Couenne in Minizinc is unable to find an optimal solution

[indrajitsg]

Hi All,

I downloaded the latest version of Couenne available from the AMPL website and was testing on a very small problem which I am posting here. If I run the solver using AMPL, it gives the correct optimal solution. However if I run the same model through Minizinc, the solution that I get is sub-optimal.

Minizinc Model: model.mzn

int: nMOVES;
set of int: MOVES = 1..nMOVES;
int: timeBound;
array[MOVES] of int: power;
array[MOVES] of int: duration;
array[MOVES] of var 0..1: occur;
var int: totalPower = sum(i in MOVES) (power[i] * occur[i]);

constraint (sum(i in MOVES) (duration[i] * occur[i])) <= timeBound;

solve maximize totalPower;

output ["Moves to train: \(occur)\n" ++ "Power: \(totalPower)"];

Minizinc Data: data.dzn

nMOVES = 5;
timeBound = 10;
power = [6, 8, 5, 3, 4];
duration = [4, 5, 3, 2, 3];

AMPL Model: model.mod

set MOVES;
param timeBound > 0;
param power {MOVES} > 0;
param duration {MOVES} > 0;
var occur{i in MOVES} binary;

maximize totalPower: sum{i in MOVES} power[i] * occur[i];

subject to constTime: sum{i in MOVES} duration[i] * occur[i] <= timeBound;

AMPL Data: data.dat

data;

set MOVES := M1 M2 M3 M4 M5;

param timeBound:= 10;

param: power duration :=
     M1     6     4
     M2     8     5
     M3     5     3
     M4     3     2
     M5     4     3;

[merraksh]

Hi Indrajit,   The model is linear binary, so Couenne can solve it to global optimality although it is "overqualified" for this---Cbc would also find the optimum. I don't know the reason why Minizinc does not find the optimum. Does Minizinc interface with Couenne, and if so, how (I was unaware of any such interface)? Is there a way to debug the Minizinc/Couenne interface so that we get a .nl, .mps, or .lp file of the Minizinc model?   Pietro   Sent: Tuesday, January 12, 2021 at 3:49 AM From: "Indrajit Sen Gupta" <notifications@github.com> To: "coin-or/Couenne" <Couenne@noreply.github.com> Cc: "Subscribed" <subscribed@noreply.github.com> Subject: [coin-or/Couenne] Couenne in Minizinc is unable finds a sub-optimal solution (#53)   Hi All, I downloaded the latest version of Couenne available from the AMPL website and was testing on a very small problem which I am posting here. If I run the solver using AMPL, it gives the correct optimal solution. However if I run the same model through Minizinc, the solution that I get is sub-optimal. Minizinc Model: model.mzn int: nMOVES; set of int: MOVES = 1..nMOVES; int: timeBound; array[MOVES] of int: power; array[MOVES] of int: duration; array[MOVES] of var 0..1: occur; var int: totalPower = sum(i in MOVES) (power[i] * occur[i]); constraint (sum(i in MOVES) (duration[i] * occur[i])) <= timeBound; solve maximize totalPower; output ["Moves to train: \(occur)\n" ++ "Power: \(totalPower)"]; Minizinc Data: data.dzn nMOVES = 5; timeBound = 10; power = [6, 8, 5, 3, 4]; duration = [4, 5, 3, 2, 3]; AMPL Model: model.mod set MOVES; param timeBound > 0; param power {MOVES} > 0; param duration {MOVES} > 0; var occur{i in MOVES} binary; maximize totalPower: sum{i in MOVES} power[i] * occur[i]; subject to constTime: sum{i in MOVES} duration[i] * occur[i] <= timeBound; AMPL Data: data.dat data; set MOVES := M1 M2 M3 M4 M5; param timeBound:= 10; param: power duration := M1 6 4 M2 8 5 M3 5 3 M4 3 2 M5 4 3; — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub, or unsubscribe.

[indrajitsg]

Hi @merraksh , yes you are correct. Couenne is an overkill for such a simple problem. The way we interface Minizinc with any of the solvers available from AMPL website is to specify in the configuration to use NL format. So Minizinc must be first converting it to NL format and then solving. Now I have tested the same model with many of the solvers including CBC, Bonmin, CPLEX, Gurobi and Baron. All of them are giving correct results. Hence the model translation is not an issue in my view.

