Programming

Optimisation and Operations Research algorithms

with Julia

Prof. Dr. Xavier Gandibleux

Université de Nantes
Département Informatique – UFR Sciences et Techniques
France

Lesson 7 – May-June 2021

Programming with Julia Xavier [email protected] 1

 Optimisation
JuMP

version 0.21.8 and later

Programming with Julia Xavier [email protected] 2

Overview of JuMP
A modeling language for Mathematical Optimization (linear,
mixed-integer, conic, semidefinite, nonlinear):
§ User friendliness: syntax that mimics natural mathematical
expressions.

§ Speed: similar speeds to special-purpose modeling languages

such as AMPL.

§ Solver independence: JuMP uses MathOptInterface (MOI), an

abstraction layer designed to provide a unified interface to
mathematical optimization solvers.
Currently supported solvers: Artelys Knitro, Baron, Bonmin, Cbc,
Clp, Couenne, CPLEX, FICO Xpress, GLPK, Gurobi, SCIP, etc.

§ Ease of embedding: JuMP itself is written purely in Julia.

Solvers are the only binary dependencies.

Programming with Julia Xavier [email protected] 3

Overview of JuMP
A modeling language for Mathematical Optimization (linear,
mixed-integer, conic, semidefinite, nonlinear):
§ User friendliness: syntax that mimics natural mathematical
expressions.

§ Speed: similar speeds to special-purpose modeling languages

such as AMPL.

§ Solver independence: JuMP uses MathOptInterface (MOI), an

abstraction layer designed to provide a unified interface to
mathematical optimization solvers.
Currently supported solvers: Artelys Knitro, Baron, Bonmin, Cbc,
Clp, Couenne, CPLEX, FICO Xpress, GLPK, Gurobi, SCIP, etc.

§ Ease of embedding: JuMP itself is written purely in Julia.

Solvers are the only binary dependencies.

Programming with Julia Xavier [email protected] 3

Overview of JuMP
A modeling language for Mathematical Optimization (linear,
mixed-integer, conic, semidefinite, nonlinear):
§ User friendliness: syntax that mimics natural mathematical
expressions.

§ Speed: similar speeds to special-purpose modeling languages

such as AMPL.

§ Solver independence: JuMP uses MathOptInterface (MOI), an

abstraction layer designed to provide a unified interface to
mathematical optimization solvers.
Currently supported solvers: Artelys Knitro, Baron, Bonmin, Cbc,
Clp, Couenne, CPLEX, FICO Xpress, GLPK, Gurobi, SCIP, etc.

§ Ease of embedding: JuMP itself is written purely in Julia.

Solvers are the only binary dependencies.

Programming with Julia Xavier [email protected] 3

Overview of JuMP
A modeling language for Mathematical Optimization (linear,
mixed-integer, conic, semidefinite, nonlinear):
§ User friendliness: syntax that mimics natural mathematical
expressions.

§ Speed: similar speeds to special-purpose modeling languages

such as AMPL.

§ Solver independence: JuMP uses MathOptInterface (MOI), an

abstraction layer designed to provide a unified interface to
mathematical optimization solvers.
Currently supported solvers: Artelys Knitro, Baron, Bonmin, Cbc,
Clp, Couenne, CPLEX, FICO Xpress, GLPK, Gurobi, SCIP, etc.

§ Ease of embedding: JuMP itself is written purely in Julia.

Solvers are the only binary dependencies.

Programming with Julia Xavier [email protected] 3

Overview of JuMP
A modeling language for Mathematical Optimization (linear,
mixed-integer, conic, semidefinite, nonlinear):
§ User friendliness: syntax that mimics natural mathematical
expressions.

§ Speed: similar speeds to special-purpose modeling languages

such as AMPL.

§ Solver independence: JuMP uses MathOptInterface (MOI), an

abstraction layer designed to provide a unified interface to
mathematical optimization solvers.
Currently supported solvers: Artelys Knitro, Baron, Bonmin, Cbc,
Clp, Couenne, CPLEX, FICO Xpress, GLPK, Gurobi, SCIP, etc.

§ Ease of embedding: JuMP itself is written purely in Julia.

Solvers are the only binary dependencies.

