Abstract
In this paper, we first establish a Multi-Relation Granular Computing model by a given graph, and point out that all the reducts of the constructed Multi-Relation Granular Computing model are exactly all the minimal vertex covers of the corresponding graph. Thus, the vertex cover problem in graph theory can be converted to the knowledge reduction problem in rough set theory. Based on the conversion, we then introduce methods for dealing with the knowledge reduction problem to solve the vertex cover problem. In particular, we introduce a kind of method called the heuristic reduction algorithm based on entropy.
Introduction
Rough set theory, proposed by Pawlak [18, 19], is a useful mathematical tool for dealing uncertain, inexact and fuzzy knowledge. It has been widely used in data analysis, data mining and artificial intelligence [2, 33]. The original idea of rough set theory is based on the classification on the universe by an indiscernibility binary relation. Each object subset is approximated by a pair of definable sets called the lower and the upper approximations. The lower approximation is the greatest definable set contained in the given subset, while the upper approximation is the smallest definable set containing the given subset. Generally speaking, the upper approximation is larger than the lower approximation. If the lower approximation and the upper approximation of one object subset are just the same, then the subset is called a definable set; otherwise, called a rough set. Rough set-based data analysis starts from a data table called an information system. In an information system, there may be redundant knowledge and their removal can decrease the search space and improve some criteria. Thus, knowledge reduction problem(KRP) is essential in information systems analysis, which aims to compute a subset of relations (i.e., a reduct) while preserving or improving the classification ability of the total relations. Many approaches have been proposed for this issue, among those the one based on boolean logical operations can obtain all the reducts precisely[6].
In graph theory, there also exists an interesting and valuable topic called the vertex cover problem(VCP) [1, 8], which has been widely used in reality life [23, 37], e.g, the network measurement problem [31] and the site layout problem [23, 37]. The purpose of VCP is to select a minimum subset of vertices which can cover all the edges of a graph, i.e., each edge should be adjacent with some selected vertex. Similar to the situation of KRP, the boolean logical operations can also be employed to compute all the solutions or an optimal solution of the VCP. However, finding an optimal solution for a VCP has been proven to be an NP-hard problem. Therefore, many existing algorithms has been proposed to approximate the optimal solution from different viewpoints [1, 7].
We can see that the theoretical foundation for dealing with both the KRP and VCP is the boolean logical operation. Inspired from this idea, this paper devotes to explore this connection and realizes the conversion between them. Our main works are given as follows. With respect to a given graph G = (V, E), we induce a Multi-Relation Granular Computing (GrC) model S G = (U, A), where U and A are labeled as the object set and the relation set. The relation set A of the induced Multi-Relation GrC model is exactly the vertex set V of the graph G. We conclude that a vertex subset is a minimal vertex cover of the graph G if and only if it is a reduct of the induced Multi-Relation GrC model S G . By the above method, we can convert the VCP to the KRP and use the existing efficient methods for the KRP to solve the VCP. For instance, the positive-region reduction method which keeps the positive region unchanged [29], the heuristic entropy reduction method in which the relations are selected one by one by a heuristic function [12], and the combination entropy reduction method in which the positive regions are combined with the entropy [22]. In this paper, we introduce the heuristic entropy reduction algorithm to deal with the VCP.
The rest of this paper is as follows. In Section 2, we give the frameworks of graph theory related with the VCP and rough set theory related with the KRP. In Section 3, we explore the connection between the VCP and the KRP through boolean logical operations. We construct a Multi-Relation GrC model induced by a graph whose all minimal vertex covers are exactly all the reducts of the induced Multi-Relation GrC model. In Section 4, we introduce a relatively efficient algorithm called heuristic algorithm based on entropy in rough set theory to compute a minimal vertex cover of a graph.
Preliminary knowledge in graph theory and rough set theory
In this section, we will review some basic concepts related to boolean logical operations, the vertex cover problem in graph theory, and the knowledge reduction problem in rough set theory.
Boolean logical operations
Boolean logical operations satisfy the following operation laws: a ∨ a = a, a ∧ a = a; a ∨ b = b ∨ a, a ∧ b = b ∧ a; (a ∨ b) ∧ c = (a ∧ c) ∨ (b ∧ c), (a ∧ b) ∨ c = (a ∨ c) ∧ (b ∨ c); a ∧ (a ∨ b) = a, a ∨ (a ∧ b) = a.
