Use a Multi-Objective Approach

In this article we will compare a multi-objective approach to separately solving single objectives.

Let’s take for an example a problem trying to find a healthy diet for a reasonable price. First let’s define our objectives:

  • Minimize calories: I’m trying to lose some weight, so my healthy diet should be low in calories.

  • Minimize price: A reasonable price means that it doesn’t have to be the absolute minimum, but should be within a close range.

Of course, I want to include minimum/maximum requirements for nutrients such as proteins.

Multi-Objective Approach

Let’s use the multi-objective feature available since AIMMS 4.65 and CPLEX 12.9.

 1Procedure SolveMultiObj {
 2    Body: {
 3        ep_GMP := gmp::Instance::Generate( DietProblem );
 4
 5        p_retcode := GMP::Column::SetAsMultiObjective(
 6            GMP      :  ep_GMP,
 7            column   :  TotalCost,
 8            priority :  2,
 9            weight   :  1,
10            abstol   :  0,
11            reltol   :  0.1);
12        if not p_retcode then raise error "Unable to set TotalCost as an objective" ; endif ;
13
14        p_retcode := GMP::Column::SetAsMultiObjective(
15            GMP      :  ep_GMP,
16            column   :  TotalCalories,
17            priority :  1,
18            weight   :  1,
19            abstol   :  0,
20            reltol   :  0.0);
21        if not p_retcode then raise error "Unable to set TotalCalories as an objective" ; endif ;
22
23        GMP::Instance::Solve( ep_GMP );
24    }
25}

There are several remarks on the above:

  1. Because the TotalCost objective has a higher priority value, it will be solved first.

  2. Because the reltol argument on line 11 has value 0.1, subsequent solves will not increase the total cost by more than 10%.

Single Objective Approach

To be able to compare the multi-objective procedure to traditional single objective solves, there are also two solution procedures in this application:

  1. MainExecution, a traditional single objective solve minimizing total cost.

  2. SolveMinCalo, a traditional single objective solve minimizing total calories.

Comparing Results

The application has both WinUI and WebUI interfaces, but WebUI is featured in the following screenshots.

../../_images/mintotcost.png

Fig. 35 Results after minimizing total cost

../../_images/mintotcalo.png

Fig. 36 Results after minimizing total calories

../../_images/multiobj.png

Fig. 37 Results after running multi objective

The objectives are summarized in the table below:

Cost

Calories

Minimize total cost

21.85

2698.40

Minimize total calories

29.80

2546.40

Multi objective

23.85

2576.40

As you can see, the multi-objective values are not as good as either of the individual objective values, but the multi-objective approach provides a good balance.

You may also find the CPLEX log interesting, as found in the file log/cplex 12.9.sta. (See also Retrieve Solver Log Files in AIMMS Developer.)

 1Solve problem 'DietProblem' with 9 rows, 15 columns (0 binaries, 9 generals), and 83 nonzeros.
 2
 3MIP - Integer optimal solution:  Objective = 2.1849999998e+01
 4Solution time = 0.09 sec.  Iterations = 20  Nodes = 0
 5
 6
 7Solve problem 'MinCaloDietProblem' with 9 rows, 15 columns (0 binaries, 9 generals), and 83 nonzeros.
 8
 9MIP - Integer optimal solution:  Objective = 2.5464000000e+03
10Solution time = 0.02 sec.  Iterations = 13  Nodes = 0
11
12
13Solve problem 'DietProblem' with 8 rows, 14 columns (0 binaries, 9 generals), and 73 nonzeros.
14
15Multi-objective solve log . . .
16
17Starting optimization #1 with priority 2.
18
19
20
21Finished optimization #1 with priority 2.
22Objective =  2.1849999998e+01,  Nodes = 0,  Time = 0.05 sec. (0.45 ticks)
23Cumulative statistics:  Nodes = 0,  Time 0.05 sec. (0.45 ticks)
24
25
26Starting optimization #2 with priority 1.
27
28
29
30Finished optimization #2 with priority 1.
31Objective =  2.5764000000e+03,  Nodes = 0,  Time = 0.02 sec. (0.43 ticks)
32Cumulative statistics:  Nodes = 0,  Time 0.06 sec. (0.89 ticks)
33
34
35
36MIP - Multi-objective optimal
37Solution time = 0.08 sec.  Iterations = 43  Nodes = 0

A breakdown of above log:

  1. lines 1 - 4 for the first solve (minimize total cost)

  2. lines 7 - 10 for the second solve (minimize total calories)

  3. lines 13 - 37 for the multi objective solve.

Project download

The project can be downloaded below.