[A2]: Problems with Neural Network shown in Tutorial example

[coopa33]

Hi,
I tried to implement the Neural Network shown in the Tutorial as example. I have a couple of questions regarding the code:

Question 1

    data.ctrl = np.clip(outputs, -np.pi/2, np.pi/2)

Shouldn't the hinge positions change accumatively? So this, from what I see, reinitializes the hinge positions everytime. This causes my robot to always have very central but positive hinge positions for every hinge (around 0.5 because of the sigmoid and random sampling from the normal distribution?). This can be seen below:

Question 2

    layer1 = sigmoid(np.dot(inputs, W1))
    layer2 = sigmoid(np.dot(layer1, W2))
    outputs = sigmoid(np.dot(layer2, W3))

As far as I know, the sigmoid should only return positive values from [0, 1]. But the allowable range of hinge values is [-pi/2, pi/2]. Shouldn't we rescale the output to be centered around 0?