Solving bilinear optimization problem using Alpine.jl

Shuvomoy Das Gupta

August 30, 2022

In this blog, we will discuss how to solve bilinear optimizations problem using Alpine.

Table of contents

  1. What is Alpine?
  2. Example in consideration
  3. Construct the optimization model using JuMP
  4. Setting local and global subsolvers in Alpine
  5. Solve the problem using Alpine
  6. Solving the problem using Gurobi
  7. Getting tightened bounds on the variables using Alpine
    1. The first way
    2. The second way
  8. References

What is Alpine?

Alpine is an open-source Julia package to solve mixed-integer nonlinear optimization problems (MINLPs) to global optimality. It is an open-source implementation of the Adaptive Multivariate Partitioning (AMP) Algorithm proposed by Nagarajan et al in [Nagarajan2019]. AMP uses an adaptive, piecewise convexification scheme and constraint programming methods to solve MINLPs to global optimality. The benefit of Alpine over other spatial branch-and-bound solvers is that it is entirely built upon JuMP [Dunning2017] and MathOptInterface [Legat2021] in Julia, which provides significant flexibility and modeling power.

Example in consideration

We want to solve the following bilinear problem:

maximizex1xd−x2xd−1subject to∑i=1dxi≤10xixi+1≤2,i=1,…,d−1∑i=1d−1xixi+1=1x≥0,(BLP) \begin{array}{ll} \textrm{maximize} & x_{1}x_d - x_2 x_{d-1}\\ \textrm{subject to} & \sum_{i=1}^{d}x_{i}\leq10\\ & x_{i}x_{i+1}\leq2,\quad i=1,\ldots,d-1\\ & \sum_{i=1}^{d-1}x_{i}x_{i+1}=1\\ & x\geq0, \end{array} \quad \textrm{(BLP)}

where x∈Rdx\in\mathbf{R}^d is the decision variables. For this example, we let d=10d=10

Now, we will solve (BLP) step by step in Alpine.

Construct the optimization model using JuMP

First, we load the packages.

## Load packages
using Alpine, JuMP, Gurobi, Ipopt

Now, we create a function that will construct the optimization model in JuMP. For the optimization model to be classified as an MINLP, we need at least one nonlinear constraint or objective declared specifically in the JuMP model. This can be achieved by using @NLobjective or @NLconstraint macro.

function BLPSolver(; solver = AlpineSolver) 
# solver is either AlpineSolver or GurobiSolver (defined below)

    d = 10
  
    # declare the JuMP model
    # ----------------------
    m = Model(solver)

    # declare the variable x ≧ 0
    # ----------------------
    @variable(m, x[1:d] >= 0) 
  
    # declare the constraints
    # -----------------------

    # add objective x[1] x[d] - x[2] x[d-1], which is to be maximized

    if solver ==  GurobiSolver
        @objective(m, Max, x[1] * x[d] - x[2] * x[d-1])
    elseif solver == AlpineSolver
        @NLobjective(m, Max, x[1] * x[d] - x[2] * x[d-1])
    else 
       @error "correct solver type not given"
    end

    # add the constraints
    # -------------------

    # constraint ∑_{i=1:d} x[i] ≦ 10, conSum is the name of the constraint
    @constraint(m, conSum, sum(x[i] for i in 1:d) <= 10) 

    # add the constraints  (∀i ∈ [1:d-1]) x[i] x[i+1] ≤ 2, 
    # where the i-the constraint is called congBilinear1[i]
    @constraint(m, conBilinear1[i=1:d-1], x[i] * x[i+1] <= 2)

    # add constraint ∑_{i ∈ [1:d-1]} x[i] x[i+1] = 1, we call it 
    @constraint(m, conBilinear2, sum(x[i] * x[i+1] for i in 1:d-1) == 1)
  
    # return the JuMP model
    # ---------------------

    return m

end

Setting local and global subsolvers in Alpine

Now we will solve this problem using Alpine. To implement the MAP algorithm, which is the core algorithm in Alpine , we need:

(i) One local solver, which usually implements some variant of the nonlinear interior point algorithm to find a locally optimal solution to the MINLP. Solvers of this type that are supported by Alpine via MathOptInterface are: Ipopt and KNITRO.