The configuration file (couenne.msc) for Couenne (or any other solver from AMPL) needs to be placed in the directory: \MiniZinc\share\minizinc\solvers. The config file should have the following information:

{
  "id": "org.coin-or.couenne",
  "name": "Couenne",
  "description": "Couenne executable",
  "version": "0.5.7",
  "executable": "../../../bin/couenne-win64/couenne",
  "supportsFzn": false,
  "supportsNL": true
}

Regards,
Indrajit

[merraksh]

I believe the next step is to compare the .nl files produced by AMPL and by Minizinc. Maybe run Couenne by adding the option "problem_print_level 4" in the option file couenne.opt. If there are substantial differences in the .nl files this should expose them.     Sent: Thursday, January 14, 2021 at 5:56 PM From: "Indrajit Sen Gupta" <notifications@github.com> To: "coin-or/Couenne" <Couenne@noreply.github.com> Cc: "Pietro Belotti" <petr7b6@gmail.com>, "Mention" <mention@noreply.github.com> Subject: Re: [coin-or/Couenne] Couenne in Minizinc is unable to find an optimal solution (#53)   Hi @merraksh , yes you are correct. Couenne is an overkill for such a simple problem. The way we interface Minizinc with any of the solvers available from AMPL website is to specify in the configuration to use NL format. So Minizinc must be first converting it to NL format and then solving. Now I have tested the same model with many of the solvers including CBC, Bonmin, CPLEX, Gurobi and Baron. All of them are giving correct results. Hence the model translation is not an issue in my view. The configuration file (couenne.msc) for Couenne (or any other solver from AMPL) needs to be placed in the directory: \MiniZinc\share\minizinc\solvers. The config file should have the following information: { "id": "org.coin-or.couenne", "name": "Couenne", "description": "Couenne executable", "version": "0.5.7", "executable": "../../../bin/couenne-win64/couenne", "supportsFzn": false, "supportsNL": true } Regards, Indrajit — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or unsubscribe.

[kistamas]

Hi,
I have checked in detail this example, and it seems that the problem is with the interpretation of the solution by the interface between minizinc and couenne.
The solution file is

3
1
1
0
2
0
6
6
0
0.9999999999999999
1
1
0
16
objno 0 3

The value of the 5 decision variables are in lines -6 .. -2, and the second one is 0.99999. So, my guess is that the number 0.999999 rounded to 0, which is a problem.
Regards,
Tamas

[merraksh]

Hi Tamás,   indeed Couenne doesn't refine the solution to round off binary variables (I'm not sure if Cbc, the underlying branch-and-bound, does). so rather than floor() one should use a rounding function.   Regards, Pietro   Sent: Thursday, January 21, 2021 at 1:12 PM From: "Tamas Kis" <notifications@github.com> To: "coin-or/Couenne" <Couenne@noreply.github.com> Cc: "Pietro Belotti" <petr7b6@gmail.com>, "Mention" <mention@noreply.github.com> Subject: Re: [coin-or/Couenne] Couenne in Minizinc is unable to find an optimal solution (#53)   Hi, I have checked in detail this example, and it seems that the problem is with the interpretation of the solution by the interface between minizinc and couenne. The solution file is 3 1 1 0 2 0 6 6 0 0.9999999999999999 1 1 0 16 objno 0 3 The value of the 5 decision variables are in lines -6 .. -2, and the second one is 0.99999. So, my guess is that the number 0.999999 rounded to 0, which is a problem. Regards, Tamas — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or unsubscribe.

[indrajitsg]

Hi @merraksh , I have placed the couenne.opt file in the same location as couenne.exe. But I don't think it is being used by couenne. What should I do? This is when I am running through Minizinc.

[merraksh]

I think the couenne.opt file should be placed in the same folder where the main command is run from, i.e. not necessarily where the Couenne executable is but rather the current working directory.     Sent: Thursday, January 21, 2021 at 2:30 PM From: "Indrajit Sen Gupta" <notifications@github.com> To: "coin-or/Couenne" <Couenne@noreply.github.com> Cc: "Pietro Belotti" <petr7b6@gmail.com>, "Mention" <mention@noreply.github.com> Subject: Re: [coin-or/Couenne] Couenne in Minizinc is unable to find an optimal solution (#53)   Hi @merraksh , I have placed the couenne.opt file in the same location as couenne.exe. But I don't think it is being used by couenne. What should I do? — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or unsubscribe.

[indrajitsg]

@merraksh - thanks it is working now.