A Survey for Solving Mixed Integer Programming via Machine Learning

Jiayi Zhang1 , Chang Liu1 , Junchi Yan1∗

Xijun Li2 , Hui-Ling Zhen2 , Mingxuan Yuan2
1
Department of CSE, and MoE Key Lab of Artificial Intelligence, Shanghai Jiao Tong University
2
Noah’s Ark Lab, Huawei Ltd.
{zhangjiayirr, only-changer, yanjunchi}@sjtu.edu.cn
{xijun.li,zhenhuiling2,yuan.mingxuan}@huawei.com

Abstract the right NP-hard problem needed to solve the task, which
arXiv:2203.02878v1 [cs.AI] 6 Mar 2022

demonstrate the importance of MIP in the area of combinato-

This paper surveys the trend of leveraging ma- rial optimization.
chine learning to solve mixed integer programming When dealing with mathematical modeling problems in in-
(MIP) problems. Theoretically, MIP is an NP- dustrial applications mentioned above, decision variables and
hard problem, and most of the combinatorial op- integer variables are sometimes inevitable, which are mod-
timization (CO) problems can be formulated as the eled as MIP models. For example: x cars, y packages. x, y
MIP. Like other CO problems, the human-designed are integer variables here, and decimals are meaningless. One
heuristic algorithms for MIP rely on good initial so- MIP problem may contain both integer and continuous vari-
lutions and cost a lot of computational resources. ables. If the problem contains an objective function with no
Therefore, we consider applying machine learn- quadratic term, then the problem is termed a Mixed Integer
ing methods to solve MIP, since ML-enhanced ap- Linear Programming (MILP). Formally, it is given as:
proaches can provide the solution based on the typ-
ical patterns from the historical data. In this pa- min cT x
per, we first introduce the formulation and prelimi- (1)
s.t. Ax ≥ b,x ∈ Rn xj ∈ Z, ∀j ∈ I
naries of MIP and several traditional algorithms to
solve MIP. Then, we advocate further promoting When there is at least a quadratic term in the objec-
the different integration of machine learning and tive, the problem terms to a Mixed Integer Quadratic Pro-
MIP and introducing related learning-based meth- gram (MIQP). If the model has any constraint containing a
ods, which can be classified into exact algorithms quadratic term, regardless of the objective function, the prob-
and heuristic algorithms. Finally, we propose the lem is termed as a Mixed Integer Quadratic Constrained Pro-
outlook for learning-based MIP solvers, direction gram (MIQCP). In this paper, we only focus on the basic
towards more combinatorial optimization problems MILP, and the MIP term denotes MILP from now on.
beyond MIP, and also the mutual embrace of tradi- A lower bound for optimum of MILP is provided by solv-
tional solvers and machine learning components. ing the corresponding LP relaxation of the MILP. The LP re-
laxation of is obtained by omitted the integer requirements:
1 Introduction min cT x
(2)
Mixed integer programming problems (MIP) are significant s.t. Ax ≥ b, x ∈ Rn
parts of combinatorial optimization (CO) problems. Benefit-
ing from the development of academic theory and commer- Exactly solving the MILP is NP-hard, that is, exponential
cial software, MIP has become a vital capability that powers time solvable. While practitioners have more interested in
a wide range of applications, including planning [Pochet and getting solutions of good quality as quickly as possible than in
Wolsey, 2010; Wu et al., 2013], scheduling [Sawik, 2011; finding a provably optimal solution. The research on how to
Keha et al., 2009], routing [Malandraki and Daskin, 1992; obtain high-quality solutions in a limited time is of practical
Schouwenaars et al., 2001] and bin packing [Gajda et al., great significance.
2022]. Moreover, [Paulus et al., 2021] manages to integrate In this survey, we focus on ML-based algorithmic tech-
integer programming solvers into neural network architec- niques to solve MIP. These basic and well-studied techniques
tures as a differentiable layer capable of learning both the cost can be divided into two main categories, the branch and
terms and the constraints. The resulting end-to-end trainable bound based algorithms for exact solving and the heuristic
architectures are able to simultaneously extract features from algorithms for approximately solving. We will introduce one
raw data and learn a suitable set of constraints that specify by one in the next sections, and discuss other interesting di-
any combinatorial problem. This architecture can learn to fit rections later. In modern MIP solvers, most of them are based
on the branch-and-bound method, and the cutting plane tech-
∗
Correspondence author is Junchi Yan. nique is often integrated into the branch-and-bound method
 +∞ node expansion Algorithm 1: Branch-and-Bound (X, f )
training examples: x = 5/2#
order
+∞ +∞ y≤1
−13/2 y = 3/2
y≥2
global lower and 1 Set L = {X} and
< upper bound
−13/3 −16/3
x = 5/3# +∞
ub = +∞#
lb = −16/3 +∞ x = 5/3# optimal node 2 while L 6= ∅ do
prune:
+∞ y=1 −13/3 −16/3 y=2
fathomed node 3 Select a subproblem S from L to explore
if a solution x̂0 ∈ {x ∈ S | f (x) < f (x̂)} can be
−13/3 x≤1 x≥2 x≤1 x≥2
ub = −3# ub = −3# 4
found then
lb = −16/3 lb = −22/5
−3 +∞
x = 1# INF x = 1# INF

Set x̂ = x̂0
−3 y=1 –22/5 y = 12/5
min −2x − y#
5
y≤2 y≥3
s.t. 3x − 5y ≤ 0#
if S cannot be pruned then
ub = −4#
lb = −4 3x + 5y ≤ 15# 6
x = 1# −4 −3 x = 0#
y=2 −4 −3 y=3
x ≥ 0, y ≥ 0#
x, y ∈ Z 7 Partition S into S1 , S2 , . . . , Sr
8 Insert S1 , S2 , . . . , Sr into L
Figure 1: Using branch-and-bound to solve an integer linear programming minimization.