(ii) One mixed-integer optimization solver, that usually implements some sort of branch-and-bound algorithm to solve mixed-integer optimization problems (MIPs). During every iteration of the MAP algorithm, Alpine solves an MIP sub-problem. Because MIPs are usually hard to solve, Alpine's runtime heavily depends on the runtime of these solvers. The MIP solvers that are supported by Alpine are: Cbc, CPLEX, Gurobi, and Bonmin. In my numerical experiments, I have found Gurobi to be the fastest and the Alpine developers also recommend Gurobi as well.

We set the local and global subsolvers, followed by setting Alpine as the global solver as follows. 

# function to set local subsolver as Ipopt
# ----------------------------------------
IpoptSolverAlpine = JuMP.optimizer_with_attributes(
                                      Ipopt.Optimizer,
                                      MOI.Silent() => true
                                      )

# function to set global subsolver as Gurobi
# ------------------------------------------
GRB_ENV  = Gurobi.Env()

GurobiSolverAlpine = JuMP.optimizer_with_attributes(
                                        () -> Gurobi.Optimizer(GRB_ENV),
                                        MOI.Silent() => true,
                                        "Presolve" => 0
                                        )

# function to set Alpine as the global solver
# -------------------------------------------
AlpineSolver = JuMP.optimizer_with_attributes(
                                              Alpine.Optimizer,
                                              "nlp_solver"   => IpoptSolverAlpine, # local solver
                                              "mip_solver"   => GurobiSolverAlpine, # global solver
                                              "presolve_bt"  => true, 
                                              # "partition_scaling_factor"   => 10,
                                              "apply_partitioning" => true,
                                              "log_level" => 1
                                              )

Define a Gurobi solver to solve the bilinear problem to global optimality. 

GurobiSolver = JuMP.optimizer_with_attributes(
                                        () -> Gurobi.Optimizer(GRB_ENV),
                                        # MOI.Silent() => true,
                                        "NonConvex" => 2 
                                        # means we are going to use Gurobi's spatial branch-and-bound algorithm
                                        )

The options used in declaring the AlpineSolver solver above are as follows (Source: https://lanl-ansi.github.io/Alpine.jl/latest/parameters/)

"nlp_solver" => IpoptSolverAlpine sets Ipopt as the local solver

"mip_solver" => GurobiSolverAlpine sets Gurobi as the global solver

"presolve_bt" => true performs sequential optimization-based bound tightening (OBBT) at the presolve step. For more details about sequential OBBT, see Section 3.1.1 of https://arxiv.org/pdf/1707.02514.pdf.

partition_scaling_factor => 10 is used to measure the width of new partitions relative to the active partition chosen in the sequentially solved lower-bounding MIPs. This value can substantially affect the run time for global convergence; this value can be set to different integer values (>= 4) for various classes of problems.

"apply_partitioning" => true applies Alpine's built-in MIP-based partitioning algorithm only (MAP) when activated; else terminates with the presolve solution.

"log_level" => 100 enables detailed debugging mode of Alpine. The option log_level (default = 0) controls the verbosity level of Alpine output; choose 1 for turning on logging, else 100 for detailed debugging mode.

Solve the problem using Alpine

Now we solve the problem using Alpine.

m = BLPSolver(solver = AlpineSolver)

JuMP.optimize!(m)

The output is as follows.

PROBLEM STATISTICS
  Objective sense = Max
  # Variables = 10
  # Bin-Int Variables = 0
  # Constraints = 11
  # NL Constraints = 10
  # Linear Constraints = 1
  # Detected convex constraints = 0
  # Detected nonlinear terms = 11
  # Variables involved in nonlinear terms = 10
  # Potential variables for partitioning = 10
SUB-SOLVERS USED BY ALPINE
  NLP local solver = Ipopt
  MIP solver = Gurobi
ALPINE CONFIGURATION
  Maximum iterations (upper-bounding MIPs) =  99
  Relative global optimality gap = 0.01%
  Discretization ratio = 10
  Bound-tightening presolve = true
  Maximum iterations (OBBT) = 25
PRESOLVE 
  Doing local search
  Local solver returns a feasible point with value 23.9792
  Starting bound-tightening
  Actual iterations (OBBT): 9
  Post-presolve optimality gap: 0.569%
  Completed presolve in 0.33s
UPPER-BOUNDING ITERATIONS
====================================================================================================
| Iter   | Incumbent       | Best Incumbent      | Upper Bound        | Gap (%)         | Time      
| 1      | 23.9792         | 23.9792             | 24.009             | 0.124           | 0.34s            
| 2      | 24.0            | 24.0                | 24.009             | 0.038           | 0.38s            
| finish | 24.0            | 24.0                | 24.0021            | 0.009           | 0.46s            
====================================================================================================

## *** Alpine ended with status OPTIMAL *** ##

Solving the problem using Gurobi

Alternatively, we can also solve the problem in consideration to global optimality using Gurobi as follows.