Programming with Julia Xavier [email protected] 3

Getting started

Install:

using Pkg
Pkg.add("JuMP")

Pkg.add("GLPK")

Setup:

using JuMP

using GLPK

Programming with Julia Xavier [email protected] 4

Getting started

Install:

using Pkg
Pkg.add("JuMP")

Pkg.add("GLPK")

Setup:

using JuMP

using GLPK

Programming with Julia Xavier [email protected] 4

Getting started

Install:

using Pkg
Pkg.add("JuMP")

Pkg.add("GLPK")

Setup:

using JuMP

using GLPK

Programming with Julia Xavier [email protected] 4

Writing a model (1/5)

Example:
» fi
max z px q “ x1 ` 3x2 p0q
—
— s .t x1 ` x2 ď 14 p1q ffi
ffi
—
— ´2x1 ` 3x2 ď 12 p2q ffi
ffi
– 2x1 ´ x2 ď 12 p3q fl
x1 , x2 ě 0 p4q

Programming with Julia Xavier [email protected] 5

Writing a model (2/5)

Creating a Model:

modName = Model(solver)

julia> model = Model(GLPK.Optimizer)

Result:
» fi
— ffi
— ffi
— ffi
– fl

Programming with Julia Xavier [email protected] 6

Writing a model (3/5)

Defining Variables:

@variable(modName, varName definition )

julia> @variable(model, x1 >= 0)

julia> @variable(model, x2 >= 0)

Result:
» fi
— ffi
— ffi
— ffi
– fl
x1 , x2 ě 0 p4q

Programming with Julia Xavier [email protected] 7

Writing a model (4/5)

Defining Objective:

@objective(modName, min/max, objectiveFunction )

julia> @objective(model, Max, x1 + 3x2)

Result:
» fi
max z px q “ x1 ` 3x2 p0q
— ffi
— ffi
— ffi
– fl
x1 , x2 ě 0 p4q

Programming with Julia Xavier [email protected] 8

Writing a model (5/5)

Defining Constraints:

@constraint(modName, cstName, cstDefinition )

julia> @constraint(model, cst1, x1 + x2 <= 14)

julia> @constraint(model, cst2, -2x1 + 3x2 <= 12)
julia> @constraint(model, cst3, 2x1 - x2 <= 12)

Result:
» fi
max z px q “ x1 ` 3x2 p0q
— s.t x1 ` x2 ď 14 p1q ffi
´2x1 ` 3x2 ď 12 p2q
— ffi
— ffi
2x1 ´ x2 ď 12 p3q
– fl
x1 , x2 ě 0 p4q

Programming with Julia Xavier [email protected] 9

Details on

The model:
§ print a summary of the problem:

julia> @show(model)

§ print the formulation of the model:

julia> print(model)

Programming with Julia Xavier [email protected] 10

Solve a model

optimize!(modName )

julia> optimize!(model)

Programming with Julia Xavier [email protected] 11

Querying the solution

julia> @show termination_status(model)

julia> @show primal_status(model)
julia> @show dual_status(model)

julia> @show objective_value(model)

julia> @show value(x1)
julia> @show value(x2)

julia> @show dual(cst1)

julia> @show dual(cst2)
julia> @show dual(cst3)

julia> solution_summary(model)

Programming with Julia Xavier [email protected] 12

Querying the solution

julia> @show termination_status(model)

julia> @show primal_status(model)
julia> @show dual_status(model)

julia> @show objective_value(model)

julia> @show value(x1)
julia> @show value(x2)

julia> @show dual(cst1)

julia> @show dual(cst2)
julia> @show dual(cst3)

julia> solution_summary(model)

Programming with Julia Xavier [email protected] 12

Querying the solution

julia> @show termination_status(model)

julia> @show primal_status(model)
julia> @show dual_status(model)

julia> @show objective_value(model)

julia> @show value(x1)
julia> @show value(x2)

julia> @show dual(cst1)

julia> @show dual(cst2)
julia> @show dual(cst3)

julia> solution_summary(model)

Programming with Julia Xavier [email protected] 12

Querying the solution

julia> @show termination_status(model)

julia> @show primal_status(model)
julia> @show dual_status(model)

julia> @show objective_value(model)

julia> @show value(x1)
julia> @show value(x2)

julia> @show dual(cst1)

julia> @show dual(cst2)
julia> @show dual(cst3)

julia> solution_summary(model)

Programming with Julia Xavier [email protected] 12

Details on
termination_status:

termination_status(modName )

Common return values

§ OPTIMAL:
The algorithm found a globally optimal solution

§ INFEASIBLE:
The algorithm concluded that no feasible solution exists.

§ TIME_LIMIT:
The algorithm stopped after a user-specified computation time.

§ NUMERICAL_ERROR:
The algorithm stopped because a numerical error.

§ etc.

