Two neural net approches. Both nets are exactly the same:

    3 hidden layers a 30 neurons
    2 inputs neurons, one output neuron
    All activations are sigmoid
    Stochastic Gradient descent algorithm as learning algorithm with eta=3.0
    quadratic cost function : cost_function = tf.scalar_mul(1.0/(N_training_set*2.0),tf.reduce_sum(tf.squared_difference(y,y_)))
    batch_size of 10
    weight initialization: The weights which connect the lth and l+1th layer are initialized with sigma=1/sqrt(N_l), where N_l is the number of neurons in the lth layer.

The data is saved in TH2D_A00_TB10.root.

you can run the nets as follows:
tensorflow: python regression.py
Michael_Nielsen: python start2.py

Some more details:
- before the training phase the inputs and outputs are normalized with the largest value in the whole set. This is necessary because 
  one uses sigmoid neurons.
- After the each training phase the net is tested with the test/validation data set.
- each net prints the total error on the test set after the last training epoch. In the tensorflowcase this is O(1 million) in the Michael Nielsen case this is O(100k).