def evaluate(env, nn, sess, weights_values, render=False):
    state = env.reset()
    reward_sum = 0
    done = False

    # if this was in the loop, took a lot of time. Why?
    y = nn.forward(nn.x, nn.weights)
    prediction = tf.nn.sigmoid(y)
    action_node = tf.greater(prediction, tf.random_uniform(tf.shape(prediction)))
    d = dict(zip(nn.weights, weights_values)) # weights

    while not done:
        if render:
            env.render()
        d[nn.x] = [state] # input
        action = sess.run(action_node, d)[0][0]
        state, reward, done, _ = env.step(action)
        reward_sum = reward_sum + reward

    return reward_sum