A boolean expression that only contains conjunction (respectively, disjunction) operations is called a conjunction (respectively, disjunction) normal form. A conjunction (respectively, disjunction) normal form with the minimal number of variables in the boolean expression is called a minimal conjunction (respectively, disjunction) normal form.
The vertex cover problem in graph theory
In a graph G = (V, E), where V is the vertex set and E is the edge set of G, we say a vertex is one end of an edge if the vertex is adjacent with the edge. For two vertices u, v, if an edge e ∈ E is adjacent with u and v, u, v are called the ends of e, denoted as V (e) = {u, v}. A vertex covers an edge if the vertex is one end of the edge. A vertex subset is a vertex cover of a graph if the subset covers all the edges of the graph. A minimal vertex cover is a vertex cover such that any of its proper subsets is no longer a vertex cover. The vertex cover with the least vertex cardinal is the minimum vertex cover. Obviously, a minimum vertex cover is minimal, but not vice versa [3].
Let G = (V, E) be a graph. The boolean function of G is defined as follows:
Theorem 1 indicates that one can obtain all the minimal vertex covers of a given graph just by using boolean logical operations to convert the boolean function F (G) to a minimal disjunction normal form which contains some minimal conjunction normal forms as variables. Each minimal conjunction normal form is just a minimal vertex cover of G. The following example illustrates the operation process.
The boolean function of G1 is F (G1) = (v1 ∨ v2) ∧ (v2 ∨ v3) ∧ (v2 ∨ v4) .
Using boolean logical operations to convert F (G1) to a minimal disjunction normal form is given as follows:
F (G1) = ((v1 ∧ v2) ∨ (v1 ∧ v3) ∨ v2 ∨ (v2 ∧ v3))∧ (v2 ∨ v4) = (v2 ∨ (v1 ∧ v3)) ∧ (v2 ∨ v4) = v2 ∨ (v2∧v4) ∨ (v1 ∧ v2 ∧ v3) ∨ (v1 ∧ v3 ∧ v4) = v2 ∨ (v1 ∧ v3 ∧ v4).
By Theorem 1, we can obtain all the minimal vertex covers: {v2} , {v1, v3, v4}. The minimum vertex cover is {v2}.
Rough set theory is one of the most successful tools for handling uncertain information. Information on rough set theory is manifested in the form of information table, or information system. An information system contains data about objects of interest that are characterized by a finite set of attributes, and is usually represented by a triple S = (U, A, {V
a
| a ∈ A}) [18, 19], where: U is a nonempty finite set of objects called the domain set; A is a nonempty finite set of attributes such that a : U → V
a
for any a ∈ A, i.e., a (x) ∈ V
a
, x ∈ U, where V
a
is called the domain of attribute a.
An information system S = (U, A, {V a | a ∈ A}) is usually written as S = (U, A) simplicity.
Each nonempty subset B ⊆ A in an information system (U, A) determines an indiscernibility relation as follows:
A partition U/R B is also called a knowledge, and an information table can be seen as an defined equivalence relation cluster, i.e. knowledge base [28].
With the theoretical development and the widening application regions of rough set theory, relations in information systems have no longer limited to equivalence relations but more general situations [10, 30]. From the viewpoint of granular computing, Lin and Yao studied a granular computing model under binary relations and proposed a framework for set approximation in the granular computing model [14– 17, 35]. In the following, we present the notion of granular computing model.
For a binary relation a ∈ A and an object x ∈ U, N
a
(x) = {y ∈ U | (x, y) ∈ a} is called the neighborhood of x under a; N
A
(x) = ∩ a∈AN
a
(x) is called the neighborhood of x under A; Pair (U, A) is called a Multi-Relation Granular Computing (GrC) model.
Attribute reduction also called knowledge reduction is one of the basic contents in rough set theory. In a Multi-Relation GrC model, there are always redundant binary relations and removing them has no influence on the representation of concepts. From this point of view, in the following, we introduce the notion of knowledge reduction in a Multi-Relation GrC model.