Programming with Julia Xavier [email protected] 13

Details on
termination_status:

termination_status(modName )

Common return values

§ OPTIMAL:
The algorithm found a globally optimal solution

§ INFEASIBLE:
The algorithm concluded that no feasible solution exists.

§ TIME_LIMIT:
The algorithm stopped after a user-specified computation time.

§ NUMERICAL_ERROR:
The algorithm stopped because a numerical error.

§ etc.

Programming with Julia Xavier [email protected] 13

Recommended workflow
For solving a model and querying the solution:

julia> if termination_status(model) == MOI.OPTIMAL

zOpt = objective_value(model)
@printf(" z=%5.2f x1=%5.2f x2=%5.2f \n",
zOpt,
value(x1),
value(x2))
@printf(" u1=%5.2f u2=%5.2f u3=%5.2f \n",
dual(cst1),
dual(cst2),
dual(cst3))
elseif termination_status(model) == DUAL_INFEASIBLE
println("problem unbounded")
elseif termination_status(model) == MOI.INFEASIBLE
println("problem infeasible")
end

Programming with Julia Xavier [email protected] 14

Details on
Variables:
§ by default, the variables are continous and unbounded:

julia> @variable(model, x) # x is free

§ possible to setup lower and/or upper bounds on a variable:

julia> @variable(model, xělb)) # x is bounded

julia> @variable(model, xďub)
julia> @variable(model, lbďxďub)
julia> @variable(model, x==2) # x is fixed

§ possible to specify the type of a variable:

julia> @variable(model, xě0, Int) # x PN

julia> @variable(model, x, Bin) # x P t0, 1u

Programming with Julia Xavier [email protected] 15

Details on
Variables:
§ by default, the variables are continous and unbounded:

julia> @variable(model, x) # x is free

§ possible to setup lower and/or upper bounds on a variable:

julia> @variable(model, xělb)) # x is bounded

julia> @variable(model, xďub)
julia> @variable(model, lbďxďub)
julia> @variable(model, x==2) # x is fixed

§ possible to specify the type of a variable:

julia> @variable(model, xě0, Int) # x PN

julia> @variable(model, x, Bin) # x P t0, 1u

Programming with Julia Xavier [email protected] 15

Details on
Variables:
§ by default, the variables are continous and unbounded:

julia> @variable(model, x) # x is free

§ possible to setup lower and/or upper bounds on a variable:

julia> @variable(model, xělb)) # x is bounded

julia> @variable(model, xďub)
julia> @variable(model, lbďxďub)
julia> @variable(model, x==2) # x is fixed

§ possible to specify the type of a variable:

julia> @variable(model, xě0, Int) # x PN

julia> @variable(model, x, Bin) # x P t0, 1u

Programming with Julia Xavier [email protected] 15

Writing an implicit model (1/5)
Example (cont’d):
» fi
max z px q “ x1 ` 3x2 p0q
—
— s .t x1 ` x2 ď 14 p1q ffi
ffi
—
— ´2x1 ` 3x2 ď 12 p2q ffi
ffi
– 2x1 ´ x2 ď 12 p3q fl
x1 , x2 ě 0 p4q

Ó
¨ ˛
1 1
c “ p1, 3q, d “ p14, 12, 12q, T “ ˝ ´2 3 ‚
2 ´1
» ř2 fi
max z px q “ cx p0q
řj2“1 j j
– s .t j “1 tij xj ď di i “ 1, 3 p1 ´ 3q fl
xj ě 0 j “ 1, 2 p4q

Programming with Julia Xavier [email protected] 16

Writing an implicit model (2/5)

Example (cont’d):
¨ ˛
1 1
c “ p1, 3q, d “ p14, 12, 12q, T “ ˝ ´2 3 ‚
2 ´1

c = [1, 3]
d = [14, 12, 12]
T = [1 1 ; -2 3; 2 -1]
n,m = size(T)

Programming with Julia Xavier [email protected] 17

Writing an implicit model (2/5)

Example (cont’d):
» ř2 fi
max z px q “ cj xj p0q
— řj2“1 ffi
– s.t j “1 tij xj ď di i “ 1, 3 p1 ´ 3q fl
xj ě 0 j “ 1, 2 p4q

julia> md = Model(GLPK.Optimizer)
julia> @variable(md, x[1:m]ě0)
julia> @objective(md, Max,sum(c[j]*x[j] for j=1:m))
julia> @constraint(md, cst[i=1:n],sum(T[i,j]*x[j]
for j=1:m)ďd[i])

