Clustering Approach for Solving Traveling

Salesman Problems via Ising Model Based Solver

Akira Dan, Riu Shimizu, Takeshi Nishikawa, Song Bian, Takashi Sato
Kyoto University
{dan, riushimizu, sbian}@easter.kuee.kyoto-u.ac.jp, [email protected], [email protected]

Abstract—Ising model based solver have gained increasing 1.0

attention due to their efficiency in finding approximate solutions
for combinatorial optimization problems. However, when solving 0.8
doubly constrained problems, such as traveling salesman problem

Probability
using the Ising model-based solver, both the execution speed 0.6
and the quality of solutions deteriorate significantly due to
the quadratically increasing spin counts and strong constraints 0.4
placed on the spins. In this paper, we propose a recursive
clustering approach that accelerates the calculations of the Ising 0.2
model and that also helps to obtain high-quality solutions.
Through evaluations using the TSP benchmarks, the qualities 0.0
with the proposed method have been improved by up to 67.1% 6 7 8 9 10 11 12 13
The number of vertices
and runtime were reduced by 73.8x. Fig. 1. The probability of finding an optimal solution using the Ising model
solver for different sizes of TSPs.
I. I NTRODUCTION
Since combinatorial optimization problems are abundant
in the real world, development of efficient solvers for such spins, which represent the city number and visiting order.
problems is important. In general, due to the combinatorial The interconnections between spins expresses the constraints
explosion of the search space, it is difficult to find exact and distance. The objective function is formulated as the
solutions to those optimization problems as the problem size summation of the distances along the route and the number of
increases. To address these problems, various approaches have constraint violations. However, there are several issues when
been proposed [1]–[3]. using this setup. First, the graph becomes quadratically large
Ising model-based solvers [4]–[6] have recently attracted in terms of node and edge counts, thus making it difficult
increased attention as efficient methods for finding approxi- to implement when using hardware that physically places
mate solutions to combinatorial optimization problems. The nodes as flip-flops and routes the wires with the weights
Ising model is a mathematical model, which is commonly representing the interactions between the nodes. Second, most
used in statistical mechanics to simulate the properties of of the solutions given by the Ising model fail to satisfy the
ferromagnetism [7]. In the Ising model-based solvers, a spin constraints when the solution quality is addressed. Alterna-
having one of two states, either {+1, −1} is connected to tively, if the constraint term is emphasized, the solution quality
adjacent spins with interactions. Two procedures for each will significantly deteriorate. Furthermore, as long as naive
spin are alternatively carried out to find the spin states that Ising model solvers are used, the probability of obtaining an
minimize the total energy of the system: 1) a deterministic optimal solution is very low for TSPs of any useful size. Fig. 1
state transition called annealing that lowers the local energy shows the probability of finding the optimal solution using a
of a spin calculated by the interaction with adjacent spins, naive Ising model-based solver. For problems larger than 12
and 2) a stochastic state transition called random flip that cities, it is nearly impossible to obtain an optimal solution,
inverts the states of randomly selected spins to help escape and even a satisfactory solution is not possible for problems
from local minima. Various Ising model-based solvers have that involve more than 30 cities. For this reason, there is not
been proposed thus far, such as quantum annealing computers much published work reporting on the applications of Ising
and CMOS annealing machines [8]–[10]. model-based solvers for such strongly constrained problems
The traveling salesman problem (TSP) is a well-known such as the TSP.
combinatorial optimization problem. In the TSP, we find the In this paper, we propose an efficient algorithm for solv-
route with the lowest total travel cost for visiting all cities ing TSPs via the Ising model-based solver. Starting from a
once, given the distance between the cities. The TSP can be coarse outline, the travel route is refined by a hierarchical
applied to a wide range of problems such as printed circuit application of clustering and Ising-based optimization. More
board wiring, transportation route planning, scheduling, and specifically, the vertices are clustered to limit the visiting order
protein structure analysis. The development of solvers that of the vertices, and thereby eliminates the spins and their
obtain high-quality approximate solutions has continued in interactions. With this preprocessing approach, the probability
terms of both hardware and software [11]–[13]. of obtaining solutions that satisfy the constraints is improved
When solving TSPs using the Ising model-based solver, and the solution quality is also improved. Our contributions
n-cities are modeled by a lattice-like graph that uses n2 are summarized as follows.

