Restrictandrelax search for 01 mixedinteger programs
Authors
 First Online:
 Received:
 Accepted:
DOI: 10.1007/s136750130007y
 Cite this article as:
 Guzelsoy, M., Nemhauser, G. & Savelsbergh, M. EURO J Comput Optim (2013) 1: 201. doi:10.1007/s136750130007y
Abstract
A highly desirable characteristic of methods for solving 01 mixedinteger programs is that they should be capable of producing highquality solutions quickly. We introduce restrictandrelax search, a branchandbound algorithm that explores the solution space not only by fixing variables (restricting), but also by freeing, or unfixing, previously fixed variables (relaxing). Starting from a restricted 01 mixedinteger program, the branchandbound algorithm may, at any node of the search tree, selectively relax, or unfix, previously fixed variables, restrict, or fix, additional variables, or unfix and fix variables at the same time using dual or structural information (problemspecific information). This process yields a dynamic search that is likely to find highquality feasible solutions more quickly than a traditional search and that is also capable of proving optimality. A proofofconcept computational study demonstrates its potential. A straightforward generic implementation in SYMPHONY, an open source solver for mixedinteger programs, is shown to generate highquality solutions for many MIPLIB instances as well as for largescale multicommodity fixedcharge network flow instances much more quickly than SYMPHONY itself.
Keywords
01 Integer programming Branchandbound Feasible solutions RestrictandrelaxMathematics Subject Classification
90C11 65K05Introduction
A highly desirable characteristic of methods for solving 01 mixedinteger programs (MIPs) is that they should be capable of producing highquality solutions in the early stages of the computation. From a practical perspective, finding highquality solutions and doing so quickly is frequently all that is desired. From a methodological perspective, it is important because many of the techniques embedded in integer programming solvers rely on the availability of both lower and upper bounds. Therefore, it is not surprising that substantial effort has been dedicated, especially in the last decade, to enhancing the ability of methods for solving 01 MIPs to find highquality solutions quickly, either by embedding techniques specifically designed for this purpose, or by modifying the methods themselves.
One of the techniques specifically designed to find highquality solutions and embedded in stateoftheart commercial branchandbound solvers, such as CPLEX, XPRESS, and Gurobi, is relaxation induced neighborhood search (RINS) [2]. RINS explores a part of the search space defined by a known feasible solution and a linear programming solution. Variables with the same value in both solutions are fixed, and the resulting restricted MIP is solved. The hope is that the reduced search space defined by the fixed variables contains highquality integer solutions and that it can be explored efficiently by solving the restricted MIP. A variant that does not require a known feasible solution is relaxation enforced neighborhood search (RENS) [1].
Local branching [5] is a related, but quite different, example of the effective use of restricted MIPs. In local branching, given a reference solution \(\bar{x}\), a neighborhood of the reference solution is defined by those feasible solutions that differ in at most \(k\) of the binary variables, i.e., the feasible solutions with a Hamming distance of less than or equal to \(k\) from the reference solution. This concept (or more specifically the local branching constraint) is used to define a branching scheme in which in one branch a restricted MIP is solved to find a best solution in the neighborhood, and in the other branch a Hamming distance of greater than or equal to \(k+1\) is enforced. The resulting branchandbound algorithm favors improving the incumbent solution, hence producing highquality solutions in the early stages of the computation.
Yet another example of the effective use of restricted MIPs is branchandprice guided search (BPGS) [8]. BPGS is based on an extended formulation of the MIP to be solved, in which restrictions of that MIP are enumerated. The extended formulation is solved with a branchandprice algorithm, which automatically defines and solves restricted MIPs and is guaranteed to produce a provable optimal solution when run to completion. Computational experience demonstrates that the pricing problems quickly produce restricted MIPs that yield highquality solutions.
The common characteristic of these approaches is that they explore portions of the solution space by defining a restricted MIP and then solving that MIP using a general MIP solver. The idea is that by fixing certain variables to obtain a restricted MIP, the restricted MIP can be solved more quickly and if the variables that were fixed were chosen appropriately, the resulting solution will be a highquality solution to the original, unrestricted MIP.