Figure 1: Use branch-and-bound to solve an integer linear 9 Remove S from L
programming minimization. (Credit to [He et al., 2014]) Output: x̂
• Easy incorporation of heuristics. Most hand-designed strategies handle only a few heuris-
tics, and they set weights on different heuristics by domain knowledge or manual experi-
to mentation.
increaseIn ourthemodel,
efficiency of branch-and-bound
multiple heuristics algorithm.
can be simply plugged in as state features for
the policy, allowing a hybrid “heuristic” to be learned effectively.
These two techniques constitute the branch-and-cut frame- half a century ago [Land and Doig, 1960], B&B is an exact
work.thatThe
We assume common
a small MIP
set of solved solvers
problems includes
are given SCIP
at training time [and
Bestuzheva
the problems to bealgorithm that is proved to reach the optimal solution eventu-
solvedetat test
al.,time
2021 ], the
are of same type.[Cplex,
CPLEX We learn a2009 ] and policy
node selection GUROBI [Gurobi
and a node pruning policy
from solving the training problems. The node selection policy repeatedly picks a node from theally. In this section, we first give an overview of the general
Optimization, 2020 ] . Due to space limitations, this
queue of all unexplored nodes, and the node pruning policy decides if the popped nodepaper willis worthB&B algorithm, and then we divide the existing works into
not cover
expanding. classical
We formulate methods
B&B search useddecision-making
as a sequential in commercial solvers
process. We designfora simple
oracle that knows the optimal solution in advance and only expands nodes containing the optimalbranching variable selection, node selection, cutting plane se-
solving MIP.
solution. We then use imitation learning to learn policies that mimic the oracle’s behavior withoutlection and discuss these methods combined with machine
There are
perfect information; theseemerging
policies mustopen-source
even mimic how thelibraries to facilitate
oracle would thethe ora-
act in states that
cle would not itself reach, as such states may be encountered at test time. We apply our approach tolearning approaches in detail.
research at the intersection of machine learning and mixed
linear programming (LP) based B&B for solving mixed integer linear programming (MILP) prob- Branch-and-bound is a classic tree search method. By
1
lems, integer
and achieveprogramming.
better solutions fasterMIPLearn
on 4 MILP problemoffers a customized
libraries li- fast
than Gurobi, a recent
brary for learning-based solver configuration
commercial solver competitive with Cplex, and SCIP, one of the fastestwhich
open-source solvers [8]. divide-and-conquer, it partitions the search space by branch-
supports ing on variables’ values and smartly uses bounds from prob-
both GUROBI and CPLEX. It can be framed as a border- lem relaxations to prune unpromising regions from the tree.
2 The
lineBranch-and-Bound Framework:which
case of an MDP framework An Application in Mixed
is the focus Integer
of another
Linear Programming [ ]
emerging library Ecole Prouvost et al., 2020 . Ecole is de- Define an optimization problem as P = (X, f ), where
signed
Consider to cooperate
an optimization problem with the state-of-the-art
of minimizing f over a feasible setopen-source
F, where F is usuallyMIPdiscrete. (called the search space) is a set of valid solutions, and
X
B&B solver SCIP
uses a divide
Sp and where
conquer Ecole
strategy: acts
F is as an
recursivelyimprovement
divided into its to
subsets F 1 , F2 , . . . , Fpf : X → R is the objective function. The goal is to find an
existing
such that F = i=1 Fi . The recursion tree is an enumeration tree of all feasible solutions, whoseoptimal solution x ∈ arg minx∈X f (x). In order to solve P,
∗
nodessolvers.
are subproblemsBesides,
and edgesORGym
[Hubbs
are the partition et al.,
conditions. 2020
Slightly
] notation,
abusingand Open- we will usea search tree T of subsets of the search space is build. Ad-
Fi to GraphGym [Zheng
refer to both the subset and itsetcorresponding
al., 2020B&B ] arenodeGym-like
from now on.libraries
A (convex)for relaxation
of each subproblem is solved to provide an upper/lower bound for that node and its descendants. Weditionally, a feasible solution x̂ ∈ X (called the incumbent
learning heuristics for a collection of combinatorial optimiza-
denote the upper and lower bound at node i by `ub (Fi ) and `lb (Fi ) respectively where `ub and `lbsolution) is stored globally. At each iteration, the algorithm
tion problems
are bounding functions. including MIP.
selects a new subsets S ⊆ X to explore from a list L of un-
Difference
A common setting where B&B to existing surveys.
is ubiquitously ThereA MILP
applied is MILP. are excellent sur- has
optimization problem
assume weexplored subsets; if a solution x̂ ∈ S (called a candidate
0
linear veys inand
objective theconstraints,
area ofandmachine
also requireslearning for combinatorial
specified variables to be integer. We op-
are minimizing the objective function in MILP from now on. At each node, we drop the integralityincumbent) can be found with a better objective value than x̂
timization
constraints [Bengio
and solve its et We
LP relaxation. al.,present
2021 ], andexample
a concrete exactly learning
in Figure for
1. The optimization
[Kleinert (i.e., f (x̂0 ) < f (x̂) ), the incumbent solution is updated. On
problemsolving
is shownmixed integer
in the lower programming
right corner. At node i, a local lower boundet(shownal., in2021;
lower half of
each circle) is found by the LP solver. A local upper bound (shown in upper part of the circle) isthe other hand, if it can be proven that no solution in S has
Lodi and Zarpellon, 2017; Huang et al., 2021a]. Compared
available if a feasible solution is found at this node. We automatically get an upper bound if the LPa better objective value than x̂( i.e. , ∀x ∈ S, f (x) ≥ f (x̂)),
with
solution thetoprevious
happens survey
be integer feasible, [Lodi
or we by heuristics. 2017] which
and itZarpellon,
may obtain
the subset is pruned (or fathomed), and the subset is termi-
B&B only
maintains covers
a queuetheL oftopic of learning-supported
active nodes, starting with a single rootnode
node on selection
it. At each step,
we pop techniques
a node Fi from due to the
L using a nodelimited
selectiondevelopment fiveitsyears
strategy, and compute bounds.ago. A node Final. Otherwise, child subsets are generated by partitioning
S into an exhaustive (but not necessarily mutually exclusive)
While the other [Kleinert et al., 2021] is focused on bi-level
2 set of subset S1 , S2 , . . . , Sr , which are then inserted into T .