978-1-7281-1085-1/20/$31.00 ©2020 IEEE

Authorized licensed use limited to: Shanghai Jiaotong University. Downloaded on February 26,2023 at 09:10:24 UTC from IEEE Xplore. Restrictions apply.
 𝑢 H Spin update Algorithm 1 Annealing
1: Initialize spin randomly.
2: T ← Ts
𝑙 𝑖 𝑟 3: while ⌊n2 T + 0.5⌋ > 0 do
4: NRF ← ⌊n2 T ⌋
Random flip
𝑑 5: Do random flip.
Solution space 6: Select ⌊n2 R⌋ spins randomly (σ1 , σ2 , · · · , σ⌊n2 R⌋ ).
Fig. 2. Typical structure of the Fig. 3. Optimization by annealing 7: for z = 1 to ⌊n2 R⌋ do
Ising model. with random flip. 8: σz ← argmin H( σz )
9: end for
10: T ← αT
• We propose a novel clustering-based algorithm that sig- 11: if H < HBest then
nificantly reduces both the number of spins and interac- 12: HBest ← H
tions for solving the TSP. This significantly increases the 13: SBest ← S
size of tractable problems for Ising model-based solvers. 14: end if
• We propose a hierarchical application of the Ising model- 15: end while
based solver to obtain coarse to fine-grained solutions, 16: Output spin state SBest as solution.
thereby accelerating the calculations.
The rest of this paper is organized as follows. In Sec. II, the
preliminary steps for applying the Ising model solver to TSP
2) Annealing with random flip: Fig. 3 shows a conceptual
are explained. Then, Sec. III describes the proposed clustering-
picture of the optimal solution search by annealing and random
based method. Sec. IV evaluates the proposed method com-
flips. In annealing, the search is performed by the local energy
pared with the existing methods. In Sec. V, the conclusions
minimization of each spin using Eq. (1). This critical action
of this paper are presented.
is called spin update.
II. BACKGROUND Annealing is performed according to Alg. 1, where temper-
A. Ising model-based solver ature T represents the ratio of spins inverted in random flips,
A part of a two-dimensional Ising model is depicted in Ts is the initial temperature, R is the update coefficient that
Fig. 2. The spins, each having either an upward (+1) or represents the ratio of spins updated, α is the cooling rate of
downward (−1) state, are arranged in a lattice-like form. the temperature to reduce the ratio of the number of spins
Through interactions with adjacent spins and combined with selected by random flips, and HBest and SBest are the best
the external bias for each spin, the local energy of spin i is result and its spin state, respectively.
expressed as: Although the spin updates decrease the local energy, it
∑ is highly likely that the local energy will fall into a local
Hi (σi ) = − Jij σi σj − hi σi , (1) minimum from the total energy standpoint. In order to help
j∈{u,d,r,l} the model to escape from a local minimum, it is necessary to
temporarily increase the energy. For that purpose, Ising model
where j ∈ {u, d, r, l} is the subscripts of the adjacent spins,
based solvers occasionally execute random flip, in which the
σi , σj ∈ {−1, +1} are the spin values, Jij is the interaction
values of randomly selected spins are inverted. By gradually
between spins i and j, and hi is the external magnetic field
reducing the number of spins, the spin states in the model
(bias) applied to spin i.
converge to the global minimum value.
The total energy of the Ising model is expressed as shown
in Eq. (2) as the sum of all spin energies. 3) Interpretation: Interpret the all the spin states and con-
∑∑ ∑ vert them backward to the solution of the original problem.
H=− Jij σi σj − hi σi (2)
i j i B. Solving the TSP using the Ising model-based solver
Minimization of the total energy is achieved by updating each TSPs are combinatorial optimization problems that find the
spin in the direction that lowers its energy. The spin energy smallest sum of weights (distance) in a circuit that passes
evaluations through the interactions with adjacent spins can through all vertices in a graph once. A n-vertices TSP is
be carried out in parallel. Hence, the total energy of the Ising formulated as:
model can be minimized very efficiently with a hardware ∑∑
implementation. minimize Wkl aik a(i+1)l
In general, solving a combinatorial optimization problem k̸=l i
∑ ∑
using the Ising model-based solver proceeds in the following subject to ∀i, aik = 1, ∀k, aik = 1, aik ∈ {0, 1}.
steps, Mapping, Annealing with random flip, and Interpreta- k=1 i=1
tion [14]. (3)
1) Mapping: The formulated combinatorial optimization The variable aik represents that vertex k is (is not) visited as
problem is transformed into the form of Eq. (2) [15]. The the i-th city when aik = 1(0). Wkl represents the weight (or
mapping of the TSP is described in detail in the next subsec- distance) between vertices k and l. Since the total number of
tion. possible routes is (n − 1)!/2, it is currently difficult to obtain