In this paper, we introduce restrictandrelax search, a branchandbound algorithm that explores the solution space not only by fixing variables (restricting), but also by freeing, or unfixing, previously fixed variables (relaxing). Starting from a restricted 01 mixedinteger program, the branchandbound algorithm may, at any node of the search tree, selectively relax, or unfix, previously fixed variables, restrict, or fix, additional variables, or unfix and fix variables at the same time using dual or structural information (problemspecific information). This process yields a dynamic search which tends to find highquality feasible solutions more quickly than a traditional search. Restrictandrelax search always works with a restricted, but dynamically changing, set of variables. As a consequence, restrictandrelax search can be used to tackle instances that are so large that they cannot be fully loaded into memory. Producing highquality solutions quickly to large MIPs was the main motivation for developing restrictandrelax search. Another was to investigate whether it is possible to effectively exploit information obtained during branchandbound to refine an initial restricted MIP, e.g., either by fixing additional variables or by unfixing previously fixed variables.
It is insightful to highlight the main differences between restrictandrelax search and some of the other approaches exploiting the benefits of solving restricted MIPs. Restrictandrelax search is fundamentally different from standard branchandbound enhanced with local search to find highquality solutions in the early stages of the computation (e.g., branchandbound enhanced with RINS). The latter approach always works on the full MIP, but, at certain nodes in the search tree, solves a specific restricted MIP. Restrictandrelax search, instead, always works on a restricted, but changing, MIP, and might only look at the full MIP at the end of the search if it is desired to prove optimality. Restrictandrelax search never solves a specific restricted MIP. Restrictandrelax search also differs from traditional IPbased local search, in which some sequence of restricted MIPs is solved, since it can, if desired, yield a proof of optimality, and bounds can be obtained throughout the search.
A computational study demonstrates the potential of restrictandrelax search. Our generic implementation in SYMPHONY [11], an open source solver for mixedinteger programs, is shown to generate highquality solutions for many MIPLIB instances as well as for largescale multicommodity fixedcharge network flow instances much more quickly than SYMPHONY itself.
In the rest of the paper, we first present the core ideas underlying restrictandrelax search. Then we discuss our proofofconcept implementation of restrictandrelax search and we present the results of an extensive computational study. We close with some final remarks and directions for further research.
Restrictandrelax search
The benefit of solving the restricted 01 MIP given by (2) is that it is smaller than the original 01 MIP and therefore can likely be solved much faster. Unfortunately, there is no guarantee that an optimal, highquality, or even feasible, solution is found. That depends on the choice of the index set \(F\) of variables that is fixed, and the choice of the values to which these variables are fixed. To eliminate this uncertainty, restrictandrelax search, at any node of the search tree, may unfix previously fixed variables. Furthermore, to maintain the benefit of solving a small, restricted 01 MIP, restrictandrelax search, at any node of the search tree, may also fix additional variables.
Observe that the choice of variables to unfix (and fix) is based on local information, i.e., information that is relevant to the restricted 01 MIP defined by the node in the branchandbound tree. The hope and expectation is that this quickly guides the search into promising parts of the solution space; therefore, producing highquality solutions early in the search (and thus avoids exploring unnecessary parts of the solution space). Note too that restricting and relaxing serve two different purposes in restrictandrelax search. Restricting serves to increase the efficiency by reducing the size of the linear programs that need to be solved, whereas relaxing serves to increase the solution quality by expanding the space that is explored.
Recognizing that in a traditional branchandbound tree, the levels corresponds to the number of variables fixed by branching, restrictandrelax search can be interpreted as starting at a node at level \(k\), where \(k\) is the number of variables fixed initially, and then jumping up and down the tree as variables are unfixed and fixed. The rules used for pruning nodes are the same as in traditional branchandbound search, unless restrictandrelax search is turned into a heuristic by pruning some nodes without proof that they cannot contain an optimal solution (see below for more details).
The key to success for restrictandrelax search is choosing the variables to fix and to unfix effectively. It is natural to use the primal and dual information available at a node to do so. This reflects one of the main thrusts underlying restrictandrelax search, namely exploiting local information to guide the search for highquality integer solutions.