Now solve the problem. 

mGurobi = BLPSolver(solver = GurobiSolver);

JuMP.optimize!(mGurobi)

Output is as follows.

Presolve time: 0.00s
Presolved: 61 rows, 22 columns, 120 nonzeros
Presolved model has 20 bilinear constraint(s)
Variable types: 22 continuous, 0 integer (0 binary)

Root relaxation: objective 4.955556e+01, 22 iterations, 0.00 seconds (0.00 work units)

    Nodes    |    Current Node    |     Objective Bounds      |     Work
 Expl Unexpl |  Obj  Depth IntInf | Incumbent    BestBd   Gap | It/Node Time

     0     0   49.55556    0   11          -   49.55556      -     -    0s
H    0     0                      23.9791576   49.55556   107%     -    0s
     0     0   24.99322    0   10   23.97916   24.99322  4.23%     -    0s
     0     0   24.99231    0    8   23.97916   24.99231  4.23%     -    0s
     0     0   24.99029    0   10   23.97916   24.99029  4.22%     -    0s
     0     0   24.99026    0   11   23.97916   24.99026  4.22%     -    0s
     0     2   24.99026    0   11   23.97916   24.99026  4.22%     -    0s
H    5     6                      23.9946925   24.95311  3.99%  11.0    0s
*   37    10               7      23.9961128   24.05999  0.27%   6.5    0s
*   45    12               8      23.9972011   24.01419  0.07%   5.6    0s
*   50    12               9      23.9980299   24.01286  0.06%   5.3    0s
*   57    12               9      23.9998195   24.01286  0.05%   4.8    0s
H   64    14                      23.9999797   24.01166  0.05%   4.3    0s
H  141    65                      23.9999994   24.00776  0.03%   2.2    0s
*  172    47              24      24.0000004   24.00776  0.03%   1.9    0s
*  185    47              24      24.0000004   24.00776  0.03%   1.7    0s

Cutting planes:
  RLT: 13

Explored 562 nodes (382 simplex iterations) in 0.05 seconds (0.01 work units)
Thread count was 10 (of 10 available processors)

Solution count 10: 24 24 24 ... 23.9792

Optimal solution found (tolerance 1.00e-04)
Best objective 2.399999999991e+01, best bound 2.400042876397e+01, gap 0.0018%

User-callback calls 1347, time in user-callback 0.00 sec

Note that we have the same optimal value.

Getting tightened bounds on the variables using Alpine

In certain applications, we may be interested in computing valid but tightened bounds on the variables rather than finding an optimal solution. One nice feature of Alpine is that, just by using its sequential OBBT algorithm, we can compute valid tightened bounds on the variables without solving the full problem. We can achieve this in two ways.

  • The first way is to set "apply_partitioning" => false, which will terminate with the presolve solution but with tighter bounds.

  • The second way is keeping "apply_partitioning" => true, but setting the option "rel_gap" => δ, where δ is a number larger than or equal to the the default relative gap 1e-4. This will lead to Alpine terminating at the relative gap δ, so with a suboptimal solution with optimality gap δ%, but with significantly tighter bounds on the variables.

For example, by inspecting BLP, we see that 0≤x≤100 \leq x \leq 10. We will now show how to tighten these bounds significantly by using the sequential OBBT feature of the AMP algorithm that is implemented in Alpine.