Authorized licensed use limited to: Shanghai Jiaotong University. Downloaded on February 26,2023 at 09:10:24 UTC from IEEE Xplore. Restrictions apply.
its optimal solution by brute force methods in a realistic time → Vertex
A B C D E F
↓ Order
frame as n becomes larger. 1
1) Mapping: In this paper, without loss of generality, we E 2
assume two-dimensional Cartesian coordinates are given for
D
n vertices (cities) in the TSP. We explain a mapping for C
3

solving the TSP via the Ising model-based solver. Following 4

the definition in Eq. (3), we introduce the following binary F A 5

variable. B
{ 6
1 (Vertex k is visited as i−th city)
aik = (4)
0 (otherwise) (a) (b)

If we consider the energy function of the quadratic, uncon- Fig. 4. Ising spins in conventional mapping.
strained binary optimization (QUBO) format [16] using this
binary variable, the TSP is expressed by:
∑∑ fully connected by the links to all spins in the same row
H = A Wkl aik a(i+1)l σi∗ , to those in the same column σ∗,k , and those in the
k̸=l i upper and lower rows σi−1,∗ and σi+1 (the row numbers
∑∑ ∑∑ are wrapped around for the circuit to close). Hence, mapping
+ B ( aik − 1)2 + C ( aik − 1)2 . (5)
using hardware becomes very difficult as n becomes larger.
i k k i
The minor embedding is also known as NP-hard [18] and is
The first term represents the total weight of the circuit, which thus difficult to apply.
is the objective function. The second term is a constraint that 2) Random flip for TSP: According to the definition of
represents a penalty for visiting multiple vertices ∑as the i-th random flips in the Ising model, the spins are randomly
visit, which takes a minimum value of 0 when k aik = 1. selected and their values are inverted [9]. The number of
The third term is also a constraint representing the penalty for selected spins is gradually reduced as annealing is repeated.
∑ k twice or more, and takes the minimum
visiting the vertices During these procedures, random flips bring the spins to a
value 0 when i aik = 1. A, B, and C are hyperparameters state that does not satisfy the constraints in Eqs. (8) and (9).
of positive values, which balance the relative strength of the In order to cause the spins to satisfy the constraints, we assign
objective function and the constraints [17]. heavier weights for HB and HC . In that case, the constraints
Here, using Eq. (6), aik ∈ {0, 1} is converted to Ising model may be satisfied but the solution obtained deteriorates. If the
spin σik ∈ {−1, 1}. hyperparameters for the constraints are weakened, we may end
σik + 1 up with no solution that satisfies the constraints.
aik = (6) 3) Interpretation: After repeating annealing with random
2
flips, the spin states will be interpreted as the solution for the
Then, the first, second, and third terms of Eq. (5) are repre-
TSP. The legal solution corresponds to the states in which
sented as follows, respectively.
for all rows and all columns, there is only one spin that is
A ∑∑ A ∑∑ valued at 1, and those of other spins are all −1. According
