUniform()
model = keras.Sequential()
model.add(keras.layers.Dense(48, input_shape = (201,), activation='relu', kernel_initializer=init))
model.add(keras.layers.Dense(96, activation='relu', kernel_initializer=init))
model.add(keras.layers.Dense(201, activation='linear', kernel_regularizer=init))
model.compile(loss=tf.keras.losses.Huber(), optimizer=tf.keras.optimizers.Adam(learning_rate=learning_rate), metrics=['accuracy'])
return model
model = NN_model()
all_actions = model.predict(state_space, verbose=0)
max_index = np.where(all_actions==np.amax(all_actions))
listOfCordinates = list(zip(max_index[0], max_index[1]))
for cord in listOfCordinates:
cord
action=cord[0]
我尝试了几种输入和输出形状，但没有一个能满足我的需求。
预期输出示例：
[0, 1, 2, ........., 198, 199, 200]