« Previous « Start » Next »

5  Algorithmic description

5.1  Overview

Classical methods for the solution of MINLP problems decompose the problem by separating the nonlinear part from the integer part. This approach is largely due to the existence of packaged software for solving Nonlinear Programming (NLP) and Mixed Integer Linear Programming problems.

In contrast, an integrated approach to solving MINLP problems is considered here. This new algorithm is based on branch-and-bound, but does not require the NLP problem at each node to be solved to optimality. Instead, branching is allowed after each iteration of the NLP solver. In this way, the nonlinear part of the MINLP problem is solved whilst searching the tree. The nonlinear solver that is considered in this manual is a Sequential Quadratic Programming solver.

A numerical comparison of the new method with nonlinear branch-and-bound is presented and a factor of up to 3 improvement over branch-and-bound is observed.

5.2  Introduction

This manual considers the solution of Mixed Integer Nonlinear Programming (MINLP) problems. Problems of this type arise when some of the variables are required to take integer values. Applications of MINLP problems include the design of batch plants (e.g. [17] and [22]), the synthesis of processes (e.g. [7]), the design of distillation sequences (e.g. [29]), the optimal positioning of products in a multi-attribute space (e.g. [7]), the minimization of waste in paper cutting [28] and the optimization of core reload patterns for nuclear reactors [1]. For a comprehensive survey of applications of MINLP applications see Grossmann and Kravanja [16].

MINLP problems can be modelled in the following form where y are the integer variables modelling for instance decisions ( y { 0 , 1 }), numbers of equipment (y { 0, 1, … ,N }) or discrete sizes and x are the continuous variables.

Classical methods for the solution of (P) “decompose” the problem by separating the nonlinear part from the integer part. This approach is made possible by the existence of packaged software for solving Nonlinear Programming (NLP) and Mixed Integer Linear Programming (MILP) problems. The decomposition is even more apparent in Outer Approximation ([7] and [10]) and Benders Decomposition ([15] and [13]) where an alternating sequence of MILP master problems and NLP subproblems obtained by fixing the integer variables is solved (see also the recent monograph of Floudas [14]). The recent branch-and-cut method for 0-1 convex MINLP of Stubbs and Mehrotra [27] also separates the nonlinear (lower bounding and cut generation) and integer part of the problem.
@percentBranch-and-bound [6] can also be viewed as a “decomposition” in which the tree-search (integer part) is largely separated from solving the NLP problems at each node (nonlinear part).

In contrast to these approaches, we considers an integrated approach to MINLP problems. The algorithm developed here is based on branch-and-bound, but instead of solving an NLP problem at each node of the tree, the tree-search and the iterative solution of the NLP are interlaced. Thus the nonlinear part of (P) is solved whilst searching the tree. The nonlinear solver that is considered in this manual is a Sequential Quadratic Programming (SQP) solver. SQP solves an NLP problem via a sequence of quadratic programming (QP) problem approximations.

The basic idea underlying the new approach is to branch early – possibly after a single QP iteration of the SQP solver. This idea was first presented by Borchers and Mitchell [4]. The present algorithm improves on their idea in a number of ways:

1. In order to derive lower bounds needed in branch-and-bound Borchers and Mitchell propose to evaluate the Lagrangian dual. This is effectively an unconstrained nonlinear optimization problem. In Section 5.4 it is shown that there is no need to compute a lower bound if the linearizations in the SQP solver are interpreted as cutting planes.
2. In [4] two QP problems are solved before branching. This restriction is due to the fact that a packaged SQP solver is used. By using our own solver we are able to remove this unnecessary restriction, widening the scope for early branching.
3. A new heuristic for deciding when to branch early is introduced which does not rely on a fixed parameter, but takes the second order rate of convergence of the SQP solver into account when deciding whether to branch or continue with the SQP iteration.