Unfixing variables at a node is more involved. There are three cases to consider: (1) the LP is feasible and the LP value is less than the value of the best known feasible solution; (2) the LP is feasible and the LP value is greater than or equal to the value of the best known feasible solution; and (3) the LP is infeasible. In case (1), we choose variables to unfix that are likely to cause a decrease in the LP value. For a fixed variable \(x_i\), \(i \in F_t\), we have that if \(x^*_i = 0\) and \(r^*_i < 0\) or if \(x^*_i = 1\) and \(r^*_i > 0\), then unfixing this variable may result in the current solution no longer being optimal and thus yielding a new problem with a strictly lower LP solution value. Hence, we only unfix variables with this property and, similar to fixing variables, we choose to unfix variables in nonincreasing order of the absolute value of their reduced costs. In case (2), we cannot fathom the node as normally would be done in a branchandbound algorithm, since it is possible that by unfixing some of the fixed variables, the LP value will become less than the value of the best known feasible solution. Therefore, in this situation, we unfix variables as described above. However, care has to be taken with the implementation. Within a branchandbound tree, MIP solvers use the dual simplex algorithm to solve LPs and stop as soon as the LP value exceeds the value of the best known feasible solution, and therefore can terminate without producing a primal LP solution. In fact, there may not even be a feasible primal solution. Therefore, if there is a cutoff, we remove it and resolve the LP to obtain its true status. If it is feasible, we proceed as before, i.e., we use dual information to unfix variables. Case (3) is similar to case (2), but the dual is now unbounded. In this case, we have found that more drastic action is needed to see if we can fathom the node. If the number of fixed variables is not too big, we unfix all previously fixed variables (i.e., all but the variables fixed by branching) and resolve the LP. If the LP is feasible and has an objective value below the value of the best known feasible solution, we unfix the variables that have a value different from their previously fixed value. For instance, if a variable was fixed to 1, but now has a value 0.6, then it will be unfixed. If the LP is feasible but has an objective value above the value of the best known feasible solution, or the LP is infeasible, the node is discarded, since it has been completely processed.
Observe that the scheme for handling nodes with an infeasible LP ensures that only nodes that can truly be discarded are discarded and thus the proposed method is capable of solving the instance to proven optimality. However, the scheme involves solving the full LP, which we may not want to (or may not be able to) do for very large instances. In these situations, restrictandrelax search can be converted into a heuristic by either accepting the LP infeasibility immediately, or by exploring what happens when a small number of previously fixed variables is unfixed.
The success of restrictandrelax search may depend on the quality of the initial restriction. Three natural choices for initial fixings are: (1) choose the variables to fix based on some known feasible solution; (2) choose the variables to fix based on the solution to the linear programming relaxation; or (3) choose the variables to fix based on the initial feasible solution to the linear programming relaxation, i.e., the solution after Phase I. The last option is of interest when dealing with extremely large instances or when dealing with instances for which it is very timeconsuming to solve the linear programming relaxation to optimality.
We have employed strategies (2) and (3). That is, we fix variables with integral values in an LP solution. More specifically, if \(x^\mathrm{LP}_i = 0\), then we assign a score \(s_i = c_i\), and if \(x^\mathrm{LP}_i = 1\) then we assign a score \(s_i = c_i\). We then fix variables in nondecreasing order of these scores, so that we first fix variables that, when relaxed, increase the value of LP solution more than others. Note that this implies that fixing is random for variables with the same objective coefficients. We fix at most 90 % of the binary variables.
The use of cutting planes in modern integer programming solvers has contributed substantially to their success. Cutting planes are generated and added at selected nodes of the branchandbound tree to obtain a tighter polyhedral approximation of the convex hull of the integer feasible set.
The use of cutting planes in restrictandrelax search requires careful consideration because inequalities generated at a node \(t\) for the subproblem defined by \(F^j_t\), i.e., the subproblem given by (5), may not be valid for the subproblem defined by \(F^{j+1}_t\) (unless \(F^{j+1}_t \subseteq F^{j}_t\)) and for the subproblem associated with a node \(s\) in the subtree rooted at \(t\) (unless \(F_s \subseteq F^{j}_t\)). The valid inequalities also might cut off feasible solutions when variables are relaxed.