Programming with Julia Xavier [email protected] 18

Structured variables in JuMP
Scalar variables:

julia> @variable(md, xě0)

Structured variables:
§ Arrays (one-based integer ranges)

julia> @variable(md, x[1:4]ě0)

julia> @variable(md, x[1:2, 1:2]ě0, Int)

§ DenseAxisArrays (indices are not one-based integer ranges)

julia> @variable(md, x[-4:4]ě0)

julia> @variable(md, y[[:tram,:train]], Bin)

§ SparseAxisArrays (indices do not form a rectangular set)

julia> @variable(model, x[i=1:2, j=i:2])

Programming with Julia Xavier [email protected] 19

Structured variables in JuMP
Scalar variables:

julia> @variable(md, xě0)

Structured variables:
§ Arrays (one-based integer ranges)

julia> @variable(md, x[1:4]ě0)

julia> @variable(md, x[1:2, 1:2]ě0, Int)

§ DenseAxisArrays (indices are not one-based integer ranges)

julia> @variable(md, x[-4:4]ě0)

julia> @variable(md, y[[:tram,:train]], Bin)

§ SparseAxisArrays (indices do not form a rectangular set)

julia> @variable(model, x[i=1:2, j=i:2])

Programming with Julia Xavier [email protected] 19

More on variables

Providing a primal starting solution (a MIP-start; a warmstart)

julia> @variable(model, x, start = 44)

julia> set_start_value(x, 44)

Deleting variables:

julia> delete(model, x)

Programming with Julia Xavier [email protected] 20

More on variables

Providing a primal starting solution (a MIP-start; a warmstart)

julia> @variable(model, x, start = 44)

julia> set_start_value(x, 44)

Deleting variables:

julia> delete(model, x)

Programming with Julia Xavier [email protected] 20

More on the objective

Modify an objective:
call @objective with the new objective function:

julia> @objective(model, Min, 2x)

Modify an objective coefficient:

use set_objective_coefficient:

set_objective_coefficient(model, variable, coef )

julia> set_objective_coefficient(model, x2, 5)

Programming with Julia Xavier [email protected] 21

More on the objective

Modify an objective:
call @objective with the new objective function:

julia> @objective(model, Min, 2x)

Modify an objective coefficient:

use set_objective_coefficient:

set_objective_coefficient(model, variable, coef )

julia> set_objective_coefficient(model, x2, 5)

Programming with Julia Xavier [email protected] 21

More on constraints
Modify the coefficient of a variable for a given constraint:

set_normalized_coefficient(cstID, var, coef )

julia> set_normalized_coefficient(model, x2, 5)

Set the right-hand side term of a constraint to a value:

set_normalized_rhs(constraintID, value )

julia> set_normalized_rhs(cst2, 10)

Delete a constraint from a model:

delete(model, constraintID )

julia> delete(model, cst2)

Programming with Julia Xavier [email protected] 22
More on constraints
Modify the coefficient of a variable for a given constraint:

set_normalized_coefficient(cstID, var, coef )

julia> set_normalized_coefficient(model, x2, 5)

Set the right-hand side term of a constraint to a value:

set_normalized_rhs(constraintID, value )

julia> set_normalized_rhs(cst2, 10)

Delete a constraint from a model:

delete(model, constraintID )

julia> delete(model, cst2)

Programming with Julia Xavier [email protected] 22
More on constraints
Modify the coefficient of a variable for a given constraint:

set_normalized_coefficient(cstID, var, coef )

julia> set_normalized_coefficient(model, x2, 5)

Set the right-hand side term of a constraint to a value:

set_normalized_rhs(constraintID, value )

julia> set_normalized_rhs(cst2, 10)

Delete a constraint from a model:

delete(model, constraintID )

julia> delete(model, cst2)

Programming with Julia Xavier [email protected] 22
More on solutions
Primal solution values for a scalar variable:
value(variable )

julia> value(x1)

Primal solution values for a structured variable:

value(variable [index ])

julia> value(x[3])

value.(variable )

julia> value.(x)

Programming with Julia Xavier [email protected] 23

More on solutions
Primal solution values for a scalar variable:
value(variable )

julia> value(x1)

Primal solution values for a structured variable:

value(variable [index ])

julia> value(x[3])

value.(variable )

julia> value.(x)