The ideas presented here have a similar motivation as a recent algorithm by Quesada and Grossmann [26]. Their algorithm is related to outer approximation but avoids the re-solution of related MILP master problems by interrupting the MILP branch-and-bound solver each time an integer node is encountered. At this node an NLP problem is solved (obtained by fixing all integer variables) and new outer approximations are added to all problems on the MILP branch-and-bound tree. Thus the MILP is updated and the tree-search resumes. The difference to the approach considered here is that Quesada and Grossmann still solve NLP problems at some nodes, whereas the new solver presented here usually solves one quadratic programming (QP) problem at each node.

The algorithmic description is organized as follows. Section 5.3 briefly reviews the SQP method and branch-and-bound. In Section 5.4 the new algorithm is developed for the special case of convex MINLP problems. Section 5.4.4 introduces a heuristic for handling nonconvex MINLP problems. Finally, Section 5.5 presents a comparison of the new method with nonlinear branch-and-bound. These results are very encouraging, often improving on branch-and-bound by a factor of up to 3.

5.3  Background

5.3.1  Sequential Quadratic Programming

A popular iterative method for solving NLP problems is Sequential Quadratic Programming (SQP). The basic form of SQP methods date back to Wilson [30] and were popularized by Han [19] and Powell [25], see Fletcher [9, Chapter 12.4] and Conn, Gould and Toint [3] for a selection of other references. In its basic form, SQP solves an NLP problem via a sequence of quadratic programming (QP) approximations obtained by replacing the nonlinear constraints by a linear first order Taylor series approximation and the nonlinear objective by a second order Taylor series approximation augmented by second order information from the constraints. It can be shown under certain conditions that the SQP method converges quadratically near a solution.

It is well known that the SQP method may fail to converge if it is started far from a local solution. In order to induce convergence, many popular methods use a penalty function which is a linear combination of the objective function f and some measure of the constraint violation. A related idea is an augmented Lagrangian function in which a weighted penalty term is added to a Lagrangian function. A step in an SQP method is then accepted when it produces a sufficient decrease in the penalty function.

Two frameworks exist which enforce sufficient decrease, namely line-search in the direction of the QP solution or a trust-region that limits the QP step that is computed (see e.g. [9, Chapter 12.4] and references therein). In our implementation global convergence is promoted through the use of a trust-region and the new concept of a “filter” [11] which accepts a trial point whenever the objective or the constraint violation is improved compared to all previous iterates. The size of the trust-region is reduced if the step is rejected and increased if it is accepted.

5.3.2  Branch-and-bound

Branch-and-bound dates back to Land and Doig [23]. The first reference to nonlinear branch-and-bound can be found in Dakin [6]. It is most conveniently explained in terms of a tree-search.

Initially, all integer restrictions are relaxed and the resulting NLP relaxation is solved. If all integer variables take an integer value at the solution then this solution also solves the MINLP. Usually, some integer variables take a non-integer value. The algorithm then selects one of those integer variables which take a non-integer value, say y_i with value y_i, and branches on it. Branching generates two new NLP problems by adding simple bounds y_i ≤ [ y_i ] and y_i ≥ [ y_i ] + 1 respectively to the NLP relaxation (where [a] is the largest integer not greater than a).

One of the two new NLP problems is selected and solved next. If the integer variables take non-integer values then branching is repeated, thus generating a branch-and-bound tree whose nodes correspond to NLP problems and where an edge indicates the addition of a branching bound. If one of the following fathoming rules is satisfied, then no branching is required, the corresponding node has been fully explored (fathomed) and can be abandoned. The fathoming rules are

FR1

An infeasible node is detected. In this case the whole subtree starting at this node is infeasible and the node has been fathomed.

FR2

An integer feasible node is detected. This provides an upper bound on the optimum of the MINLP; no branching is possible and the node has been fathomed.

FR3

A lower bound on the NLP solution of a node is greater or equal than the current upper bound. In this case the node is fathomed, since this NLP solution provides a lower bound for all problems in the corresponding sub-tree.

Once a node has been fathomed the algorithm backtracks to another node which has not been fathomed until all nodes are fathomed.

Many heuristics exist for selecting a branching variable and for choosing the next problem to be solved after a node has been fathomed (see surveys by Gupta and Ravindran [18] and Volkovich et.al. [24]).