We can see that a reduct of a Multi-Relation GrC model is a minimal subset of relations which remains the original neighborhood of each object. In other words, a reduct keeps the neighborhood discernibility between each pair of objects. In the following, similar to information systems based on equivalence relation, we define the discernibility set of a pair of objects in Multi-Relation GrC models as same as the discernibility set in information systems.
d (x
i
, x
j
) = {a ∈ A | x
j
∉ N
a
(x
i
)} is called the discernibility set of (x
i
, x
j
); is called the discernibility matrix of the Multi-Relation GrC model (U, A).
It should be noted that in a discernibility matrix, the discernibility relation between each pair of objects is determined explicitly. Therefore, a Multi-Relation GrC model and its discernibility matrix correspond one to one.
The reducts and consistent sets can be characterized by the discernibility matrix . In the following, a reduction method based on discernibility matrix is proposed to compute all the reducts of a Multi-Relation GrC model.
⇐Suppose B∩ d (x i , x j ) ≠ ∅ for any d (x i , x j )≠ ∅. Then a′ ∈ B satisfying x j ∉ Na′ (x i ), which follows x j ∉ N B (x i ). Moreover, d (x i , x j )≠ ∅ equals to x j ∉ N A (x i ). Thus, we have if x j ∉ N A (x i ), then x j ∉ N B (x i ). That means N B (x i ) ⊆ N A (x i ). On the other hand, since B ⊆ A, it is obvious that N A (x i ) ⊆ N B (x i ). Therefore, N B (x i ) = N A (x i ), which means B is a consistent set. This completes the proof.
In information systems based on equivalence relation, an efficient theoretical reduction method is based on boolean logical operations. If d (x
i
, x
j
)≠ ∅, the partial pair (x
i
, x
j
) can be distinguished by any relation in d (x
i
, x
j
), thus through ∨d (x
i
, x
j
), x
i
and x
j
can be distinguished. The expression ∧ {∨ d (x
i
, x
j
)} is the conjunction of all ∨d (x
i
, x
j
) which implies that each partial pair (x
i
, x
j
) can be distinguished by it. These also apply to Multi-Relation GrC models. Thus, one can obtain all the reducts of a Multi-Relation GrC model by computing the discernibility function:
Based on the neighborhoods of all objects in Table 1, the discernibility matrix of the Multi-Relation GrC model can be uniquely determined. In the first row of Table 1, we have x1 ∈ N a 1 (x1) , N a 2 (x1) and N a 3 (x1), hence d11 =∅. Moreover, x2 ∈ N a 1 (x1), N a 2 (x1) and N a 3 (x1), so d12 =∅; x3 ∉ N a 1 (x1), thus d13 = {a1}; x4 ∉ N a 1 (x1) and N a 2 (x1), so d14 = {a1, a2}. Similarly, we can obtain the discernibility matrix of the Multi-Relation GrC model as shown in Table 2.
By Table 2, we can obtain the discernibility function:
Using boolean logical operations, we can obtain:
Hence the Multi-Relation GrC model has only one reduct: {a1, a2}.
From the above discussion, we know that all the reducts of a Multi-Relation GrC model can be obtained by computing the discernibility function and all the minimal vertex covers of a given graph can also be obtained by computing the boolean function. Assume that if the noempty discernibility sets in a Multi-Relation GrC model are exactly all the end sets of the edges in a graph, then, by using boolean logical operations, the discernibility function of the Multi-Relation GrC model and the boolean function of the graph can obtain the same result. i.e., all the reducts of the Multi-Relation GrC model are just all the minimal vertex covers of the graph. Base on it, in this section, we first induce a Multi-Relation GrC model by a graph whose all end sets of edges are exactly all the nonempty discernibility sets of the induced Multi-Relation GrC model and give an example to illustrate it. We then discuss the relationship between them.
In the following, we construct a Multi-Relation GrC model for a given graph.
Given a graph G = (V, E) with V = {a1, a2, . . . , a
m
} and E = {e1, e2, . . . , e
k
}. For unifying symbols, here we use a
i
, 1 ≤ i ≤ m instead of v
i
to represent the vertices, and for convenience, we use e ∈ E to represent its end set V (e) in the following. The induced Multi-Relation GrC model of the graph G can be constructed by the following steps: Label an object set U = {x1, x2, . . . , x|U|}; Label a relation set A = V = {a1, a2, . . . , a
m
}; Construct a discernibility matrix
where d (x
i
, x
j
) is denoted as the discernibility set of the pair (x
i
, x
j
) as shown in Table 3.
