def Mont(X,Y):
    summ=[]
    a=[]
    for i in range(n):
        summ.append(tf.cast(k(X,Y)[i],tf.float32)*NN(X)[0][0])
    a.append((betta[0]/n)*tf.math.reduce_sum(summ,axis=0))
    return(a)

def loss():
    l = []
    g1 = np.array([[g(X)]], dtype='float32')
    return tf.reduce_mean(tf.square(NN(X)[0][0][:n] + Mont(X,Y) - g1[0][0]))
loss()

def train():
    with tf.GradientTape() as tape:
        loss_f = loss()
    trainable_variables = list(weights.values()) + list(biases.values())
    gradients = tape.gradient(loss_f, trainable_variables)
    optimizer.apply_gradients(zip(gradients, trainable_variables))

l1=[]
l3=[]
for i in range(2000):
    train()
    if i % 10 == 0:
         l1.append(loss())
         l3.append(tf.reduce_mean(tf.square((tf.ones(n)-NN(X)[0][0][:n]))))
         print('loss',i,loss())
         print(tf.reduce_mean(tf.square((tf.ones(n)-NN(X)[0][0][:n]))))
l2=tf.convert_to_tensor(l1)
l4=tf.convert_to_tensor(l3)

The error value is fixed from somewhere and the program is stopped. Monte Carlo approximation and loss function and train are explained in the code. I am not sure that the above functions are correct…