Branch-and-bound can be inefficient in practice since it requires the solution of one NLP problem per node which is usually solved iteratively through a sequence of quadratic programming problems. Moreover, a large number of NLP problems are solved which have often no physical meaning if the integer variables do not take integer values.

5.4  Integrating SQP and branch-and-bound

An alternative to nonlinear branch-and-bound for convex MINLP problems is due to Borchers and Mitchell [4]. They observe that it is not necessary to solve the NLP at each node to optimality before branching and propose an early branching rule, which branches on an integer variable before the NLP has converged. Borchers and Mitchell implement this approach and report encouraging numerical results compared to Outer Approximation [5]. In their implementation the SQP solver is interrupted after two QP solves (Ideally one would prefer to interrupt the SQP method after each QP solve. Borchers and Mitchell use an SQP method from the NAG library which does not allow this.) and branching is done on an integer variable which is more than a given tolerance away from integrality.

The drawback of the early branching rule is that since the NLP problems are not solved to optimality, there is no guarantee that the current value of f(x,y) is a lower bound. As a consequence, bounding (fathoming rule FR3) is no longer applicable. Borchers and Mitchell seek to overcome this difficulty by evaluating the Lagrangian dual for a given set of multiplier estimates, λ. The evaluation of the Lagrangian dual amounts to solving where the Y ⊂ Y now also includes bounds that have been added during the branching.

Note that this evaluation requires the solution of a bound constrained nonlinear optimization problem. In order to diminish the costs of these additional calculations, Borchers and Mitchell only evaluate the dual every sixth QP solve.

Clearly, it would be desirable to avoid the solution of this problem and at the same time obtain a lower bounding property at each QP solve. In the remainder of this section it is shown how this can be achieved by integrating the iterative NLP solver with the tree search.

Throughout this section it is assumed that both f and g are smooth, convex functions. The nonconvex case is discussed in Section 5.4.4 where a simple heuristic is proposed. The ideas are first presented in the context of a basic SQP method without globalization strategy. Next this basic algorithm is modified to allow for the use of a line-search or a trust-region. It is then shown how the early branching rule can take account of the quadratic rate of convergence of the SQP method. Finally, a simple heuristic for nonconvex MINLP problems is suggested.

5.4.1  The basic SQP algorithm

This section shows how the solution of the Lagrangian dual can be avoided by interpreting the constraints of the QP approximation as supporting hyperplanes. If f and g are convex, then the linear approximations of the QP are outer approximations and this property is exploited to replace the lower bounding.

Consider an NLP problem at a given node of the branch-and-bound tree, where the integer restrictions have been relaxed and Y ⊂ Y contains bounds that have been added by the branching. Let f denote the solution of (P).

Applying the SQP method to (P) results in solving a sequence of QP problems of the form where f^(k) = f(x^(k),y^(k)) etc. and approximates the Hessian of the Lagrangian,
where λ_i is the Lagrange multiplier of g_i(x,y) ≤ 0.

Unfortunately, the solution of (QP^k) does not underestimate f, even if all functions are assumed to be convex. This implies that fathoming rule FR3 cannot be used if early branching is employed. However, it turns out that it is not necessary to compute an underestimator explicitly. Instead, a mechanism is needed of terminating the sequence of (QP^k) problems once such an underestimator would become greater than the current upper bound, U, of the branch-and-bound process. This can be achieved by adding the objective cut to (QP^k), where є > 0 is the optimality tolerance of branch-and-bound. Denote the new QP with this objective cut by (QP_c^k).

Note that (QP_c^k) is the QP that is generated by the SQP method when applied to the following NLP problem where a nonlinear objective cut has been added to (P). This NLP problem is infeasible, if fathoming rule FR3 holds at the solution to (P).

It is now possible to replace fathoming rule FR3 applied to (P) by a test for feasibility of (QP_c^k) (replacing the bounding in branch-and-bound). This results in the following lemma.