HA = Wkl σik σ(i+1)l + Wkl σik +const.
(7)
4 2 to the definition, σik = 1 can be interpreted as “k-th vertex
k̸=l i k̸=l i
is the i-th city to visit.” Note that the illegal solutions that
B ∑∑∑ n − 2 ∑∑ do not satisfy any of the constraints are practically useless.
HB = σik σil + B σik +const.(8)
4 i 2 i Due to the use of n2 spins, finding a good hyperparameter set
k l k
C ∑∑∑ n − 2 ∑∑ that gives the allowed solutions is difficult as it contradicts the
HC = σik σjk + C σik +const.(9) parameters that improve the quality of the solutions.
4 i j
2 i
k k
III. H IERARCHICAL I SING M ODEL S OLVER
In Eqs. (7),(8), and (9), the first term is the interaction between
spins, the second term is the bias applied to the spin, and the A. Key Concept
third term is a constant that can be ignored in minimization. In the conventional Ising model-based solvers, the solutions
Thus, Eq. (10) is the energy function when solving the TSP that are both legitimate and of high quality are increasingly
via the Ising model-based solver. difficult to obtain as the size of the problem becomes larger.
To alleviate this issue, we propose a hierarchical clustering
H = HA + HB + HC (10)
approach.
As shown in Fig. 4, Eq. (10) is represented by n × n spins We consider TSPs, Vp = {vp,1 , vp,2 , . . . , vp,np }, where vp,k
that are arranged in a lattice form. By comparing Eq. (2) is a coordinate of the vertices and p distinguishes the TSPs in
and Eq. (10), the interactions between spins and the bias different hierarchies. The number of the vertices is np for the
of each spin are determined. Typically, for hardware-based TSP of the p-th hierarchy.
Ising model-based solvers, the number of interactions that can As shown in Fig. 5, the proposed method is divided into
be coupled to a spin is limited due to routing congestion. two phases, hierarchical vertex clustering and incremental
When an n-vertex TSP is mapped, the number of spins is annealing. Starting from the original problem V0 , the vertices
n2 . In addition, it is also as shown in Fig. 4; a spin σik is are recursively clustered to reduce the number of nodes. After

Authorized licensed use limited to: Shanghai Jiaotong University. Downloaded on February 26,2023 at 09:10:24 UTC from IEEE Xplore. Restrictions apply.
Algorithm 2 Optimization with Hierarchical clustering Hierarchical vertex clustering
1: Set K1 , K2 , . . . , Km (K1 > K2 > · · · > Km > 0) V0 V1 V2
2: for p = 0 to m − 1 do
3: Clustering Vp into Kp+1 clusters.
4: Compute the centroids of the vertices of each cluster.
5: Define the centroids as a new T SP Vp+1 .
6: end for
7: Mapping Vm to the Ising model.
8: Do Algorithm 1.
9: for p = 0 to m − 1 do
10: Determine the order of cluster visits for Vm−p−1 .
11: Mapping Vm−p−1 to the Ising model to correspond to
the clustering results and the cluster visit order.
12: Do Algorithm 1.
13: end for
14: Output spin state SBest as solution of the original problem Incremental annealing
V0 . Fig. 5. Example of hierarchical clustering and incremental annealing.