All existing cutting planes are removed as soon as variables are relaxed. (Of course new cutting planes can be generated for the new subproblem.)

All relaxed variables are lifted in all existing cutting planes using some lifting algorithm.

Cutting planes are generated in such a way that they are globally valid, where globally valid is with respect to \(\mathcal{S}\) in problem (1), i.e., the cutting planes are generated assuming that all variables (other than variables fixed by branching) are relaxed.
Implementation
Restrictandrelax search parameters
Mindepth (\(d^\mathrm{min}\)) 
Fix and unfix only at nodes below tree level \(d^\mathrm{min}\) (default: 0) 
Maxdepth (\(d^\mathrm{max}\)) 
Fix and unfix only at nodes above tree level \(d^\mathrm{max}\) (default: 50) 
Levelfrequency (lf) 
Fix and unfix at a node if the node level is a multiple of lf (default: 3) 
Triallimit (tl) 
At a node, fix and unfix at most tl times (default: 5) 
Unfixratio (ur) 
In a trial, unfix at most ur % of the fixed variables (default: 5) 
Fixratio (fr) 
In a trial, fix at most fr % of the unfixed variables (default: 5) 
Prunebybound (pb) 
If enabled, fix and unfix at nodes pruned by bound regardless 
of the node level (default: enabled)  
Prunebyinfeasibility (pi) 
If enabled, fix and unfix at nodes pruned by infeasibility regardless 
of the node level (default: enabled) 
Let \(t\) be a node with a tree level that is multiple of lf, that is greater than \(d^\mathrm{min}\), and that is less than \(d^\mathrm{max}\), or a node that is to be pruned by bound (if pb is enabled), or a node that is to be pruned by infeasibility (if pi is enabled). Then, we fix and relax at node \(t\) at least once and at most tl times until the node LP relaxation is feasible and has a solution value strictly less than the current upper bound.
Varying the parameter values can have significant impact on the performance on specific instances. The default values provide a compromise that seems to perform reasonably well across the instances in our test set.

all integer variables (other than the variables fixed by branching) are unfixed and the node LP yields an integer feasible solution;

all integer variables (other than the variables fixed by branching) are unfixed and the node LP is infeasible or the node LP value is greater than the value of the best known feasible solution; or

fixing and unfixing has been performed tl times and the node LP remains either infeasible or has a value that is greater than the value of the best known feasible solution.
We have embedded restrictandrelax search in the development version of SYMPHONY [11] with CLP (the COIN LP solver) because SYMPHONY provides the capability to start from a restricted problem and to fix and unfix variables at selected nodes in the search tree (functionality that is not available in the commercial solvers CPLEX, Gurobi, and XPRESS).
SYMPHONY itself contains various techniques aimed at finding feasible solutions quickly, e.g., two rounding heuristics, six diving heuristics, an implementation of the feasibility pump [4], local branching and RINS, and restrictandrelax search benefits from (and to some extend relies on) their effectiveness.
Computational study
We have conducted a computational study to assess the potential of restrictandrelax search. The goal of the computational study is to compare traditional branchandbound with restrictandrelax search with regard to finding good solutions quickly. The computational study has two parts. In the first part, we use restrictandrelax search to quickly find highquality solutions to 01 MIPs from the wellknown test set MIPLIB [10]. In the second part, we use restrictandrelax search to quickly find highquality solutions to instances of the multicommodity fixedcharge network flow problem. All experiments were run on an Intel Xeon E5520 processor at 2.27 GHz.

