Previous Up Next

16.4  Repeated Solving of an Eplex Problem

Part of the power of using the eplex library comes from being able to solve an eplex problem repeatedly after modification. For example, we can solve the original transportation problem, add the extra constraint, and resolve the problem. Remember that as eplex_solve/1 instantiates its argument, we need to use a new variable for each call:

        .... % setup the constraints for the original problem as before
        prob: (A3 + B3 + C3 + D3 =< 40),

        prob: eplex_solver_setup(min(....)), % as before

        prob: eplex_solve(Cost1),     % h. solve original problem
        prob: (A1 $= A2),
        prob: eplex_solve(Cost2),     % i. solve modified problem

Note that posted constraints behave logically: they are added to an eplex instance when posted, and removed when they are backtracked over.

In the examples so far, the solver has been invoked explicitly. However, the solver can also behave like a normal constraint, i.e. it is automatically invoked when certain conditions are met.

As an example, we implement the standard branch-and-bound method of solving a MIP problem, using the external solver as an LP solver only. Firstly we outline how this can be implemented with the facilities we have already encountered. We then show how this can be improved usin more advanced features of lib(eplex)

With the branch-and-bound approach, a search-tree is formed, and at each node a `relaxed' version of the MIP problem is solved as an LP problem. Starting at the root, the problem solved is the original MIP problem, but without any of the integrality constraints:

:- eplex_instance(mip).

main5(Cost, Vars) :-
      % set up variables and constraints, but no integers/1 constraints
      % assume minimise for simplicity
      mip: eplex_solver_setup(min(Obj)), 
      mip: eplex_solve(RelaxedCost),
      mip: (Cost $>= RelaxedCost),  % RelaxedCost is lower bound 

Figure 16.5: Labelling a variable at a MIP tree node

In general, this initial LP solution contains non-integer assignments to integer variables. The objective value of this LP is a lower bound on the actual MIP objective value. The task of the search is to find integer assignments for the integer variables that optimises the objective function. Each node of the search-tree solves the problem with extra bound constraints on these variables. At each node, a particular variable is `labelled' as shown in Figure 16.5. The integer variable in this case has been assigned the non-integer value of 4.2. In the subsequent nodes of the tree, we consider two alternate problems, which creates two branches in the search. In one problem, we impose the bound constraint X ≤ 4, and in the other, X ≥ 5: these are the two nearest integer values to 4.2. In each branch, the problem is solved again as an LP problem with its new bound for the variable:

branching(IntVars) :-
      % for each integer variable X which violates the integer constraint
      mip: eplex_var_get(X, solution, XVal),
      Split is floor(XVal),
      % choice: branch on the two ranges for X
      (mip: (X $=< Split) ; mip: (X $>= Split + 1)),
      mip: eplex_solve(RelaxedCost),
      ...% repeat until there are no integer violations

A choice-point for the two alternative branchings is created in the above code, the problem is solved with one of the branchings (X $=< Split). The program then proceeds to further labelling of the variables. The alternative branch is left to be tried on backtracking.

Eventually, if the problem has a solution, all the integer variables will be `labelled' with integer values, resulting in a solution to the MIP problem. However, this will generally not be optimal, and so the program needs to backtrack into the tree to search for a better solution by trying the other branches for the variables, using the existing solution value as a bound. This `branch-and-bound' search technique is implemented in lib(branch_and_bound).

Remember that ECLiPSe provides libraries that make some programming tasks much easier. There is no need to write your own code when you can use what is provided by an ECLiPSe library.

Figure 16.6: Reminder: use ECLiPSe libraries!

In the code, the external solver is invoked explicitly at every node. This however may not be necessary as the imposed bound may already be satisfied. As stated at the start of this section, the invocation of the solver could be done in a data-driven way, more like a normal constraint. This is done with eplex_solver_setup/4: eplex_solver_setup(+Obj,-ObjVal,+Options,+Trigs), a more powerful version of eplex_solver_setup/1 for setting up a solver. The Trigs argument specifies a list of `trigger modes' for triggering the solver.
See the ECLiPSe reference manual for a complete description of the predicate.
For our example, we add a bound constraint at each node to exclude a fractional solution value for a variable. The criterion we want to use is to invoke the solver only if this old solution value is excluded by the new bounds (otherwise the external solver will solve the same problem redundantly). This is done by specifying deviating_bounds in the trigger modes. The full code that implements a MIP solution for the example transportation problem is given below:

:- lib(eplex).
:- lib(branch_and_bound).