(4) We obtain a Multi-Relation GrC model S G = (U, A) induced by the given graph G, where U is the arbitrarily labeled object set, A is the relations set which is just the vertex set V, and d (x i , x j ) is the discernibility set of the objects pair (x i , x j ) ∈ U × U.
i.e. d (x
i
, x
j
) =
For the above construction, we have several illustrations as follows:
(a) In the above construction, Step (3) regards all e t (1 ≤ t ≤ k) as the nonempty discernibility sets of the object pairs. This is a key step which can determine the neighborhood of each object. ∅ is assigned to the diagonal entries d (x i , x i ) (1 ≤ i ≤ |U|) of the discernibility matrix and e1, . . . , e k are assigned to the undiagonal entries d (x i , x j ) (i ≠ j) one by one until the last element e k is assigned. Then ∅ is associated to the rest elements in .
(b) Each pair of objects is associated with one discernibility set, thus |U| objects need at least |U| × (|U|-1) discernibility sets. |U| needs to satisfy |U| × (|U|-1) ≥ k, which is equal to that . Here we take when 1 + 4k is complete square like 9,25,81,121,169,..., and otherwise .
(c) From the above discussion, a Multi-Relation GrC model can be uniquely determined by a given discernibility matrix. Thus, from step (3), we can obtain the induced Multi-Relation GrC model (U, A) in which the neighborhood of each object under each relation is N a i (x j ) = {x t ∈ U | a i ∉ d (x j , x t )}, 1 ≤ i ≤ m, 1 ≤ j ≤ |U|,1 ≤ t ≤ |U|.
In the following example, we illustrate how to construct the induced Multi-Relation GrC model of a given graph.
(1) Since k = |E|=5, , thus label an object set U = {x1, x2, x3};
(2) Label a relation set A = V = {a1, a2, a3, a4, a5, a6};
(3) Construct the discernibility matrix : ∅ is assigned to d (x1, x1), d (x2, x2) and d (x3, x3); e1 is assigned to d (x1, x2); e2 is assigned to (x1, x3); e3 is assigned to d (x2, x1); e4 is assigned to d (x2, x3) and e5 is assigned to d (x3, x1). Because e5 is the last one, ∅ is assigned to d (x3, x2). Consequently, we can obtain the discernibility matrix D as shown in Table 4.
(4) Because a Multi-Relation GrC model can be uniquely determined by a discernibility matrix, we now obtained the induced Multi-Relation GrC model S G 2 = (U, A) where U = {x1, x2, x3}, A = {a1, a2, a3, a4, a5, a6} and the discernibility matrix is as shown in Table 4.
By Table 4, we can obtain the discernibility function:
Using boolean logical operations to compute the reducts, we can obtain:
.
Hence, {a1, a3, a4, a5}, {a2, a5}, {a1, a3, a4, a6} and {a2, a4, a6} are four reducts of the induced Multi-Relation GrC model (U, A) and {a2, a5} is the unique minimum reduct.
The relationship between the graph G = (V, E) and its induced Multi-Relation GrC model S G = (U, A) is described as follows.
Theorem 3 implies that the SCP in graph theory and the KRP in rough set theory can be converted to each other.
Thus, in Example 3, we can conclude that {a1, a3, a4, a5}, {a2, a5}, {a1, a3, a4, a6} and {a2, a4, a6} are four minimal vertex covers and {a2, a5} is the unique minimum vertex cover of the graph G2.
A heuristic minimum vertex cover algorithm based on entropy for the VCP
In this paper, based on the conversion, we aim to use the existing methods for dealing with the KRP in rough set theory to deal with the VCP in graph theory. Such as the discernibility matrix based reduction method, the positive-region based reduction method, the information entropy based heuristic reduction method which can gradually reduce the time required by eliminating either unlikely possibilities or irrelevant states, or by selecting the required relations gradually [10, 20]. Entropy was firstly introduced into reduction method by Skowron [24], then Slezak used it to deal with the reduction problem in information system and Wang et al. introduced the conditional entropy into decision system [30], Liang et al. defined a kind of entropy and applied it to compute a reduct [12], Qian et al. developed a common accelerator into entropy reduction with which the data needed computing is gradually reduced in the process of the algorithm [22]. To measure the uncertainty of knowledge in Multi-Relation GrC model, Hu et al. proposed a kind of neighborhood entropy [9] as follows.