Instances with general integer variables (72);

Instances which are infeasible, the instances labeled I in MIPLIB (15);

Instances which are extra large, the instances labeled X in MIPLIB (9);

Instances which are unstable, the instances labeled U in MIPLIB (16);

Instances which are unstable for either CLP or SYMPHONY (23);

Instances which are easy, where easy is defined as being solved to optimality by SYMPHONY in less than 500 s (34).
 1.
Running the default solver (SYMPHONY) on the full MIP (denoted by Deffull);
 2.
Running the default solver (SYMPHONY) on the initial restricted MIP (denoted by Defrestricted);
 3.
Running restrictandrelax search in which the fixing of variables during the search is disabled (denoted by RRrelaxonly);
 4.
Running restrictandrelax search (denoted by RR).
Performance statistics for RRrelaxonly
Deffull 
Defrestricted  

100 s 
250 s 
500 s 
1,000 s 
2,000 s 
100 s 
250 s 
500 s 
1,000 s 
2,000 s  
Feas 
8 
8 
12 
4 
4 
14 
21 
26 
25 
27 
\(<\) 
35 
48 
53 
56 
51 
39 
45 
49 
50 
50 
\(=\) 
11 
11 
13 
14 
20 
7 
3 
3 
4 
4 
\(>\) 
26 
20 
17 
21 
23 
20 
18 
17 
16 
17 
Nofeas 
4 
4 
3 
4 
4 
0 
2 
1 
1 
0 
Performance statistics for RR
Deffull 
Defrestricted  

100 s 
250 s 
500 s 
1,000 s 
2,000 s 
100 s 
250 s 
500 s 
1,000 s 
2,000 s  
Feas 
7 
7 
10 
3 
3 
13 
21 
24 
24 
26 
\(<\) 
34 
46 
49 
52 
47 
42 
43 
52 
52 
54 
\(=\) 
14 
13 
14 
17 
20 
6 
3 
3 
4 
2 
\(>\) 
24 
21 
21 
23 
27 
18 
20 
15 
15 
15 
Nofeas 
4 
3 
2 
3 
4 
0 
2 
0 
0 
0 
We observe that both variants of restrictandrelax search perform significantly better than the default solver when either given the full MIP or the initial restricted MIP. Furthermore, the benefits of restrictandrelax search are most clearly seen with a time limit of 500 s. When the time limit is smaller, a relatively large portion of the time is consumed by root node processing and the benefits of the dynamic exploration of the search space cannot be fully realized. When the time limit is larger, the benefits of the dynamic exploration of the search space, i.e., finding a highquality solution quickly, is less significant, as the default solver has sufficient time to find a highquality solution as well. The ability of restrictandrelax search to quickly find feasible solutions can be observed for the smaller time limits, since the number of instances where restrictandrelax search finds a feasible solution and the default solver does not is higher than the number of instances where the default solver finds a feasible solution and the restrictandrelax search does not.
Performance statistics for RR
RRrelaxonly  