:- eplex_instance(mip).

main6(Cost, Vars) :-
        % b. create the problem variables and set their range
        Vars = [A1,A2,A3,B1,B2,B3,C1,C2,C3,D1,D2,D3], 
        mip: (Vars :: 0.0..1.0Inf),

        % c. post the constraints for the problem to the eplex instance
        mip: (A1 + A2 + A3 $= 21),
        mip: (B1 + B2 + B3 $= 40),
        mip: (C1 + C2 + C3 $= 34),
        mip: (D1 + D2 + D3 $= 10),

        mip: (A1 + B1 + C1 + D1 $=< 50),
        mip: (A2 + B2 + C2 + D2 $=< 30),
        mip: (A3 + B3 + C3 + D3 $=< 40),
        mip: (A1 $= A2),

        % j. post the objective function as a constraint 
        ObjFunc = 10*A1 + 7*A2 + 200*A3 + 
                   8*B1 + 5*B2 + 10*B3 +
                   5*C1 + 5*C2 +  8*C3 + 
                   9*D1 + 3*D2 +  7*D3,
        mip: (ObjFunc  $= Cost),

        % k. this is a more flexible method for setting up a solver.
        %    [deviating_bounds] specifies that the external solver should be
        %    invoked when any solution value is outside the variable bounds 
        mip: eplex_solver_setup(min(ObjFunc), Cost, [], [deviating_bounds]),

        % l. Use the branch_and_bound library to do the branch and bound
        bb_min(( branching(Vars), 
                 mip: eplex_get(cost, Cost)
                 (foreach(V, Vars) do mip: eplex_var_get(V,solution,V))
               ), Cost, _).

branching(IntVars) :-
        % Find a variable X which does not have an integer solution value
        (integer_violation(IntVars, X, XVal) ->
            % m. try the closer integer range first
            Split is round(XVal),
            (Split > XVal ->
                (mip: (X $>= Split) ;  mip: (X $=< Split - 1))
                (mip: (X $=< Split) ; mip:  (X $>= Split + 1))
            % cannot find any integer violations; found a solution

% returns Var with solution value Val which violates the integer constraint
integer_violation([X|Xs], Var, Val) :-
        mip: eplex_var_get(X, solution, RelaxedSol),
        % m. we are dealing with floats here, so need some `margin' for a
        %    float value to be considered integer (1e-5 on either side)
        (abs( RelaxedSol - round(RelaxedSol) ) >= 1e-5 ->
            Var = X, Val = RelaxedSol
            integer_violation(Xs, Var, Val)

The setup of the solver is done in line k, with the use of the deviating_bounds trigger mode. There are no explicit calls to trigger the solver – it is triggered automatically. In addition, the first call to eplex_solve/1 for an initial solution is also not required, because when trigger modes are specified, then by default, eplex_solver_setup/4 will invoke the solver once the problem is setup.

Besides the
deviating_bounds trigger condition, the other argument of interest in our use of eplex_solver_setup/4 is the second argument, the objective value of the problem (Cost in the example): recall that this was returned previously by eplex_solve/1. Unlike in eplex_solve/1, the variable is not instantiated when the solver returns. Instead, one of the bounds (lower bound in the case of minimise) is updated to the optimal value, reflecting the range the objective value can take, from suboptimal to the `best' value at optimal. The variable is therefore made a problem variable by posting of the objective as a constraint in line j. This informs the external solver needs to be informed that the Cost variable is the objective value.

In line m, the branch choice is created by the posting of the bound constraint, which may trigger the external solver. Here, we use a simple heuristic to decide which of the two branches to try first: the branch with the integer range closer to the relaxed solution value. For example, in the situation of Figure 16.5, the branch with X $=< 4 is tried first since the solution value of 4.2 is closer to 4 than 5.

By using lib(branch_and_bound)'s
bb_min/3 predicate in m, there is no need to explicitly write our own branch-and-bound routine. However, this predicate requires the cost variable to be instantiated, so we call eplex_get(cost, Cost) to instantiate Cost at the end of each labelling of the variables. We also get the solution values for the variables, so that the branch-and-bound routine will remember it. The final value returned in Cost (and Vars for the solution values) is the optimal value after the branch-and-bound search, i.e. the optimal value for the MIP problem.

  • Use Instance:eplex_solver_setup(+Obj,-ObjVal,+Opts,+Trigs) to set up an external solver state for instance Instance. Trigs specifies a list of trigger conditions to automatically trigger the external solver.
  • Instance:eplex_var_get(+Var,+What,-Value) can be used to obtain information for the variable Var in the eplex instance.
  • Instance:eplex_get(+Item, -Value) can be used to retrieve information about the eplex instance's solver state.

Figure 16.7: More advanced modelling in eplex

Of course, in practice, we do not write our own MIP solver, but use the MIP solver provided with the external solvers instead. These solvers are highly optimised and tightly coupled to their own LP solvers. The techniques of solving relaxed subproblems described here are however very useful for combining the external solver with other solvers in a hybrid fashion.
See chapter 17 for more details on hybrid techniques.

Previous Up Next