tailored approaches that exploit MIP techniques to solve bi-
Once no unexplored subset remain, the best incumbent so-
level optimization problems. In this paper, we aim to give a
lution is returned; since the subset is only fathomed if they
comprehensive and up-to-date review in the area of learning
contain no solution better than x̂, it must be the case that
for solving MIP, especially seeing the rapid development in
x̂ ∈ arg minx∈X f (x). Pseudocode for the generic B&B pro-
recent years. The recent paper [Huang et al., 2021a] mainly
cedure is given in Alg. 1, and an example of solving MIP by
focuses on the development in recent decades of the branch-
B&B [He et al., 2014] is shown in Fig. 1.
and-bound algorithm instead of MIP, but it does not empha-
size the importance of adapting machine learning in MIP. One critical reason why B&B is difficult to formalize re-
sides in its inherent exponential nature: millions of Branch-
ing Variable Selection (BVS) decisions could be needed to
2 Branch-and-Bound solve a MILP, and a single bad one could result in a doubled
In general, mixed integer linear programming problems can tree size and no improvement in the search. Such a complex
be solved using a linear-programming based branch-and- and data-rich setting, paired with often a lack of formal un-
bound (B&B) algorithm [Land and Doig, 2010]. Proposed derstanding of the particular problem structure, makes B&B
an appealing ground for machine learning techniques, which
1
https://anl-ceeesa.github.io/MIPLearn has lately been thriving in discrete optimization. Employing
 Table 1: Summary of methods that combine machine learning with B&B. “Selection” denotes which part involves learning.
Method Selection Learning Network Representation Remark
[Gasse et al., 2019] Variable Reinforcement GCN Bipartite graph Imitate strong branching
[Gupta et al., 2020] Variable Supervised GCN Bipartite graph Accelerate via dynamic embedding
[Sun et al., 2020] Variable Reinforcement PD policy Subproblem set Evolution strategy for training
[He et al., 2014] Node Reinforcement Standalone Standalone Imitate optimal oracle
[Yilmaz et al., 2020] Node Supervised MLP Handcraft Prune leaf
[Khalil et al., 2016] Variable Supervised SVM Handcraft Learning to rank
[Shen et al., 2021] Variable Supervised GCN Bipartite graph Combined with DFS
[Huang et al., 2021b] Cutting Supervised MLP Handcraft Large scale
[Zarpellon et al., 2020] Variable Reinforcement MLP Handcraft Imitate strong branching
[Tang et al., 2020] Cutting Reinforcement Attention & LSTM Handcraft Evolution strategy for training
[Nair et al., 2021] Variable Reinforcement GCN Bipartite graph Imitate strong branching
[Ding et al., 2019] Variable Supervised GCN Tripartite graph Extract connection information
[Alvarez et al., 2014] Variable Supervised ExtraTrees Handcraft Imitate strong branching

Algorithm 2: Branching Variable Selection

Input: Subproblem of the current node S with its
optimal LP solution x̂ ∈ / XM ILP
1 Define branching candidates set C = {i ∈ I | x̂i ∈
/ Z}
2 for each candidate i ∈ C do
3 Calculate its score value si ∈ R
Output: A subscript i = arg mini∈C si of an integer
variable with fractional value x̂i ∈
/Z
Figure 2: Transform an MIP instance to a bipartite graph. The
variables (in green) correspond to one side and the constraints
a clever branching rule is critical for MIP solvers. (in red) refers to the other. (Credit to [Nair et al., 2021])
The majority of works focus on decision in the branch-and-
bound algorithm [Balcan et al., 2018; Khalil et al., 2016]. rally, the idea of adopting an imitation learning [Ho and Er-
Two most crucial ones are branching variable selection and mon, 2016] strategy to learn a fast approximation of branch-
node selection. Recent works explore “learning to branch”, ing rules came into being. Imitation learning aims at using
including learning for branching variable selection [Khalil expert experience to conduct the learning of neural networks,
et al., 2016; Song et al., 2020b; Liberto et al., 2016; which can acquire the neural networks that can reach the same
Gasse et al., 2019; Sun et al., 2020; Zarpellon et al., 2020; or even better performance than the initial expert. [Alvarez
Etheve et al., 2020], node selection [He et al., 2014; Song et al., 2014; 2017; Khalil et al., 2016; Gasse et al., 2019;
et al., 2017; 2020b; Yilmaz et al., 2020], and cutting plane Gupta et al., 2020] learn branching policies by imitating the
selection [Huang et al., 2021b; Tang et al., 2020]. In partic- strong branching rule. This kind of approach adopts a high-
ular, branching variable selection is the main stream of the quality but expensive heuristic scheme, which is the earliest
presented articles, since the selection of the next variable is and most widely studied method for machine learning.