100 s 
250 s 
500 s 
1,000 s 
2,000 s  
Feas 
1 
1 
2 
1 
2 
\(<\) 
19 
21 
25 
26 
26 
\(=\) 
39 
41 
40 
40 
41 
\(>\) 
21 
24 
28 
28 
29 
Nofeas 
0 
1 
1 
1 
1 
The performance profiles highlight different aspects of the performance of the solution approaches. When \(\tau = 1\), the performance profile shows the fraction of instances for which a solution approach performs best. We see that when the time limit is 100 s, the fraction of instances for which a solution approach performs best is similar for RR, RRrelaxonly, and Deffull, a little over 40 %. As we observed before, when the time limit is only 100 s, the benefits of restrictandrelax search cannot always be realized because a relatively large fraction of the time is spent on processing the root node. When the time limit is between 250 and 1,000 s, we see that RR and RRrelaxonly do substantially better than Deffull, performing best in about 50 % of the instances compared to about 30 % for Deffull. The difference in performance between RR and RRrelaxonly and Deffull is a bit less when the time limit is 2,000 s, but still substantial. It is important to realize that when given enough time, Deffull should be at least as good as RR and RRrelaxonly because it will find an optimal solution (and possibly even better because our implementation of RR and RRrelaxonly does not guarantee that they find an optimal solution). However, apparently the time limit of 2,000 s is not enough for Deffull to catch up. When \(\tau = 32\), the performance profile essentially shows the fraction of instances for which a solution approach has found a feasible solution (out of the set of instances for which at least one of the solution approaches has found a feasible solution). We see that when the time limit is either 250 or 500 s RR and RRrelaxonly do noticeably better than Deffull, whereas for the other time limits the difference is minor, although it exists. This clearly demonstrates the ability of restrictandrelax search to quickly find highquality feasible solutions; if only limited time is available to find a feasible solution, restrictandrelax search should be the method of choice. When \(\tau \) is between 2 and 4, we get an indication of the average performance of the solution approaches. A high value indicates that a solution approach may not necessarily have been the best, but it has not been far from the best. The fact that we observe a relatively large difference between RR and RRrelaxonly and Deffull in this range, again especially when the time limit is 250 or 500 s, shows the robustness of restrictandrelax search. Overall, the performance profiles demonstrate decidedly that restrictandrelax search has the ability to quickly find highquality solutions. The performance profiles also demonstrate indisputably that relaxing is crucial to the success of restrictandrelax search since the performance of Defrestrict, which solves the initial restricted integer program, is far worse than the other three approaches.
Optimality gaps for solutions produced by RRrelaxonly in 500 s
Deffull 
Defrestricted  

\(<\) 
\(= \) 
\(>\) 
\(<\) 
\(=\) 
\(>\)  
\([0, 1)\) 
11 
9 
1 
18 
2 
1 
\([1, 5)\) 
15 
1 
6 
16 
0 
6 
\([5, 10)\) 
10 
2 
1 
12 
0 
1 
\([10, 20)\) 
10 
0 
5 
13 
0 
2 
\([20, 50)\) 
8 
1 
1 
9 
1 
0 
\([50, 100)\) 
3 
0 
3 
6 
0 
0 
\([100, \infty )\) 
8 
0 
0 
1 
0 
7 
Optimality gaps for solutions produced by RR in 500 s
Deffull 
Defrestricted  

\(<\) 
\(= \) 
\(>\) 
\(<\) 
\(=\) 
\(>\)  
\([0, 1)\) 
11 
11 
1 
21 
1 
1 
\([1, 5)\) 
18 
1 
6 
19 
0 
6 
\([5, 10)\) 
8 
1 
2 
11 
0 
0 
\([10, 20)\) 
6 
0 
4 
10 
0 
0 
\([20, 50)\) 
6 
1 
4 
8 
2 
1 
\([50, 100)\) 
2 
0 
3 
5 
0 
0 
\([100, \infty )\) 
8 
0 
1 
2 
0 
7 
We see that the quality of many of the solutions produced by RR is quite good, but that for some instances the quality of the solution is still quite poor. It is important to remember that some of the instances in the test set are quite large and that only relatively few nodes have been evaluated in the 500 s.
In the second part of our computational study, we investigate whether the generic restrictandrelax search implementation is able to find highquality solutions to a set of largescale instances of integer multicommodity fixedcharge network flow problems (see Hewitt et al. [7] for a description of the problem and the instances). The size of these instances varies between 150,000–600,000 variables, 180,000–700,000 constraints and 750,000–3,000,000 nonzero elements.
Even solving the LP relaxation for some of these instances in a reasonable amount of time can be challenging. Therefore, we construct the initial restricted IP using the primal feasible LP solution obtained by the Phase I simplex algorithm. The LP relaxation of the restricted IP is then solved to obtain the dual information used for fixing and unfixing at the root of the search tree.
Recall that when an infeasible LP is encountered, the generic implementation of restrictandrelax search unfixes all previously fixed variables (i.e., all but the variables fixed by branching) and resolves the LP. Because we do not want to solve the full LP, when an infeasible LP is encountered during the search, unfixing is done using dual information associated with the LP solution at the parent node. If the resulting LP is feasible and has an objective value below the value of the best known feasible solution, then variables that have a value different from their previously fixed value are unfixed (otherwise the node is fathomed).
Finally, based on some preliminary experimentation, the following parameter values were used: maxdepth, 1,000; triallimit, 100; unfixratio, 0.50 %; and fixratio, 1.0 %.
Results for multicommodity fixedcharge network flow instances with a 15min time limit
\(z_\mathrm{LP}\) 
Deffull 
Defrestricted 
RRrelaxonly 
RR  

