Originally posted by: FiFila

Hi;

I want to model by CP optimizer and NN a problem (I want to improve the NN by combination of it with constraint programming). As the input of NN is not linear, I can not use CPLEX, I have force to use a CP optimizer. Because I am in primary steps of learning a CP optimizer, would you please send me a sample or reference to see how I can model some constraints by CP optimizer as  a nonlinear input of  my NN.

Another question: Is it possible to write the neural network section in Matlab and its constraints section in CPLEX?

Thanks

Hi,

about float expressions in CPO have a look at the example floatexpr.

About calling CPLEX from MATLAB you may have a look at

http://www-01.ibm.com/support/docview.wss?uid=swg21399987

regards

Originally posted by: FiFila

Hi Alex;

Thanks for your help but I want to use real amount (tanh). Have you any idea?

Regards

Fila

Originally posted by: FiFila

Hi;

As I know we are be able to have some constraint  for real number like "e" in CP optimizer but I do not know how I can do it. Alex gave me a guidance to take a look at floatexpr.

Thank you  but as I have seen the example of floatexpr it is related to p/q (p and q are integer) but I want to work with some other types of input like "e" which I am not be able to use this method. Does any one have an idea about it? 

Regards

Fila

Originally posted by: ol

Hello

you can use one of the two kinds of formulations below. The first one is represented as a floating point, so it is not the exact value of e. The second one is represented as an interval that is guaranted to contain the real value.

Regards,

Olivier

IloNumExpr e=IloNumExpr(env,2.718);

IloNumExpr e=IloExponent(IloNumExpr(env,1.0));

Originally posted by: FiFila

Thanks Olivier;

But I really want to calculate the tanh(X)? What about your idea?

Originally posted by: Petr Vilím

Hello FiFila,

tanh can be computed as ( exp(2*x) - 1 )  / ( exp(2*x) + 1 ).

However it will probably not propagate very well because x is used twice in the expression (that's something better to avoid in CP).

I would first try this expression. Then, if necessary, you may try to approximate tanh by a piecewise linear function. Or you can write your own constraint (in C++).

Petr

Originally posted by: FiFila

Hi Petr;

Thanks for your great help.

Because I have for instance two layers in NN (two layer with tanh), I think this computation encounters more fault. Do you have any idea?

Best

Fila

Originally posted by: Petr Vilím

What do you mean by "more fault"? Less precise result? The computation is done internally in "double" type, so it should be quite precise. Do you have a specific example where it behaves badly?

Best, Petr

Originally posted by: FiFila

Thanks Petr;

Yes, Less precise I mean. No, I have not written my code but I think because of properties of NN  it may be with less precise. Because we multiple tanh several times in input of each layer. This is the reason I have doubt to is CP or CPLEX a good environment to implementation (If yes, how I can reduce the error because of this estimation) or I should switch to others like or-tools, etc?

Originally posted by: Petr Vilím

We care about precision of the computation a lot, the error using exp function should be really minimal (maybe one or two last digits of the sixteen digit mantissa). You can easily try and is it for yourself.

Best, Petr