The basic effect of the entropy is to compare two classes of knowledge and to measure the uncertainty degree of one class of knowledge with respect to another.
In a Multi-Relation GrC model, we have the following descriptions for the relations.
a is a core iff for any reduct R ∈ RED, a ∈ R, i.e. a ∈ ∩ RED; a is unnecessary iff for any reduct R ∈ RED, a ∉ R, i.e. a ∈ (A - ∪ RED); a is relative necessary iff a belongs only part of the reducts, i.e. a ∈ (∪ RED - ∩ RED).
For a relation a ∈ B ⊆ A, we say a is dispensable in B iff for any x ∈ U, NB-{a} (x) = N B (x); Otherwise, a is indispensable.
a is a core iff a is indispensable in A, a is unnecessary iff a is dispensable in B for all consistent sets B that contains a, a is relative necessary iff a is dispensable in A, and there exists a consistent set B which contains a such that a is indispensable in B.
⇐ If a is not a core, we have that there exists a reduct B satisfying a ∉ B. Thus B ⊂ A - {a} ⊂ A and N A (x) ⊆ NA-{a} (x) ⊆ N B (x) for all x ∈ U. Note that N B (x) = N A (x), therefore NA-{a} (x) = N A (x), a is dispensable in A, and this is a contradiction with.
(2) ⇒ In fact for any consistent set a ∈ B, there exists a reduct C ⊆ B and if a is unnecessary, it follows a ∉ C, then C ⊂ B - {a} ⊂ B. Since N C (x) = N B (x) = N A (x) for any x ∈ U, it follows that N A (x) = N B (x) ⊆ NB-{a} (x) ⊆ N C (x) = N A (x), thus NB-{a} (x) = N B (x). So a is dispensable in B.
⇐ For any reduct C, we can see that B = C ∪ {a} is a consistent set. If a is dispensable in B, then NB-{a} (x) = N B (x) = N A (x), that is NC-{a} (x) = N B (x) = N A (x) for any x ∈ U. Therefore C - {a} is consistent, moreover C is a reduct, thus C - {a} = C, that is a ∉ C, so a is unnecessary.
(3) We may conclude it from (1) and (2).
The followings characterize the relations from the view point of neighborhood entropy.
⇒ By Definition 6,
. Since , if NH (A | B) =0, then for all x ∈ U, it follows , besides N B (x) ⊇ N A (x), so N B (x) = N A (x) for any x ∈ U.
⇐ If N B (x) = N A (x) for all x ∈ U, then . By Definition 6, NH (A | B) =0.
B is a consistent set iff NH (A | B) =0; B is a reduct iff NH (A | B) =0 and NH (A | (B - {a})) ≠0 for all a ∈ B.
(2) It holds from (1).
ais dispensable inBiffNH (B | (B - {a})) =0. ais indispensable inB iff NH (B | (B - {a})) ≠0,
Therefore,
NH (B | (B - {a})) =
⇐ If
NH (B | (B - {a})) =
, then for all x ∈ U, it follows NB-{a} (x) ⊆ N a (x) and with the fact of NB-{a} (x) ⊆ N B (x), then NB-{a} (x) = N B (x), thus a is dispensable.
(2) It is obvious from (1).
ais a core iffNH (A | (A - {a})) ≠0, ais unnecessary iffNH (A | B) =0 ⇒ NH (A | (B - {a})) =0 for any B ⊆ A, ais relative necessary iffNH (A | (A - {a})) =0 and there exists someB ⊆ Athat containsasuch thatNH (A | B) =0 and NH (A | (B - {a})) ≠0.
(2) From Theorem 5.(1), NH (A | B) =0 ⇔ B is a consistent set, NH (A | (B - {a})) =0 ⇔ (B - {a}) is a consistent set. The sufficiency condition means that if B is a consistent set, then (B - {a}) is also consistent, witch implies a is unnecessary.
(3) It holds from (1) and (2).
For a Multi-Relation GrC model S = (U, A) and two relation sets B, C ⊆ A, we say that B is finer than C if any x ∈ U satisfies N B (x) ⊆ N C (x), denoted by B ⪯ C.
Corollary 3 means that for a fixed relation set D, when we add another relation set B or C to D, respectively, if B ∪ D are finer than C ∪ D, then NH (B | D) ≥ NH (C | D). Hence, we could use the neighborhood entropy to measure the uncertainty degree of one set of relations to another.