Proof: If (QP_c^k) is infeasible, then it follows that there exists no step d such that Since (12) is an outer approximation of the nonlinear constraints of (P) and (11) underestimates f, it follows that there exists no point (x,y) X × Y such that Thus any lower bound on f at the present node has to be larger than U − є. Thus to within the tolerance є, f ≥ U and fathoming rule FR3 holds. q.e.d.




In practice, an SQP method would use a primal active set QP solver which establishes feasibility first and then optimizes. Thus the test of Lemma 1 comes at no extra cost. The basic integrated SQP and branch-and-bound algorithm can now be stated.

Algorithm 1: Basic SQP and branch-and-bound

Initialization: Place the continuous relaxation of (P) on the tree and set U = ∞.

while (there are pending nodes in the tree) do
	Select an unexplored node.
	repeat (SQP iteration)
		1. Solve (QPck) for a step d(k) of the SQP method.
		2. if ((QPck) infeasible) then fathom node, exit
		3. Set (x(k+1), y(k+1)) = (x(k), y(k)) + (dx(k), dy(k)).
		4. if ((x(k+1), y(k+1)) NLP optimal) then
			if (y(k+1) integral) then
				Update current best point by setting
				(x*,y*) = (x(k+1), y(k+1)), f* = f(k+1) and U = f*.
			else
				Choose a non-integral yi(k+1) and branch.
			endif
			exit
		4. endif
		5. Compute the integrality gap θ = maxi | yi(k+1) − anint(yi(k+1)) |.
		6. if (θ > τ) then
			Choose a non-integral yi(k+1) and branch, exit
		6. endif
end while

Here anint(y) is the nearest integer to y. Step 2 implements the fathoming rule of Lemma 1 and Step 6 is the early branching rule. In [4] a value of τ = 0.1 is suggested for the early branching rule and this value has also been chosen here.

A convergence proof for this algorithm is given in the next section where it is shown how the algorithm has to be modified if a line-search or a trust-region is used to enforce global convergence for SQP. The key idea in the convergence proof is that (as in nonlinear branch-and-bound), the union of the child problems that are generated by branching is equivalent to the parent problem. The fact, that only a single QP has been solved in the parent problem does not alter this equivalence. Finally, the new fathoming rule implies FR3 and hence, any node that has been fathomed by Algorithm 1 can also be considered as having been fathomed by nonlinear branch-and-bound. Thus convergence follows from the convergence of branch-and-bound.

Algorithm 1 has two important advantages over the work in [4]. Firstly, the lower bounding can be implemented at no additional cost (compared to the need to solve the Lagrangian dual in [4]). Secondly, the lower bounding is available at all nodes of the branch-and-bound tree (rather than at every sixth node).
@percent Note that if no further branching is possible at a node, then the algorithm resorts to normal SQP at that node.

5.4.2  The globalized SQP algorithm

It is well known that the basic SQP method may fail to converge if started far from a solution. In order to obtain a globally convergent SQP method, descent in some merit function has to be enforced. There are two concepts which enforce descent: a line-search and a trust-region. This section describes how the basic algorithm of the previous section has to be modified to accommodate these global features.

The use of a line-search requires only to replace d^(k) by α^(k) d^(k), where α^(k) is the step-length chosen in the line-search. Similarly a Levenberg-Marquardt approach would require minimal change to the algorithm.

An alternative to using a line-search is to use a trust region. Trust-region SQP methods usually possess stronger convergence properties and this leads us to explore this approach. In this type of approach the step that is computed in (QP_c^k) is restricted to lie within a trust-region. Thus the QP that is being solved at each iteration becomes where ρ^k is the trust-region radius which is adjusted according to how well the quadratic model of (QP_c,∞^k) agrees with the true function (see [9, Chapter 14.5] for details).

The drawback of using a trust-region in the present context is that the trust-region may cause (QP_c,∞^k) to be infeasible even though the problem without a trust-region (QP_c^k) has a non-empty feasible region. Thus Lemma 1 is no longer valid. However, if (QP_c,∞^k) is infeasible and the trust-region is inactive, then the argument of Lemma 1 still holds and the node can be fathomed.

