Small fix for neural net constraint

[Alexander-Schiendorfer]

The definition of fzn_neural_net seems odd by using the last element of first_edge instead of EDGE for the last neuron.

forall(i in NODE diff INPUTS) (
            let { int: first = first_edge[i];
                  int: last = if i = max(NODE) then max(index_set(first_edge))

which should be

forall(i in NODE diff INPUTS) (
            let { int: first = first_edge[i];
                  int: last = if i = max(NODE) then max(index_set(EDGE))

That might explain why a dummy element is needed according to the documentation:

To debug all of that, I've copied the definition of fzn_neural_net and added a few traces:
nn_example_mzn.zip

With the original definition, I get an empty set of parents for the last neuron

Neuron 4 first 3, last 4
Inputs [X_INTRODUCED_0_, X_INTRODUCED_1_], weights [-3.0, 2.0], -2.0
Neuron 5 first 5, last 6
Inputs [X_INTRODUCED_18_, X_INTRODUCED_27_], weights [0.0, 1.0], 1.0
Neuron 6 first 7, last 6
Inputs [], weights [], -1.0

whereas the proposed change fixes that:

Neuron 4 first 3, last 4
Inputs [X_INTRODUCED_0_, X_INTRODUCED_1_], weights [-3.0, 2.0], -2.0
Neuron 5 first 5, last 6
Inputs [X_INTRODUCED_18_, X_INTRODUCED_27_], weights [0.0, 1.0], 1.0
Neuron 6 first 7, last 8
Inputs [X_INTRODUCED_18_, X_INTRODUCED_27_], weights [2.0, 0.0], -1.0

Additionally, it might make sense to pass an array from node ids to NEURON_TYPE (e.g., the hidden neurons might use non-linearities but only the output neurons should use LINEAR, or SOFTMAX once it becomes available :-) )

Cheers,
Alex