We give a heuristic reduction algorithm based on neighborhood entropy as followings. In the algorithm, the relation is selected one by one according to its uncertainty degree which is measured by the neighborhood entropy.
1: Let red =∅;
2: Compute NH (A | (A - {a})), if NH (A | (A - {a})) ≠0, then let red = red ∪ {a} for all a ∈ A;
3: If red =∅, then {
31: Compute NH ({a}), for all a ∈ A;
32: Let NH ({a0}) = max {NH ({a}) | a ∈ A};
33: Let red = {a0},
}
4: While (NH (A | red) ≠0){
41: Compute
NH ({a} | red) for all a ∈ (A - red);
42: Let
NH ({a0} | red) = max {NH ({a} | red),
for all a ∈ (A - red)};
43: Let red = red ∪ {a0},
}
5: Return red.
From Theorem 7.(1), NH (A | (A - {a}))≠0 ⇔ a is a core. So step 2 picks all cores. From Corollary 3, if {a} ∪ red ⪯ {b} ∪ red, then NH ({a} | red) ≥ NH ({b} | red). We can say that step 4 picks the relation which maximizes NH ({a} | red) in the remaining relation set (A - red).
1: Let red =∅;
2: Compute and obtain NH (A | (A - {a})) =0, for all a ∈ A, thus red =∅;
3: Since red =∅, then{
31: Compute NH ({a}), for all a ∈ A;
32: Obtain
NH ({a2}) = max {NH ({a}) | a ∈ A};
33: Let red = {a2};
}
4:(1)Since NH (A | {a2}) ≠0, then {
41: compute
NH ({a} | {a2}) for all a ∈ (A - {a2});
42: Obtain that
NH ({a5} | {a2}) = max {NH ({a} | {a2})
for all a ∈ (A - {a2})};
43: Let red = {a2} ∪ {a5};
}
(2)Since NH (A | {a2, a5}) =0, step 4 ends;
5. Return a reduct {a2, a5}, the algorithm ends.
The result {a2, a5} obtained by Algorithm 1 is one of the reducts obtained in Example 3. Thus, we can see that Algorithm 1 is feasible to compute a reduct in a Multi-Relation GrC model.
Theorem 3 means a vertex subset of a graph is a minimal vertex cover if and only if it is a reduct of the induced Multi-Relation GrC model by the graph. So, {a2, a5} is also a minimal vertex cover of the graph G2 in Example 3. Thus, Algorithm 1 can also be used to compute a minimal vertex cover of a graph just needing to convert the graph to its induced Multi-Relation GrC model firstly.
For convenience, here we summarize the processes by using the proposed rough set method to compute a minimal vertex cover of a graph in Algorithm 2.
Input: A graph
Output: A minimal vertex cover
step 1: Construct the induced Multi-Relation GrC model respected to the graph;
step 2: Compute a reduct of the induced Multi-Relation GrC model by Algorithm 1;
step 3: Return the reduct.
Given a graph G = (V, E) with V = {a1, a2, . . . , a m } and E = {e1, e2, . . . , e k }, the time complexity of step 1 and step 2 are and , respectively, so the time complexity of Algorithm 2 is .
Conclusions
In this study, we have converted the vertex cover problem in graph theory to the knowledge reduction problem in rough set theory and introduce a reduction algorithm to deal with the vertex cover problem. Thus, we have connected rough set theory with classical optimization theory and used the rough set-based methods to deal with the optimization problems. Moreover, there are many kinds of reduction algorithms in rough set theory. Based on the conversion, all these reduction algorithms can be used to deal with the vertex cover problem.
As we know, the vertex cover problem in graph theory has wide applications. Based on the conversion, the application of rough set theory can be expanded in some degree. There are many different types of optimization problems. Connecting them with data mining issues and realizing mutual utilization of methods from different fields are parts of our further works.
Footnotes
Acknowledgments
The authors thank the anonymous referees for their valuable comments and suggestions.
This work is supported by grants from the National Natural Science Foundation of China (Nos. 61379021, 61272021, 61303131, 61573321) and the Department of Education of Fujian Province (Nos. JA13202, JK2013027, JK2014028).
This work is also supported by Research start-up fee of Zhejiang Ocean University.