significant in B&B and matters to the total cost of B&B. Specifically, [Alvarez et al., 2014] is the first to use su-
pervised learning to learn a strong branching model. By the
2.1 Branching Variable Selection (BVS) observation of the expert, the neural networks can learn an
Branch variable selection determines which fractional vari- approximated function of the variable score, and can further
ables (also known as candidates) to branch the current node adapt it to select the suitable variable in B&B. [Khalil et al.,
into two child nodes. To indicate the quality of a candidate 2016] extends the work of [Alvarez et al., 2014] by design-
variable, a score of this variable is used to measure its effec- ing a novel pipeline that can solve the MIP on the fly. In
tiveness, and the candidate with the highest score is picked to the first 500 branches, they use the traditional SB to make
branch on. The pseudocode of BVS is presented in Alg. 2, decisions while recording the problem features to train the
which illustrates the basic variable selection idea. neural networks, and the neural networks take control of the
There are various branching rules used to measures vari- decision-maker from the 501st branch. Both of them imitate
ables in BVS, including strong branching (SB) rule [Apple- SB successfully, while the spending time is too high due to
gate et al., 1995], pseudo-cost [Bénichou et al., 1971] and the calculation of a large number of features on each node.
hybrid branching [Achterberg and Berthold, 2009]. Natu- To overcome complex feature calculation, [Gasse et al.,
 Algorithm 3: Node Selection
Input: Node list L = {P } and upper bound J¯
1 Assigns a score to each active node
2 Pop a node P i from L
3 Solve the LP relaxation


4 get solution xi , y i and lower bound J i
5 if J i ≥ J¯ then
6 Prune the node P i
Figure 3: From MIP to tripartite graph. Three types of ver- 7 else if y i ∈ Zp then
tices represent objective functions, variables, and constraints, 8 Improve upper bound J¯ = J i
respectively. (Credit to [Ding et al., 2019])
9 else
10 Branch node P i and push child nodes to list L
2019] encodes MIP into a graph-convolutional network
(GCN) [Kipf and Welling, 2017], which can extract the fea-
tures on the graph efficiently by various message passing
approaches. [Nair et al., 2021] encodes MIP to the graph- (MDP). They further design a policy net based on primal-dual
convolutional network (GCN) as a bipartite graph and com- iteration over reduced LP relaxation, which utilizes the power
putes an initial feasible solution (Neural Diving), then trains of evolution strategy to update the neural networks.
a GCN to imitate ADMM-based policy for branching (Neural
Branching). It uses RL to learn a policy that un-assigns one
2.2 Node Selection
variable at a time, interleaved with solving a sub-MIP every η As mentioned above, the branch-and-bound algorithm recur-
steps to compute a new solution. The authors formulate MIP sively divides the feasible set of a problem into disjoint sub-
as a natural bipartite graph representing variable-constraint sets, organized in a tree structure, where each node represents
relationships, which is illustrated in Fig. 2. Each MIP has a subproblem that searches only the subset at that node. The
two sets of nodes, one representing variables and another rep- main steps of the node selection algorithm are given in Alg. 3.
resenting constraints. And an edge between a variable i and If computing bounds on a subproblem does not rule out the
a constraint j means variable i appears in constraint j. A lot possibility that its subset contains the optimal solution, the
of subsequent work has referenced the bipartite graph model subset can be further partitioned (“branched”) as needed. Else
proposed by [Gasse et al., 2019]. Inspired by the utilization if the lower bound of possible solutions of a node is larger
of the GCN model, [Gupta et al., 2020] proposes a hybrid than the known upper bound, the node can be pruned. A key
architecture that uses a GNN model at only the root node of question in B&B is how to prioritize which nodes to consider.
the B&B tree and a weak but fast predictor at the remain- An effective node priority decision strategy guides the tree
ing nodes, such as a simple Multi-Layer Perception (MLP). search to promising areas and improves the chance of quickly
The model combines the superior representation framework finding a good incumbent solution, which can be used to de-
of [Gasse et al., 2019] with the computationally cheaper cide whether to discard or expand other nodes. Thus, learning
framework of [Khalil et al., 2016] to realize a time-accuracy the appropriate node selection policy of a B&B tree is worthy
trade-off in branching. Besides, [Ding et al., 2019] generates of being investigated. More precisely, a policy is a function
a tripartite graph from its MIP formulation, as illustrated in that maps a state to an action, which in this case is the next
Fig. 3. They train a GCN for variable solution prediction on node to be selected.
the collected features, labels, and tripartite graphs. In [Zarpel- There is less work in node selection of B&B than in BVS.
lon et al., 2020], the authors claim that information contained [He et al., 2014] uses imitation learning to train a node se-
in the global branch-and-bound tree state is an important fac- lection and a node pruning policy to speed up the tree search
tor in variable selection, and they use a novel neural network in the B&B process. Following the above work, [Yilmaz et
design that incorporates the branch-and-bound tree state di- al., 2020] obtains a node selector by imitation learning. The
rectly. Furthermore, they are one of the few techniques on difference is that [Yilmaz et al., 2020] learns only to choose
heterogeneous instances. a node’s children it should select. This encourages finding
Different from learning the output of a computationally ex- solutions quickly, as opposed to learning a BFS-like method.
pensive heuristic used in B&B, [Etheve et al., 2020] applies
reinforcement learning (RL) for BVS from scratch, free from 2.3 Cutting Plane
any heuristic. The core idea of RL is to learn from the in- It is known that every MIP can be relaxed to a linear program-
teractions between the agent and the environment, and here ming (LP) by dropping the integer constraints, and there are
the MIP instance is the environment and the agent is try to many traditional efficient algorithms for solving LP such as
solve it. Experience shows that the RL approach sometimes Simplex [Dantzig, 1990], Interior Point Method (IPM) [Roos
works better than the imitation learning we mentioned before. et al., 2005], etc. It means that we can relax MIP to LP and
In [Sun et al., 2020], the authors argue that strong branching solve the incident LP instead at an acceptable cost, which
is not a good expert to imitate and utilize RL by modeling leads to the cutting plane algorithm. As Fig. 4 shows, cutting
the variable selection process as a Markov Decision Process plane methods iteratively add cuts to the relaxed LP, which
Table 2: A brief summary of methods that combine machine learning with heuristic algorithms. The “Focus” column denotes
the combined heuristic algorithm. LNS means large neighborhood search and FP means feasibility pump.