@percent Otherwise, the feasibility restoration phase has to be entered and this part of the SQP method is briefly described next. The aim of the restoration phase is to get closer to the feasible region by minimizing the violation of the constraints in some norm, subject to the linear constraints x X,   y Y. where and the operation a⁺ = max( 0 , a ) is taken componentwise.

@percent In this phase, an SQP algorithm is applied to minimize (F). Methods for solving this problem include Yuan [31] for the l_∞ norm, El-Hallabi and Tapia [8] for the l₂ norm and Dennis, El-Alem and Williamson [21] for the l₁ norm. These methods converge under mild assumptions similar to those needed for the convergence of SQP methods. The details of the restoration phase are beyond the scope of this manual and we concentrate instead on the interaction between this phase and branch-and-bound. In the remainder it will be assumed that this phase converges.

In order to obtain an efficient MINLP solver it is important to also integrate the restoration phase with the tree-search. After each QP in the restoration phase four possible outcomes arise.

1. (QP_c,∞^k) is feasible. In this case the solver exits the restoration phase and resumes SQP and early branching.
2. (QP_c,∞^k) is not feasible and a step d^(k) is computed for which the trust-region is inactive ( d^(k) < ρ^k). This indicates that the current node can be fathomed by Lemma 1.
3. The feasibility restoration converges to a minimum of the constraint violation (x^*,y^*) with h⁺(x^*,y^*) > 0. This is taken as an indication that (P) has no feasible point with an objective that is lower than the current upper bound and the corresponding node can be fathomed. Note that convergence of the feasibility SQP ensures that the trust-region will be inactive at (x^*,y^*).
4. The feasibility restoration SQP continues if none of the above occur.

Thus the feasibility restoration can be terminated before convergence, if 1 or 2 hold above. In the first case early branching resumes and in the second case, the node can be fathomed.

The modifications required to make Algorithm 1 work for trust-region SQP methods are given below.

Algorithm 2: Trust-region SQP and branch-and-bound

Only steps 1. and 2. need to be changed to:

		1'. Compute an acceptable step d(k) of the trust-region SQP method.
		2'. if ((QPc,∞k) infeasible and d(k)∞ < ρk) then
			fathom node exit
		2'. elseif ((QPc,∞k) infeasible and d(k)∞ = ρk) then
			Enter feasibility restoration phase:
			repeat (SQP iteration for feasibility restoration)
				(a) Compute a step of the feasibility restoration.
				(b) if (d(k)∞ < ρk) then fathom node, exit
				(c) if ((QPc,∞k) feasible) then return to normal SQP, exit
				(d) if (ming+(x,y) > 0) then fathom node, exit
		2'. endif

The following theorem shows that Algorithm 2 converges under standard assumptions (e.g.[7] or [10]).



Proof: The finiteness of Y integer implies that the branch-and-bound tree is finite. Thus the algorithm must terminate after visiting a finite number of nodes. Nodes are only fathomed if they are infeasible, integer feasible or if the lower bounding of Lemma 1 holds. Thus an optimal node must eventually be solved and convergence follows from the convergence of the trust-region SQP method. q.e.d.

5.4.3  Quadratic convergence and early branching

When implementing Algorithm 2 the following adverse behaviour has been observed on some problems.

1. Some integer variables converge to a non-integral solution, namely y_i^(k) → y_i but θ 0.02 < τ, i.e. the integrality gap remains bounded away from zero.
2. The SQP solver converges at second order rate during this time and θ ≫ d^(k) _∞ → 0.

In other words the early branching heuristic is not activated during the second order convergence of the SQP method so that no early branching takes place in the final stages of the SQP solver (as τ = 0.1 is too large).
@percent One way to avoid this would be to reduce τ, but this would simply repeat the problem at a smaller threshold. Choosing τ to be of the order of the tolerance of the SQP solver on the other hand would trigger early branching more often which could duplicate the work to be done if integer variables are converging to an integral solution.