Next, we return to the hierarchy at one level higher. Recall

that, by repeatedly using the Ising model solver, the solution that each vertex in the already solved problem of Vp is a
of TSP Vi for hierarchy i is calculated while preserving the centroid. That is, each vertex in p-th level TSP corresponds to a
coarsened route obtained in the lower level TSP Vi+1 . group of vertices in the (p−1)-th level. Since we have already
Alg. 2 summarizes the procedure of the proposed method calculated the optimal route for Vp , we arrange the nodes in
for solving a TSP, where m is the maximum coarsening level, Vp−1 according to the visit order found by solving Vp . These
and K1 , K2 , . . . , Km are the number of clusters divided by the procedures are illustrated in Fig. 6. Fig. 6(a) shows the clusters
p-th clustering. Hierarchical vertex clustering is first conducted of nodes in level k. In the coarsest level, k + 1, the visit order
in lines 2–6, and incremental annealing follows in lines 7–13. has been determined as shown by the numbers in circles. In the
k-th level optimization, we preserve this order when mapping
the spins on the graph. Fig. 6(b) shows the mapping result for
B. Hierarchical vertex clustering (a). Because the vertex groups (A, B, and C), (D and E), and
In Fig. 5, V0 is the original TSP that we wish to solve. (F) should be visited in that order, unnecessary spins will not
We first cluster the vertices (represented by the solid dots in be generated. In the case of this example, A, B, and C are the
this figure) according to their spatial proximity. The colors of first three vertices to visit, hence the spins at A4, A5, A6, B4,
the dots in Fig. 5 represent the results of the clustering. Dots B5, B6, C4, C5, and C6 are not generated. It also implies that
with the same color are in the same cluster. In problem V1 , D, E, and F will not be visited as the first three vertices, so
clusters are indicated by using dotted ellipses. After clustering, D1, E1, F1, D2, E2, F2, D3, E3, and F3 are not generated.
we compute the centroids (represented by open circles) of the The mapping shown in Fig. 6 can be compared with the
vertices of each cluster. Here, we define a new TSP V1 whose conventional mapping shown in Fig. 4. Even in this very small
vertices are the centroids of the clusters in V0 . The new TSP, example, the number of spins has been reduced from 36 to 14,
V1 , serves as a coarsened approximation of the original TSP, and the number of interactions has been reduced from 360 to
V0 . In Fig. 5, the vertices of V1 are again clustered, and the 47. We see that clustering substantially reduces the complexity
centroids of the clusters are then calculated to form a further of the calculations.
coarsened TSP, V2 . This coarsening process is recursively The spin reductions of the proposed method can be analyzed
repeated until the number of nodes becomes sufficiently small as follows. When dividing n vertices into K clusters, we
for easy solution by the Ising model. assume each cluster has n/K vertices on average and the
number of spins required for mapping the problem on the Ising
C. Incremental annealing model then becomes n2 /K. Therefore, the required number
Once the number of vertices has become tractable, we then of spins is reduced to 1/K of the existing method when we
begin to incrementally solve the newly defined TSPs. The coarsen the vertex by one level of clustering. Hierarchical
annealing starts from the most-coarsened problem, Vp . In the clustering of m levels ideally reduces the number of spins
example of Fig. 5, the most-coarsened TSP is V2 , in which exponentially to 1/K m .
the clustering was performed recursively for p = 2 times. The
mapping of the most-coarsened problem is carried out in the D. Random flip for TSP
same way as in the conventional method, and was explained in We also propose a modified random flip method that can be
the previous section. Because the number of vertices is np , the applied during the proposed TSP optimization. The procedure
number of spins is n2p . However, after hierarchical clustering, is shown in Alg. 3. In the proposed method, a randomly
np has become small enough and the optimal solution can selected spin σik is forced to be 1, and the other spins in
easily be obtained. the same column and in the same row are forced to be −1.