Method Focus Learning Network Representation Remark
[Song et al., 2020a] LNS Reinforcement MLP Handcraft Combined with GUROBI
[Sonnerat et al., 2021] LNS Reinforcement GCN Bipartite graph Use imitation learning
[Wu et al., 2021] LNS Reinforcement GCN Bipartite graph Outperforms SCIP
[Liu et al., 2021] LNS Supervised & Reinforcement GNN Bipartite graph Adaptive neighborhood size
[Qi et al., 2021b] FP Reinforcement MLP & CNN Parameter matrix Combine [A,b] as matrix
[Ding et al., 2020] Pick Supervised GCN Tripartite graph Predict solution value for variables
[Grover et al., 2018] Pick Reinforcement Standalone Handcraft Pick heuristics in CPLEX
[Xavier et al., 2021] Pick Supervised KNN Handcraft As a warm start

proposes a cut ranking method for cut selection in the settings

of multiple instance learning. It uses neural networks to give
scores to different candidate cuts for the next step.

3 Heuristic Algorithms
Though MIP can be solved via B&B exactly, it is time and
resource-consuming due to its NP-hard nature. In many cases
especially large-scale problem instances, B&B becomes in-
tractable. Thus, heuristic-based methods are considered in-
stead. Heuristic algorithms aim to solve MIP approximately,
by the integration of greedy approach and searching. The
Figure 4: Cutting planes for MIP. common heuristic algorithms for MIP are large neighborhood
search and the feasibility pump. Besides, some works aim at
are linear constraints that can tighten the LP relaxation by better utilizing existing MIP solvers by machine learning. We
eliminating some part of the feasible region while preserv- will discuss these approaches in detail.
ing the LP optimal solution. Suppose that we add the cut set
0  0
|C | 3.1 Large Neighborhood Search (LNS)
C = αiT x ≥ βi i=1 to the original formulation in Eq. (1).
Then, the optimization formulation of MIP becomes: Large neighborhood search (LNS) is a powerful heuristic for
MIP. Given the problem instance and the initial feasible solu-
min cT x tion, LNS searches for better candidate solutions among pre-
(3) defined neighborhoods of current solution in each iteration.
s.t. Ax ≥ b, x ∈ Rn αT x ≥ β, xj ∈ Z, ∀j ∈ I
The iteration continues until the search budget (for example
As shown by the new formulation, cuts serve as the pur- the computing time) is used up. Due to the nature of LNS, it is
pose of reducing the LP solution space, which might lead important to prevent the search from falling into a poor local
to a smaller tree in the branch-and-cut algorithm so that the optimum. In general, the size of the neighborhood grows ex-
number of nodes to be searched is significantly reduced. As ponentially as the size of the input problem increase. There-
mentioned before, the cutting plane can be combined with the fore, it is necessary to optimize the LNS algorithm by the
branch and bound algorithm, which constitutes the branch- learning techniques to improve its efficiency.
and-cut framework. Branch-and-cut is known as one of the There are two critical choices to determine the effective-
most commonly used algorithms in modern solvers. ness of LNS: 1) initial solution and 2) search neighborhood
Due to the importance of the cutting plane in solving MIP, at each iteration. [Song et al., 2020a] learns a neighbor-
many researchers try to utilize machine learning technologies hood selection policy using imitation learning and reinforce-
to improve the traditional cutting plane algorithm. Unlike ment learning (RL). It uses a random neighborhood selec-
B&B, there are relatively few well-designated heuristics in tion policy to generate training data for imitation learning.
the cutting plane algorithm, which denotes imitation learn- Following the neural diving idea proposed by [Nair et al.,
ing or supervised learning cannot be directly applied here. 2021] mentioned in Section 2.1, [Sonnerat et al., 2021] adapts
Therefore, researchers begin to think of reinforcement learn- neural diving to obtain the initial solution and for select-
ing. [Tang et al., 2020] introduces a MDP formulation for ing the search neighborhood at each LNS step. Some re-
the problem of sequentially selecting cutting planes for MIP, searchers [Liu et al., 2021] conduct analysis to LNS, and it
and training a reinforcement learning (RL) agent using evo- turns out the size of neighbors is important, and the most suit-
lutionary strategies. This work shows the ability of RL to able neighbor size varies over iterations. Therefore, they pro-
improve the cutting plane algorithm and potentially opens a pose to use machine learning to automatically find the suit-
new research topic. The following work [Huang et al., 2021b] able neighbor size. [Wu et al., 2021] combines RL with LNS,
 Algorithm 4: Feasibility Pump
Input: x0 ← argmin cT x; x0 ← [x0 ]; i = 0
1 while xi is not feasiable do
2 xi+1 ← argmin ||x − xi ||
3 xi+1 ← [xi+1 ]
4 if xi+1 == xi then
5 random perturbation of xij , ∀j ∈ I
6 else Figure 5: The feasibility pump. (Credit to [Qi et al., 2021a])
7 k ←k+1
reasonable to solve MIP based on existing works since exist-
Output: xi ing MIP solvers have been improved for decades. Some re-
searchers believe that better utilization of existing MIP meth-
ods is a valuable research topic. [Ding et al., 2020] predicts
where the action of RL is to select the variable to be replaced. a solution value for each variable based on historical data
They also propose a novel feature extractor for variables and and decides whether to use heuristic algorithms or the ex-
constraints in MIP. act branching approach to solve it. The authors try to make
LNS aims to continuously improve a solution, which is a the exact branching approach focus on the hard case while
common idea in solving CO problems. Therefore, we will leaving the easy case to the heuristic algorithm. By adopt-
discuss some approaches in the CO field that shares similar ing reinforcement learning, [Grover et al., 2018] proposes to
ideas with LNS, which we hope could inspire adaptation of select the predefined heuristics in CPLEX by the features of
these methods to solve MIP. [Chen and Tian, 2019] proposes the given instance, which is an online learning framework.