The first way

## Set Alpine as the global solver but to compute tighter bounds via its presolve phase
AlpineSolver = JuMP.optimizer_with_attributes(
                                               Alpine.Optimizer,
                                              "nlp_solver"   => IpoptSolverAlpine, # local solver
                                              "mip_solver"   => GurobiSolverAlpine, # global solver
                                              "presolve_bt"  => true, 
                                              "partition_scaling_factor"  => 10,
                                              "apply_partitioning" => false,
                                              "log_level" => 1
                                              )

Solve the bound tightening problem and get the tightened bounds on the variables.

mOBBT1 = BLPSolver(solver = AlpineSolver)

JuMP.optimize!(mOBBT1)

Let us get the lower bound and upper bound on our decision variable x.

vars = all_variables(mOBBT1) # 

lowerBoundVar = MOI.get.(mOBBT1, Alpine.TightenedLowerBound(), vars) 

upperBoundVar = MOI.get.(mOBBT1, Alpine.TightenedUpperBound(), vars)

@info "[☼ ] Bound on the decision variable x is:"

for i in eachindex(vars)
    @info "$(lowerBoundVar[i]) <=  x[$(i)] <= $(upperBoundVar[i])"
end

The output is:

[ Info: [☼ ] Bound on the decision variable x is:

[ Info: 4.6326 <=  x[1] <= 5.174300000000001
[ Info: 0.0 <=  x[2] <= 0.2096
[ Info: 0.0 <=  x[3] <= 0.030600000000000002
[ Info: 0.0 <=  x[4] <= 0.030000000000000002
[ Info: 0.0 <=  x[5] <= 0.030000000000000002
[ Info: 0.0 <=  x[6] <= 0.030000000000000002
[ Info: 0.0 <=  x[7] <= 0.030000000000000002
[ Info: 0.0 <=  x[8] <= 0.030600000000000002
[ Info: 0.0 <=  x[9] <= 0.2096
[ Info: 4.633100000000001 <=  x[10] <= 5.1742

, which is significantly tighter than 0≤x≤100 \leq x \leq 10.

The second way 

## Set Alpine as the global solver but to compute tighter bounds
AlpineSolver = JuMP.optimizer_with_attributes(
                                               Alpine.Optimizer,
                                              "nlp_solver"   => IpoptSolverAlpine, # local solver
                                              "mip_solver"   => GurobiSolverAlpine, # global solver
                                              "presolve_bt"  => true, 
                                              "partition_scaling_factor"   => 10,
                                              "apply_partitioning" => true,
                                              "rel_gap" => 5e-2, # it means that we will stop when we have relative gap [(UpperBound-LowerBound)/UpperBound] <= 50%, because our objective is just getting tightened bounds on the variables rather than finding globally optimal solution
                                              "log_level" => 1
                                              )

Solve the bound tightening problem and get the tightened bounds on the variables.

mOBBT2 = BLPSolver(solver = AlpineSolver)

JuMP.optimize!(mOBBT2)

Let us get the lower bound and upper bound on our decision variable x.

vars = all_variables(mOBBT2) # 

lowerBoundVar = MOI.get.(mOBBT2, Alpine.TightenedLowerBound(), vars) 

upperBoundVar = MOI.get.(mOBBT2, Alpine.TightenedUpperBound(), vars)

@info "[☼ ] Bound on the decision variable x is:"

for i in eachindex(vars)
    @info "$(lowerBoundVar[i]) <=  x[$(i)] <= $(upperBoundVar[i])"
end

The output is:

[ Info: [☼ ] Bound on the decision variable x is:
[ Info: 3.8099000000000003 <=  x[1] <= 5.7202
[ Info: 0.0 <=  x[2] <= 0.2625
[ Info: 0.0 <=  x[3] <= 0.334
[ Info: 0.0 <=  x[4] <= 0.329
[ Info: 0.0 <=  x[5] <= 0.329
[ Info: 0.0 <=  x[6] <= 0.329
[ Info: 0.0 <=  x[7] <= 0.329
[ Info: 0.0 <=  x[8] <= 0.3508
[ Info: 0.0 <=  x[9] <= 0.264
[ Info: 4.341 <=  x[10] <= 5.7131

, which is significantly tighter than 0≤x≤100 \leq x \leq 10.

References

[Nagarajan2019] Nagarajan, Harsha, et al. "An adaptive, multivariate partitioning algorithm for global optimization of nonconvex programs." Journal of Global Optimization 74.4 (2019): 639-675.

[Legat2021] Legat, Benoît, et al. "MathOptInterface: a data structure for mathematical optimization problems." INFORMS Journal on Computing 34.2 (2022): 672-689.

[Dunning2017] Dunning, Iain, Joey Huchette, and Miles Lubin. "JuMP: A modeling language for mathematical optimization." SIAM review 59.2 (2017): 295-320.

© Shuvomoy Das Gupta. Last modified: October 11, 2024. Website built with Franklin.jl and the Julia programming language.