Programming with Julia Xavier [email protected] 23

 Optimisation
vOptSolver

Programming with Julia Xavier [email protected] 24

Overview of vOptSolver
An ecosystem for modeling and solving multiobjective linear
optimization problems (MOCO, MOIP, MOMIP, MOLP):
§ it deals with structured and non-structured optimization problems with at
least two objectives
§ it integrates several specific and generic exact algorithms for computing
efficient solutions
§ Natural and intuitive use for mathematicians, informaticians, engineers

§ Efficient, flexible, evolutive solver

§ Aims to be easy to formulate a problem, to provide data, to solve a

problem, to collect the outputs, to analyze the solutions
§ Free, open source (MIT licence), multi-platform, reusing existing
specifications
§ Using usual free (GLPK, Clp/Cbc) and commercial (GUROBI, CPLEX)
MILP solvers

Programming with Julia Xavier [email protected] 25

Overview of vOptSolver
An ecosystem for modeling and solving multiobjective linear
optimization problems (MOCO, MOIP, MOMIP, MOLP):
§ it deals with structured and non-structured optimization problems with at
least two objectives
§ it integrates several specific and generic exact algorithms for computing
efficient solutions
§ Natural and intuitive use for mathematicians, informaticians, engineers

§ Efficient, flexible, evolutive solver

§ Aims to be easy to formulate a problem, to provide data, to solve a

problem, to collect the outputs, to analyze the solutions
§ Free, open source (MIT licence), multi-platform, reusing existing
specifications
§ Using usual free (GLPK, Clp/Cbc) and commercial (GUROBI, CPLEX)
MILP solvers

Programming with Julia Xavier [email protected] 25

Getting started with vOptGeneric

Install:

using Pkg
Pkg.add("vOptGeneric")

Pkg.add("GLPK")

Setup:

using vOptGeneric

using GLPK

Programming with Julia Xavier [email protected] 26

Getting started with vOptGeneric

Install:

using Pkg
Pkg.add("vOptGeneric")

Pkg.add("GLPK")

Setup:

using vOptGeneric

using GLPK

Programming with Julia Xavier [email protected] 26

Example with vOptGeneric

For the bi-objective unidimensional 01 knapsack problem,

(
max pp1 x , p2 x q | wx ď c , x P t0, 1un

with1 n“5

p1 “ p10, 3, 6, 8, 2q

p2 “ p12, 9, 11, 5, 6q

w “ p4, 5, 2, 5, 6q

c “ 17

compute YN , the set of non-dominated points using the -constraint

method.

1
exercise 10.2, page 290 of Multicriteria Optimization (2nd edt), M. Ehrgott,
Springer 2005

Programming with Julia Xavier [email protected] 27

Setup the data

julia> p1 = [10,3,6,8,2] # coef vct of the obj 1

julia> p2 = [12,9,11,5,6] # coef vct of the obj 2
julia> w = [4,5,2,5,6] # coef vct of weights
julia> c = 17 # nominal capacity
julia> n = length(p1) # number of items

Programming with Julia Xavier [email protected] 28

Setup the model

julia> kp = vModel( GLPK.Optimizer )

julia> @variable(kp, x[1:n],Bin)
julia> @addobjective(kp,Max,sum(p1[j]*x[j] for j=1:n))
julia> @addobjective(kp,Max,sum(p2[j]*x[j] for j=1:n))
julia> @constraint(kp,sum(w[j]*x[j] for j=1:n) ď c)

Programming with Julia Xavier [email protected] 29

Solve and display results
Invoking the solver (dichotomic method):

julia> vSolve(kp,method=:epsilon,step=0.5,verbose=false)

Querying the results:

julia> Y_N = getY_N(kp)

Displaying the results (X_SE and Y_SN):

julia> for i = 1:length(Y_N)

julia> X = value.(x, i)
julia> print("X = ", findall(elt -> elt « 1, X))
julia> println(" | Z = ",Y_N[i])
julia> end

julia> printX_E(kp)

Programming with Julia Xavier [email protected] 30

Plot results

julia> using PyPlot

julia> z1, z2 = map(x -> x[1],Y_N), map(x -> x[2],Y_N)
julia> PyPlot.title("Knapsack")
julia> PyPlot.xlabel("$z_1$ to maximize")
julia> PyPlot.ylabel("$z_2$ to maximize")
julia> grid()
julia> plot(z1,z2,"bx",markersize="8",label="$Y_N$")
julia> legend(loc=1,fontsize="small")
julia> show()

Programming with Julia Xavier [email protected] 31

 Review and exercises

(notebook)

Programming with Julia Xavier [email protected] 32

Programming with Julia Xavier [email protected] 33