Authorized licensed use limited to: Shanghai Jiaotong University. Downloaded on February 26,2023 at 09:10:24 UTC from IEEE Xplore. Restrictions apply.
 → Vertex
A B C D E F 50
↓ Order
1 Optimal solution
② K1 = 1 (Conventional)
E 2 40 K1 = 24,K2 = 12,K3 = 6

Frequency
D ①
C 3
30
4

F A 5
20
B

③
6
10
(a) (b)
Fig. 6. Mapping on the Ising model using the proposed method. 0
500 1000 1500 2000 2500 3000
→ Vertex
Total distance
↓ Order
A B C D A B C D Fig. 8. Solution distribution (eil101).
1 1

2 2 • Clustering method: k-means++

3 3
• Number of trials: 100 times for each setting
Here, ∆x = xmax − xmin , and xmax and xmin are the
4 4
maximum and minimum x-coordinates at a given vertex group,
(a) (b) respectively. ymax and ymin are defined similarly.
Fig. 7. Proposed random flip for TSP. Table I summarizes the simulation results. The columns
Km show the number of clusters in the m-th clustering. The
columns, Best and Ave, show the best and average solutions
This forces the tour to visit vertex k as the i-th vertex and for 100 trials, respectively. The numbers in the parentheses
the constraints for that visit are all satisfied. Compared to are the ratios of the best and average solutions to the known
the conventional random flip, the spin states are considered optimal value in TSPLIB. The columns tc and ta are the aver-
to be close to the allowed solutions. A simple example of the age times taken for clustering and for solving the original and
proposed random flip is illustrated in Fig. 7. In Fig. 7(a), the generated TSPs via the Ising model based solver. The column
spin at 2C is selected for a random flip and set to 1 (upward #Spins is the average number of spins used to solve the given
arrow). As shown in Fig. 7(b), all other spins in row-2 and benchmark problem. K1 = 1 corresponds to the conventional
column-C, except for the spin at 2C, are set to −1. method, wherein V0 is directly solved. In this setup, since the
hyperparameters B and C are given sufficiently large values,
Algorithm 3 Random flip for TSP the constraints were satisfied in all trials.
1: Select NRF spins to flip(σi1 k1 , σi2 k2 , · · · , σiN kN ) Compared to the conventional method, the best and the aver-
RF RF
2: for r = 1 to NRF do age solutions were both improved for all of the problems. The
3: σir kr ← 1 quality of the average solution with the proposed method was
4: for j = 1 to n (where j ̸= ir ) do improved by up to 67.1% (in eil101). As expected, with larger
5: σjkr ← −1 problem sizes, greater improvements are achieved. In addition,
6: end for the number of spins, as well as the calculation times, were suc-
7: for l = 1 to n (where l ̸= kr ) do cessfully reduced owing to the clustering. In the conventional
8: σir l ← −1 method, the annealing time was the dominant component of
9: end for the total runtime, which was significantly reduced by the
10: end for proposed method. Comparing the total runtime for the problem
with 101 vertices, the proposed method obtained a solution that
was 73.78x faster than the conventional method. Because the
IV. N UMERICAL EVALUATION proposed method does not alter the solver hardware, further
We implemented the proposed hierarchical Ising model accelerations are expected if executed using hardware solvers.
solver in C++ to evaluate its performance. The following Fig. 8 shows the solution distributions obtained by the
problems from the TSP benchmark TSPLIB [19] were solved: conventional and the proposed methods (K1 = 24, K2 =
burma14, ulysses16, ulysses22, bays29, berlin52, eil76, and 12, K1 = 6) for eil101. The distribution of the proposed
eil101. The numbers embedded in the problem names indicate method is very close to the best known value in TSPLIB,
the number of vertices. The solver was run as a single-threaded while that of the conventional method is distributed over much
program on an Intel Core i9-7980XE [email protected] GHz with larger values. Even for the problem with 101 vertices, the ratio
128 GB of memory with the following settings. is 1.22 which significantly better than the ratio of 3.37 from
• Initial temperature: 1/n the conventional method.
• Cooling rate: α=0.999 Fig. 9 shows the correlation between K1 and the average
• Update coefficient: R = 0.3 √ number of spins when eil101 is mapped. When K1 = 24, the
• Hyper parameters: A = 1, B = C = n ∆x2 + ∆y 2 number of spins was reduced by 94.8%. In the conventional
• Clustering depth: m=1, 3 method, there are a large number of interactions connected

Authorized licensed use limited to: Shanghai Jiaotong University. Downloaded on February 26,2023 at 09:10:24 UTC from IEEE Xplore. Restrictions apply.
 TABLE I
O PTIMIZATION RESULTS ( BOLD FONT INDICATES BETTER RESULT. K1 = 1 CORRESPONDS TO THE CONVENTIONAL METHOD .)