a framework called local rewrite, which tries to improve a [Xavier et al., 2021] focuses on solving the MIP instances in
given solution by selecting a part of the solution and modify- a data-driven manner. Combined with existing MIP solvers,
ing it. In their paper, the local rewrite framework is proved they train a KNN to predict redundant constraints, good ini-
to be a powerful method for the vehicle routing problem and tial feasible solutions, and affine subspaces where the optimal
computing resource allocation. The ECO-DQN [Barrett et solution is likely to lie, which can lead to a significant reduc-
al., 2020] framework re-designs the action space of the rein- tion in the problem size of MIP. They conduct experiments on
forcement learning agent, which allows revoking the previous the electric grid unit commitment problem, which is an appli-
action. In other words, the action of the agent is revocable in cation of MIP. Besides, the early stopping technique is widely
the ECO-DQN framework. There are many other works [Lu adopted in the hyperparameter optimization [Makarova et al.,
et al., 2020; Fu et al., 2021] following the idea of local rewrite 2021], which uses machine learning to predict when to stop
and ECO-DQN in combinatorial optimization. the searching process without much loss of quality. It greatly
improves the efficiency and saves computing resources. The
3.2 Feasibility Pump early stopping technique might be integrated into the MIP
solvers. Specifically, one can use machine learning tech-
Feasibility pump (FP) is a heuristic algorithm that runs the
niques to predict when to early stop the B&B process mean-
following steps: 1) find the rounded optimal continuous re-
while the incumbent primal solution is still acceptable.
laxation solution of the MIP (degenerating to LP); 2) search
for the nearest solution in the relaxed feasible region; 3) per-
turb and round the new solution found at each step until the 4 Conclusion and Outlook
solution is feasible. If the limit of the maximum number of In this survey, a study of the state-of-the-art machine learning
steps is reached, the algorithm will halt and return the cur- approach for solving MIP is presented, to summarize exist-
rent feasible solution as the output. The basic steps of the FP ing work and serve as a starting point for future research in
algorithm are shown in Algorithm 4. these areas. We find that the integration of machine learning
Though FP is a relatively powerful heuristic, the efficiency techniques and traditional operational research algorithms is
and cost-effectiveness are not satisfied. Therefore, [Qi et al., a raising topic in the research field, including combination
2021a; 2021b] utilize RL to improve FP. The entry point of with the exact B&B algorithms and heuristic algorithms.
RL is to find the next non-integer solution in step 2). Instead Since MIP is an NP-hard problem, it is very difficult to
of choosing the nearest solution, the authors let the RL agent obtain an exact solution. Leveraging machine learning tech-
choose the next solution as its action. Besides, two methods niques to obtain an acceptable solution within limited com-
of MLP and CNN are designed on the representation of the puting resources is welcomed and reasonable in practical ap-
state space, where the CNN is to treat the parameters [A, b] plications. We can design a model, by instructing learning
in MIP as an image and process it. Fig. 4 shows the FP algo- models to imitate heuristic algorithms, and make some deci-
rithm and the smart FP proposed by [Qi et al., 2021a]. sions in the B&B algorithm or make adjustments to the initial
solution. It can be seen that with the development of ma-
3.3 Predict and Pick chine learning, especially for deep and reinforcement learn-
Instead of solving the MIP directly, some works aim to pre- ing, more and more models are used in MIP solving, contin-
dict and pick from the existing MIP solvers or methods. It is uously improving the efficiency and solution quality.
References [Etheve et al., 2020] M. Etheve, Z. Alès, C. Bissuel, O. Juan,
and S. Kedad-Sidhoum. Reinforcement learning for vari-
[Achterberg and Berthold, 2009] T. Achterberg and able selection in a branch and bound algorithm. Arxiv,
T. Berthold. Hybrid branching. In CPAIOR, 2009. abs/2005.10026, 2020.
[Alvarez et al., 2014] A. M. Alvarez, Q. Louveaux, and [Fu et al., 2021] Z.-H. Fu, K.-B. Qiu, and H. Zha. Generalize
L. Wehenkel. A supervised machine learning approach to a small pre-trained model to arbitrarily large tsp instances.
variable branching in branch-and-bound. In ECML, 2014. In AAAI, 2021.
[Alvarez et al., 2017] A. M. Alvarez, Q. Louveaux, and [Gajda et al., 2022] M. Gajda, A. Trivella, R. Mansini, and
L. Wehenkel. A machine learning-based approximation D. Pisinger. An optimization approach for a complex real-
of strong branching. JOC, 2017. life container loading problem. Omega, 2022.
[Applegate et al., 1995] D. Applegate, R. Bixby, V. Chvatal, [Gasse et al., 2019] M. Gasse, D. Chételat, N. Ferroni,
and B. Cook. Finding cuts in the tsp. Technical report, L. Charlin, and A. Lodi. Exact combinatorial optimiza-
1995. tion with graph convolutional neural networks. Arxiv,
abs/1906.01629, 2019.