\(z\) 
# sol found 
\(z\) 
# sol found 
\(z\) 
# sol found 
\(z\) 
# sol found  
T52150 
\(\infty \) 
\(\infty \) 
0 
16,791,739 
1 
10,633,431 
12 
6,348,474 
110 
T52200 
\(\infty \) 
\(\infty \) 
0 
18,264,978 
1 
10,295,957 
13 
6,971,284 
106 
T52.5100 
3,010,123 
3,096,873 
1 
11,421,658 
1 
6,680,455 
13 
3,816,954 
62 
T52.5150 
\(\infty \) 
\(\infty \) 
0 
16,519,891 
1 
12,725,369 
5 
7,778,874 
101 
T52.5200 
\(\infty \) 
\(\infty \) 
0 
22,803,549 
1 
19,499,182 
4 
16,915,379 
22 
T53100 
\(\infty \) 
\(\infty \) 
0 
15,336,883 
1 
10,976,480 
4 
6,990,546 
111 
T53125 
\(\infty \) 
\(\infty \) 
0 
13,856,412 
1 
8,983,917 
6 
5,822,540 
59 
T53150 
\(\infty \) 
\(\infty \) 
0 
15,464,987 
1 
12,539,533 
4 
8,173,244 
31 
T53200 
\(\infty \) 
\(\infty \) 
0 
23,302,848 
1 
20,855,570 
2 
14,119,195 
50 
T5350 
2,143,902 
2,310,557 
1 
7,539,845 
3 
3,818,608 
15 
2,858,502 
64 
T5375 
\(\infty \) 
\(\infty \) 
0 
11,478,392 
1 
6,484,868 
17 
4,622,346 
49 
Results for multicommodity fixedcharge network flow instances with a 60min time limit
\(z_\mathrm{LP}\) 
Deffull 
Defrestricted 
RRrelaxonly 
RR  