Problem K1 K2 K3 Best Ave. tc [ms] ta [s] tt [s] #Spins

burma14 1 - - 31.83 (1.03) 37.01 (1.20) - 1.179 1.184 196
burma14 4 - - 30.88 (1.00) 32.51 (1.05) 1.321 0.255 0.262 63.82
ulysses16 1 - - 75.03 (1.01) 88.30 (1.19) - 1.798 1.807 256
ulysses16 4 - - 73.99 (1.00) 79.76 (1.08) 1.306 0.455 0.464 97.98
ulysses22 1 - - 89.90 (1.19) 108.66 (1.44) - 5.019 5.035 484
ulysses22 6 - - 75.67 (1.00) 78.94 (1.04) 1.534 0.723 0.741 117.22
bays29 1 - - 13977.50 (1.50) 16181.28 (1.74) - 12.623 12.652 841
bays29 6 - - 9155.45 (1.00) 9878.29 (1.06) 1.628 1.203 1.222 154.44
berlin52 1 - - 16852.80 (2.23) 19416.83 (2.57) - 114.520 114.631 2704
berlin52 24 12 6 8046.29 (1.07) 9137.48 (1.21) 7.229 3.724 3.830 188.12
eil76 1 - - 1532.65 (2.81) 1694.31 (3.11) - 443.122 443.456 5776
eil76 24 12 6 590.21 (1.08) 642.21 (1.18) 7.305 8.257 8.366 285.16
eil101 1 - - 2125.16 (3.33) 2391.28 (3.37) - 1226.331 1227.032 10201
eil101 24 12 6 717.54 (1.12) 786.47 (1.22) 7.389 16.522 16.636 531.22