[Balcan et al., 2018] M.-F. Balcan, T. Dick, T. Sandholm,
and E. Vitercik. Learning to branch. In ICML, 2018. [Grover et al., 2018] A. Grover, T. Markov, P. Attia, N. Jin,
N. Perkins, B. Cheong, M. Chen, Z. Yang, S. Harris,
[Barrett et al., 2020] T. Barrett, W. Clements, J. Foerster, and W. Chueh, et al. Best arm identification in multi-armed
A. Lvovsky. Exploratory combinatorial optimization with bandits with delayed feedback. In AISTATS, 2018.
reinforcement learning. In AAAI, 2020.
[Gupta et al., 2020] P. Gupta, M. Gasse, E. B. Khalil, M. P.
[Bengio et al., 2021] Y. Bengio, A. Lodi, and A. Prou- Kumar, A. Lodi, and Y. Bengio. Hybrid models for learn-
vost. Machine learning for combinatorial optimization: a ing to branch. Arxiv, abs/2006.15212, 2020.
methodological tour d’horizon. EJOR, 2021. [Gurobi Optimization, 2020] Gurobi Optimization. Gurobi
[Bénichou et al., 1971] M. Bénichou, J.-M. Gauthier, optimizer reference manual. 2020.
P. Girodet, G. Hentges, G. Ribière, and O. Vincent. [He et al., 2014] H. He, H. Daume III, and J. M. Eisner.
Experiments in mixed-integer linear programming. Learning to search in branch and bound algorithms. In
Mathematical Programming, 1971. NeurIPS, 2014.
[Bestuzheva et al., 2021] K. Bestuzheva, M. Besançon, W.- [Ho and Ermon, 2016] J. Ho and S. Ermon. Generative ad-
K. Chen, A. Chmiela, T. Donkiewicz, J. van Doorn- versarial imitation learning. In NeurIPS, 2016.
malen, L. Eifler, O. Gaul, G. Gamrath, A. Gleixner, [Huang et al., 2021a] L. Huang, X. Chen, W. Huo, J. Wang,
L. Gottwald, C. Graczyk, K. Halbig, A. Hoen, C. Ho-
F. Zhang, B. Bai, and L. Shi. Branch and bound in mixed
jny, R. van der Hulst, T. Koch, M. Lübbecke, S. J. Ma-
integer linear programming problems: A survey of tech-
her, F. Matter, E. Mühmer, B. Müller, M. E. Pfetsch,
niques and trends. arXiv preprint arXiv:2111.06257, 2021.
D. Rehfeldt, S. Schlein, F. Schlösser, F. Serrano, Y. Shi-
nano, B. Sofranac, M. Turner, S. Vigerske, F. Wegschei- [Huang et al., 2021b] Z. Huang, K. Wang, F. Liu, H. ling
der, P. Wellner, D. Weninger, and J. Witzig. The SCIP Zhen, W. Zhang, M. Yuan, J. Hao, Y. Yu, and J. Wang.
Optimization Suite 8.0. ZIB-Report 21-41, Zuse Institute Learning to select cuts for efficient mixed-integer pro-
Berlin, December 2021. gramming. Pattern Recognition, 2021.
[Chen and Tian, 2019] X. Chen and Y. Tian. Learning to per- [Hubbs et al., 2020] C. D. Hubbs, H. D. Perez, O. Sarwar,
form local rewriting for combinatorial optimization. In N. V. Sahinidis, I. E. Grossmann, and J. M. Wassick. Or-
NeurIPS, 2019. gym: A reinforcement learning library for operations re-
search problems. arXiv:2008.06319, 2020.
[Cplex, 2009] I. I. Cplex. V12. 1: User’s manual for
[Keha et al., 2009] A. B. Keha, K. Khowala, and J. W.
cplex. International Business Machines Corporation,
Fowler. Mixed integer programming formulations for sin-
46(53):157, 2009.
gle machine scheduling problems. Computers & Industrial
[Dantzig, 1990] G. B. Dantzig. Origins of the simplex Engineering, 2009.
method. In A history of scientific computing, pages 141– [Khalil et al., 2016] E. B. Khalil, P. L. Bodic, L. Song,
151. 1990. G. Nemhauser, and B. Dilkina. Learning to branch in
[Ding et al., 2019] J. Ding, C. Zhang, L. Shen, S. Li, mixed integer programming. In AAAI, 2016.
B. Wang, Y. Xu, and L. Song. Optimal solution predic- [Kipf and Welling, 2017] T. Kipf and M. Welling. Semi-
tions for mixed integer programs. Arxiv, abs/1906.09575, supervised classification with graph convolutional net-
2019. works. ArXiv, abs/1609.02907, 2017.
[Ding et al., 2020] J.-Y. Ding, C. Zhang, L. Shen, S. Li, [Kleinert et al., 2021] T. Kleinert, M. Labbé, I. Ljubić, and
B. Wang, Y. Xu, and L. Song. Accelerating primal solu- M. Schmidt. A survey on mixed-integer programming
tion findings for mixed integer programs based on solution techniques in bilevel optimization. EURO Journal on
prediction. In AAAI, 2020. Computational Optimization, 2021.
[Land and Doig, 1960] A. H. Land and A. G. Doig. An auto- [Schouwenaars et al., 2001] T. Schouwenaars, B. De Moor,
matic method of solving discrete programming problems. E. Feron, and J. How. Mixed integer programming for
Econometrica, 28:497, 1960. multi-vehicle path planning. In ECC, 2001.