\(z\) 
# sol found 
\(z\) 
# sol found 
\(z\) 
# sol found 
\(z\) 
# sol found  
T52150 
5,080,051 
7,455,350 
1 
16,791,739 
1 
9,666,279 
19 
6,048,452 
133 
T52200 
5,410,777 
\(\infty \) 
0 
18,264,978 
1 
9,603,211 
16 
6,659,686 
123 
T52.5100 
3,010,123 
3,096,873 
1 
11,421,658 
1 
5,440,807 
27 
3,459,553 
75 
T52.5150 
\(\infty \) 
\(\infty \) 
0 
16,519,891 
1 
11,867,108 
9 
6,963,319 
146 
T52.5200 
\(\infty \) 
\(\infty \) 
0 
22,803,549 
1 
18,937,251 
5 
12,827,664 
83 
T53100 
\(\infty \) 
\(\infty \) 
0 
15,336,883 
1 
10,967,825 
5 
6,797,231 
120 
T53125 
\(\infty \) 
\(\infty \) 
0 
13,856,412 
1 
8,715,813 
7 
5,115,193 
99 
T53150 
\(\infty \) 
\(\infty \) 
0 
15,464,987 
1 
11,365,537 
5 
6,752,942 
88 
T53200 
\(\infty \) 
\(\infty \) 
0 
23,302,848 
1 
19,307,474 
4 
10,569,513 
144 
T5350 
2,143,902 
2,310,557 
1 
7,539,845 
3 
3,055,353 
27 
2,858,502 
64 
T5375 
3,540,911 
\(\infty \) 
0 
11,478,392 
1 
6,293,908 
23 
4,545,794 
53 
A number of observations can be made regarding these results. First, we see that solving the LP relaxation of these instances is indeed quite challenging. Even with a time limit of 60 min, CLP is unable to solve the LP relaxation for 6 out of the 11 instances. (With CPLEX version 12.4 similar results are obtained; CPLEX is unable to solve the LP relaxation within 60 min for 7 out of the 11 instances.) Secondly, we observe that restrictandrelax search is able to produce IP solutions for all of the instances within 15 min. In fact, for all of the instances, restrictandrelax search has generated many feasible IP solutions within 15 min. This clearly indicates the enormous potential of restrictandrelax search for extremely large instances of certain classes of integer programs. Thirdly, we observe that for these instances, there is a clearly observable difference between the performance of RR and RRrelaxonly. A closer examination of the results shows that RR evaluates far more nodes than RRrelaxonly (75 % more with a time limit of 15 min and 150 % more with a time limit of 60 min), which is almost certainly due to the fact that large numbers of variables are fixed at nodes in the tree resulting in smaller linear programs being solved at nodes. RR therefore explores a much larger portion of the solution space than RRrelaxonly.
Final remarks
The development of restrictandrelax search was motivated by the need to quickly find highquality solutions to very large integer programs. However, as our computational study demonstrates, restrictandrelax search has significant benefits also when solving much smaller integer programs. Our proofofconcept implementation of restrictandrelax search in SYMPHONY is effective in producing highquality solutions to many of the integer programs in MIPLIB and to largescale multicommodity fixedcharge network flow instances much more quickly than SYMPHONY itself. Perhaps even more important, it challenges the paradigm that has been at the heart of branchandbound algorithms since their inception: rather than organizing the search around restrictions only, restrictandrelax search organizes the search around restrictions and relaxations. It has long been recognized that the basic branchandbound paradigm has weaknesses (e.g., that it is difficult to recover from “unfortunate” branching decisions at the top of the tree) and dynamic strategies involving restarts have been proposed to address these weaknesses. Restrictandrelax search provides a fundamentally different perspective on how to explore the space of feasible solutions.
As mentioned above, we have used a proofofconcept implementation of restrictandrelax search to successfully demonstrate its potential. To reach its full potential, more research is needed. There is clearly a need to investigate and explore alternative methods for producing the initial restricted integer program. The choice of the initial restricted integer program may be especially important for integer programs with few feasible solutions. Schemes that incorporate ideas from recent work on finding a small set of critical variables (a “backdoor” in the terminology of [12]) to be used first for branching, e.g., [9] and [6], may be quite powerful.
When encountering an infeasible LP, the generic implementation of restrictandrelax search unfixes all previously fixed variables and resolves the LP to ascertain the status of the node and, if possible, to obtain dual information to guide the unfixing of variables. That is a pragmatic, but computationally expensive choice. When solving multicommodity fixedcharge network flow instances, we chose not to do this, but instead to use the dual information associated with LP solution at the parent node. There is a need to better understand the impact of these and other possible choices.
The generic implementation has a large number of parameters that control its behavior. We have observed that varying the parameter values can have significant impact on the performance on specific instances. The default values provide a compromise that seems to perform reasonably well across the instances in our test set. A better understanding of and better mechanisms for detecting when to restrict and when to relax during the search is essential.
An important component of a branchandbound algorithm is the node selection scheme. Node selection schemes, i.e., schemes that decide which of the active nodes to evaluate next, balance two goals: finding better solutions and proving that no better solutions exist. It is not clear that any of the node selection schemes currently employed by integer programming solvers is appropriate for restrictandrelax search. The fact that it is possible to unfix variables at a node and thus locally enlarge the search space should be taken into account in node selection schemes.
To summarize, we have introduced a new branchandbound search paradigm for 01 mixedinteger programs that is designed to find highquality solutions quickly, but is also capable of proving optimality or infeasibility. We have conducted computational experiments with a proofofconcept implementation that indicates its potential, and have identified a number of research avenues that should make restrictandrelax search even more powerful.