[2] G. A. Kochenberger, J.-K. Hao, Z. Lü, H. Wang, and F. Glover, “Solving

104
The number of spins

large scale Max Cut problems via tabu search,” J. Heuristics, vol. 19,
no. 4, pp. 565–571, 2011.
[3] K.-L. Du and M. N. S. Swamy, Search and optimization by metaheuris-
tics. Springer International Publishing, 2016.
[4] M. Yamaoka, C. Yoshimura, M. Hayashi, T. Okuyama, H. Aoki, and
H. Mizuno, “20k-spin Ising chip for combinational optimization problem
with CMOS annealing,” in Dig. Tech. Papers, IEEE Intl. Solid-State
Circuits Conf., Feb 2015, pp. 1–3.
103 [5] P. L. McMahon, A. Marandi, Y. Haribara, R. Hamerly, C. Langrock,
S. Tamate, T. Inagaki, H. Takesue, S. Utsunomiya, K. Aihara et al.,
“A fully programmable 100-spin coherent Ising machine with all-to-all
0 6 12 18 24 connections,” Science, vol. 354, no. 6312, pp. 614–617, 2016.
K1 [6] H. Gyoten, H. Hiromoto, and T. Sato, “Enhancing the solution quality of
hardware Ising-model solver via parallel tempering,” Proc. IEEE/ACM
Fig. 9. The number of required spins (eil101). Intl. Conf. on Computer-Aided Design, pp. 70:1–70:8, 2018.
[7] H. Nishimori, Statistical physics of spin glasses and information pro-
cessing. Oxford University Press, 2001.
[8] T. Kadowaki and H. Nishimori, “Quantum annealing in the transverse
to each spin, and hence it is necessary to update the spins Ising model,” Physical Review E, vol. 58, no. 5, pp. 5355–5363, 1998.
in series. Otherwise, the quality of the solution degrades due [9] M. Yamaoka, C. Yoshimura, M. Hayashi, T. Okuyama, H. Aoki, and
H. Mizuno, “A 20k-spin Ising chip to solve combinatorial optimization
to the so-called bang-bang effect. We therefore updated the problems with CMOS annealing,” IEEE J. Solid-State Circuits, vol. 51,
spins serially. By contrast, in the proposed clustering based no. 1, pp. 303–309, 2016.
approach, the spins that belong to non-adjacent clusters were [10] C. Cook, H. Zhao, T. Sato, M. Hiromoto, and S. X.-D. Tan, “GPU-
based ising computing for solving max-cut combinatorial optimization
not connected and could thus be updated simultaneously. problems,” Integration, vol. 69, pp. 335 – 344, 2019. [Online]. Available:
In the conventional method, the number of spin updates is http://www.sciencedirect.com/science/article/pii/S0167926019301348
proportional to n2 . Meanwhile, assuming that n vertices are [11] M. Erol and F. Bulut, “Real-time application of travelling salesman prob-
lem using google maps api,” Electric Electronics, Computer Science,
equally divided into K clusters, the proposed method can Biomedical Engineerings’ Meeting, pp. 1–5, 2017.
update spins in non-adjacent clusters in parallel, so the number [12] R. Takahashi, “Quantitative evaluation of iterative extended changing
of updates is proportional to 2×(n/K)2 , which leads to K 2 /2 crossover operators to solve the traveling salesman problem: Diversity
measurement and its application to selection strategies in genetic algo-
times faster spin updates than when using naive methods. rithms,” Intl. Conf. on Natural Computation, pp. 235–244, 2014.
[13] M. Dorigo and L. M. Gambardella, “Ant colony system: a cooperative
V. C ONCLUSION learning approach to the traveling salesman problem,” IEEE Transac-
In this paper, we proposed a hierarchical clustering approach tions on evolutionary computation, vol. 1, no. 1, pp. 53–66, 1997.
[14] H. Gyoten, M. Hiromoto, and T. Sato, “Area efficient annealing proces-
for solving TSPs using the Ising model-based solver. Through sor for ising model without random number generator,” IEICE Trans.
simulations, benchmark problems were solved by the existing Information and Systems, vol. E101-D, no. 2, pp. 314–323, 2018.
method and by the proposed method using the Ising model- [15] A. Zaribafiyan, D. Marchand, and S. Rezaei, “Systematic and determin-
istic graph minor embedding for cartesian products of graphs,” Quantum
based solver and were compared. The solution quality of the Information Processing, vol. 16, no. 5, p. 136, 2017.
proposed method has been improved by up to 67.1% while [16] E. Boros, P. Hammer, and G. Tavares, “Local search heuristics for
the runtimes have become 73.8x shorter than when using the quadratic unconstrained binary optimization (QUBO),” J. Heuristics,
vol. 13, no. 2, pp. 99–132, 2007.
conventional method. [17] S. Kanamaru, D. Oku, M. Tawada, S. Tanaka, M. Hayashi, M. Yamaoka,
M. Yanagisawa, and N. Togawa, “Efficient Ising model mapping to
ACKNOLEGEMENT solving slot placement problem,” IEEE Itnl. Conf. Consumer Electronics,
This work was partially supported by JSPS KAKENHI pp. 1–6, 2019.
[18] J. Cai, W. G. Macready, and A. Roy, “A practical heuristic for finding
Grant No. 17H01713. graph minors,” 2014.
[19] “TSPLIB,” softlib.rice.edu/pub/tsplib/tsp/.
R EFERENCES
[1] S. Kirkpatrick, C. D. Gelatt, and M. P. Vecchi, “Optimization by
simulated annealing,” Science, vol. 220, no. 4598, pp. 671–680, 1983.

Authorized licensed use limited to: Shanghai Jiaotong University. Downloaded on February 26,2023 at 09:10:24 UTC from IEEE Xplore. Restrictions apply.