[Land and Doig, 2010] A. H. Land and A. G. Doig. An auto- [Shen et al., 2021] Y. Shen, Y. Sun, A. Eberhard, and X. Li.
matic method for solving discrete programming problems. Learning primal heuristics for mixed integer programs. In
50 Years of Integer Programming 1958-2008: From the IJCNN, 2021.
Early Years to the State-of-the-Art, 2010. [Song et al., 2017] J. Song, R. Lanka, A. Zhao, Y. Yue, and
[Liberto et al., 2016] G. D. Liberto, S. Kadioglu, K. Leo, M. Ono. Learning to search via self-imitation with appli-
and Y. Malitsky. Dash: Dynamic approach for switching cation to risk-aware planning. In Proc. Adv. Neural Inform.
heuristics. EJOR, 2016. Process. Syst. Workshop, 2017.
[Liu et al., 2021] D. Liu, M. Fischetti, and A. Lodi. Learning [Song et al., 2020a] J. Song, R. Lanka, Y. Yue, and B. Dilk-
to search in local branching. ArXiv, 2021. ina. A general large neighborhood search framework for
[Lodi and Zarpellon, 2017] A. Lodi and G. Zarpellon. On solving integer linear programs. arXiv:2004.00422, 2020.
learning and branching: a survey. Top, 2017. [Song et al., 2020b] J. Song, R. Lanka, Y. Yue, and M. Ono.
[Lu et al., 2020] H. Lu, X. Zhang, and S. Yang. A learning- Co-training for policy learning. In Uncertainty in Artificial
based iterative method for solving vehicle routing prob- Intelligence, pages 1191–1201. PMLR, 2020.
lems. In ICLR, 2020. [Sonnerat et al., 2021] N. Sonnerat, P. Wang, I. Ktena,
[Makarova et al., 2021] A. Makarova, H. Shen, V. Perrone, S. Bartunov, and V. Nair. Learning a large neigh-
A. Klein, J. B. Faddoul, A. Krause, M. Seeger, and C. Ar- borhood search algorithm for mixed integer programs.
chambeau. Overfitting in bayesian optimization: an em- arXiv:2107.10201, 2021.
pirical study and early-stopping solution. arXiv preprint [Sun et al., 2020] H. Sun, W. Chen, H. Li, and L. Song.
arXiv:2104.08166, 2021. Improving learning to branch via reinforcement learning.
[Malandraki and Daskin, 1992] C. Malandraki and LMCA, 2020.
M. Daskin. Time dependent vehicle routing prob- [Tang et al., 2020] Y. Tang, S. Agrawal, and Y. Faenza. Re-
lems: Formulations, properties and heuristic algorithms. inforcement learning for integer programming: Learning
Transportation Science, 1992. to cut. In ICML, 2020.
[Nair et al., 2021] V. Nair, S. Bartunov, F. Gimeno, I. von [Wu et al., 2013] T. Wu, K. Akartunalı, J. Song, and L. Shi.
Glehn, P. Lichocki, I. Lobov, B. O’Donoghue, N. Son- Mixed integer programming in production planning with
nerat, C. Tjandraatmadja, P. Wang, R. Addanki, T. Ha- backlogging and setup carryover: Modeling and algo-
puarachchi, T. Keck, J. Keeling, P. Kohli, I. Ktena, Y. Li, rithms. DEDS, 2013.
O. Vinyals, and Y. Zwols. Solving mixed integer programs [Wu et al., 2021] Y. Wu, W. Song, Z. Cao, and J. Zhang.
using neural networks. arXiv:2012.13349, 2021.
Learning large neighborhood search policy for integer pro-
[Paulus et al., 2021] A. Paulus, M. Rolı́nek, V. Musil, gramming. NeurIPS, 2021.
B. Amos, and G. Martius. Comboptnet: Fit the right
[Xavier et al., 2021] Á. S. Xavier, F. Qiu, and S. Ahmed.
np-hard problem by learning integer programming con-
straints. In ICML, 2021. Learning to solve large-scale security-constrained unit
commitment problems. INFORMS Journal on Computing,
[Pochet and Wolsey, 2010] Y. Pochet and L. A. Wolsey. 33(2):739–756, 2021.
Production Planning by Mixed Integer Programming.
[Yilmaz et al., 2020] K. Yilmaz, N. Yorke-Smith, and xxx.
Springer, 2010.
Learning efficient search approximation in mixed integer
[Prouvost et al., 2020] A. Prouvost, J. Dumouchelle, branch and bound. Arxiv, abs/2007.03948, 2020.
L. Scavuzzo, M. Gasse, D. Chételat, and A. Lodi. Ecole:
A gym-like library for machine learning in combinatorial [Zarpellon et al., 2020] G. Zarpellon, J. Jo, A. Lodi, and
optimization solvers. arXiv:2011.06069, 2020. Y. Bengio. Parameterizing branch-and-bound search trees
to learn branching policies. Arxiv, abs/2002.05120, 2020.
[Qi et al., 2021a] M. Qi, M. Wang, and Z.-J. Shen. Smart
feasibility pump: Reinforcement learning for (mixed) in- [Zheng et al., 2020] W. Zheng, D. Wang, and F. Song. Open-
teger programming. arXiv:2102.09663, 2021. graphgym: A parallel reinforcement learning framework
for graph optimization problems. In International Confer-
[Qi et al., 2021b] M. Qi, M. Wang, Zuo-jun, and M. Shen. ence on Computational Science, 2020.
Reinforcement learning for (mixed) integer programming:
Smart feasibility pump. In RL4RealLife Workshop of
ICML, 2021.
[Roos et al., 2005] C. Roos, T. Terlaky, and J.-P. Vial. Inte-
rior point methods for linear optimization. 2005.
[Sawik, 2011] T. Sawik. Scheduling in supply chains using
mixed integer programming. John Wiley & Sons, 2011.