These observations lead to a new early branching heuristic which takes the quadratic rate of convergence of the SQP method into account. Steps 5. and 6. of Algorithm 2 are replaced by

		5'. Compute the integrality gap θ = maxi | yi(k+1) − anint(yi(k+1)) |
		5'. and the experimental order of convergence p := ln( d(k) ∞) / ln( d(k−1) ∞).
		6'. if (θ > τ or (p > 1.5 and d(k)∞ < θ )) then
			Choose a non-integral yi(k+1) and branch, exit
		6'. endif

For a quadratically convergent method one expects that p 2 and once quadratic convergence sets in, that d^(k+1)_∞ d^(k)_∞^p. Thus if d^(k)_∞ < θ, one cannot expect to reduce the integrality gap to zero in the remaining SQP iterations.

The new early branching rule resulted in a small but consistent reduction in the number of QP problems solved (roughly a 5% reduction) compared to the original heuristic.

5.4.4  Nonconvex MINLP problems

In practice many MINLP problems are nonconvex. In this case, the underestimating property of Lemma 1 no longer holds, as linearizations of nonconvex functions are not necessarily underestimators. Thus it may be possible that a node is wrongly fathomed on account of Lemma 1 and this could prevent the algorithm from finding the global minimum.

A rigorous way of extending Algorithm 2 to classes of nonconvex MINLP problems would be to replace the linearizations by linear underestimators (e.g. Horst and Tuy [20, Chapter VI]). However, it is not clear what effect this would have on the convergence properties of the SQP method. Instead the following heuristic is proposed, which consists of replacing 2'. by 2”:

		2”. if ((QPc,∞k) infeasible) then
			Enter feasibility restoration phase:
			repeat (SQP iteration)
				(a) Compute a step of the feasibility restoration.
				(c) if ((QPc,∞k) feasible) then return to normal SQP, exit
				(d) if (ming+(x,y) > 0) then fathom node, exit
		2”. endif

The main difference between 2'. and 2”. is that a node is no longer fathomed if the QP problem is infeasible and the step lies strictly inside the trust-region ( d < ρ). Instead a feasibility problem has to be solved to optimality (Step (d) above) or until a feasible QP problem is encountered (Step (b) above). This heuristic is motivated by the success of nonlinear branch-and-bound in solving some nonconvex MINLP problems. It ensures that a node is only fathomed if it would have been fathomed by nonlinear branch-and-bound.
@percent Clearly, there is no guarantee that the feasibility problem is solved to global optimality if the constraints or the objective are nonconvex.

5.5  Numerical Experience

Nonlinear branch-and-bound and Algorithm 2 have been implemented in Fortran 77 using filterSQP [12] as the underlying NLP solver. The implementation of Algorithm 2 uses a callback routine that is invoked after each step of the SQP method. This routine manages the branch-and-bound tree stored on a stack, decides on when to branch and which problem to solve next. The use of this routine has kept the necessary changes to filterSQP to a minimum.

---

Header	Description
problem	The SIF name of the problem being solved.
n	Total number of variables of the problem.
ni	Number of integer variables of the problem.
m	The total number of constraints (excl. simple bounds).
mn	Number of nonlinear constraints.
k	The dimension of the null–space at the solution.
# NLPs	Number of NLP problems solved (branch-and-bound).
# nodes	Number of nodes visited by Algorithm 2.
# QPs	Total number of QP (and LP) problems solved.
# f-QPs	Number of QPs in feasibility restoration.
CPU	Seconds of CPU time needed for the solve.

Table 3: Description of headers of the result tables

---

The test problems are from a wide range of backgrounds. Their characteristics are displayed in Table 4. The SYNTHES* problems are small process synthesis problems [7]. OPTPRLOC is a problem arising from the optimal positioning of a product in a multi-attribute space [7]. BATCH is a batch plant design problem that has been transformed to render it convex [22]. FEEDLOC is a problem arising from the optimal sizing and feed tray location of a distillation column [29], TRIMLOSS is a trim loss optimization problem that arises in the paper industry [28] and TRIMLON4 is an equivalent nonconvex formulation of the same problem. All problems are input as SIF files [2] plus an additional file to identify the integer variables. The test-problems are available at http://www.mcs.dundee.ac.uk:8080/~ sleyffer/MINLP_TP/.

---

problem	n	ni	m	mn	Description
SYNTHES1	6	3	6	2	Process synthesis
SYNTHES2	11	5	14	3	Process synthesis
SYNTHES3	17	8	23	4	Process synthesis
OPTPRLOC	30	25	30	25	Optimal product location
BATCH	49	24	73	1	Batch plant design
FEEDLOC	90	37	259	89	Distillation column design (nonconvex)
TRIMLOSS	142	122	75	4	Trimloss optimization
TRIMLON4	24	24	27	4	Trimloss optimization (nonconvex)

Table 4: Problem characteristics of test problems

---

Table 3 lists a description of the headers of the result tables. Table 4 list the problem characteristics. The performance of branch-and-bound and Algorithm 2 is compared in Table 5. All runs were performed on a SUN SPARC station 4 with 128 Mb memory, using the compiler options -r8 -O and SUN's f77 Fortran compiler. In both algorithms a depth first search with branching on the most fractional variable was implemented.

---

	Branch-and-bound				Algorithm 2
problem	# NLPs	# QPs	(# f-QPs)	CPU	# nodes	# QPs	(# f-QPs)	CPU
SYNTHES1	5	18	(0)	0.1	5	13	(0)	0.1
SYNTHES2	17	53	(0)	0.4	17	40	(0)	0.3
SYNTHES3	25	80	(0)	0.8	25	64	(3)	0.8
OPTPRLOC	109	491	(119)	8.5	112	232	(28)	5.3
BATCH	143	371	(10)	8.9	181	391	(11)	14.8
FEEDLOC	47	340	(35)	70.8	55	103	(4)	29.0
TRIMLOSS	1336	3820	(160)	254.3	944	1624	(1)	90.3
TRIMLON4	887	1775	(66)	19.0	566	900	(112)	13.3

Table 5: Results for branch-and-bound and Algorithm 2

---

Table 5 shows that branch-and-bound only requires about 2-6 QP solves per node that is solved making it a very efficient solver. By comparison, Algorithm 2 often visits more nodes in the branch-and-bound tree. That is not surprising, as the early termination rule is based on a weaker lower bound than FR2. However, the overall number of QP problems that are being solved is reduced by 20% – 80%. This results in a similar reduction in terms of CPU time compared to branch-and-bound. For the largest problem, Algorithm 2 is 3 times faster than nonlinear branch-and-bound.

@percent These results are roughly in line with what one might expect taking into account that branch-and-bound solves between 3 and 5 QPs per NLP node. They are somewhat better than the results in [5].

Even though problem TRIMLON4 is nonconvex, Algorithm 2 terminated with the global minimum. For the nonconvex problem FEEDLOC, Algorithm 2 did not find an integer feasible point. This behaviour is due to the fact that the linearization are not outer approximations in the nonconvex case and shows that the early termination rule does not hold for nonconvex MINLP problems. Running Algorithm 2 in nonconvex mode (see Section 5.4.4) produced the same minimum as branch-and-bound. As one would expect, the performance of Algorithm 2 in nonconvex mode is not as competitive as in convex mode (see Table 6) but still competitive with branch-and-bound. Better heuristics for nonconvex MINLP problems remain an open problem.

---

	Branch-and-bound				Algorithm 2
problem	# NLPs	# QPs	(# f-QPs)	CPU	# nodes	# QPs	(# f-QPs)	CPU
FEEDLOC	47	340	(35)	70.8	65	159	(75)	55.9
TRIMLON4	887	1775	(66)	19.0	984	1606	(259)	21.9

Table 6: Results for Algorithm 2 in nonconvex mode

---

5.6  Conclusions

We have presented an algorithm for MINLP problems which interlaces the iterative solution of each NLP node with the branch-and-bound tree search.
@percent The algorithm presented here can also be interpreted as an improved lower bounding scheme for branch-and-bound. An implementation of this idea gave a factor of up to 3 improvement in terms of CPU time compared to branch-and-bound.

« Previous « Start » Next »