CN106685686A - Network topology estimation method based on simulated annealing - Google Patents
Network topology estimation method based on simulated annealing Download PDFInfo
- Publication number
- CN106685686A CN106685686A CN201610817914.6A CN201610817914A CN106685686A CN 106685686 A CN106685686 A CN 106685686A CN 201610817914 A CN201610817914 A CN 201610817914A CN 106685686 A CN106685686 A CN 106685686A
- Authority
- CN
- China
- Prior art keywords
- node
- topology
- tree
- network
- feature
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
- H04L41/12—Discovery or management of network topologies
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
- H04L41/14—Network analysis or design
Landscapes
- Engineering & Computer Science (AREA)
- Computer Networks & Wireless Communication (AREA)
- Signal Processing (AREA)
- Data Exchanges In Wide-Area Networks (AREA)
Abstract
The invention discloses a network topology estimation method based on simulated annealing. The method comprises the following steps: S1) carrying out network tomography imaging detection on a network in a known network topology; S2) processing a detection result through a wavelet packet decomposition algorithm to obtain clustering features; S3) carrying out feature selection through a simulated annealing algorithm; and S4) serving the selected features as features used when carrying out estimation on all other unknown network topologies, and estimating network topology through an agglomerate hierarchical clustering algorithm. The clustering features are selected through a randomized optimization algorithm, that is, the simulated annealing algorithm, so that the features most beneficial to topology estimation can be selected, and the problem that in an existing network topology imaging topology detection method, since the clustering features are not selected before using the agglomerate hierarchical clustering algorithm to estimate topology, topological estimation error is increased is prevented.
Description
Technical field
The present invention be more particularly directed to a kind of Estimating topology of networks method based on simulated annealing.
Background technology
With developing rapidly for the network technologies such as internet communication, contacting also increasingly between the life of people and network
Closely, but while network brings many convenient to user, the scale of network is increasing, and the complexity of network is also more next
It is higher, great difficulty is brought to the guarantee of the service quality of network, the performance characteristic of network is increasingly subject to user and net
The concern of network supervision department.But the network measuring system majority that there is currently is carried out on the premise of known network topology
, and real network is typically changeable, if can not accurately speculate to topological structure, then also cannot be accurate
Network is supervised.For this, scholars propose the method measured to network by Estimating topology of networks.Network topology
Estimation has become the important component part of modern network management system, has in communication network scientific development very important
Status.
Estimating topology of networks is referred to by the way that some of network element is scanned for and captured, and is found between element
Mutual relation, then this relation is shown appropriate topological structure.The angle of the element-specific paid close attention to from us goes out
Send out, network topology can be divided into physical topology and the class of logical topology two.Physical topology is represented in network between each entity device
Annexation, and logical topology then describes how flow transmits in a network.Say that Estimating topology of networks refers to from this angle
Be that method is measured to network of concern and its logical topology is speculated.
Network tomography technology be it is a kind of for detecting internet in logical topology new technique, be tomography
The once cross-cutting application of technology.Network tomography technology be based on it is a kind of during end to end technology is to obtain network those
The information that can not be observed directly.It considers that the routing node of network internal to be detected can't return information to observer.It
Mutually send between some controllable nodes of network edge to be detected, receive detection bag, and can be through treating in transmitting procedure
Detection network, therefore result of detection can reflect network internal characteristic, can reversely deduce the internal structure of network to be detected.
At present network tomography is mainly made up of two parts:Part I is the collection of detection data, wherein mainly
Research how inside collection network related useful information;Part II is statistical inference, and it is mainly according to by collection
Data are finding the information and rule of network internal.
The content of the invention
It is an object of the invention to overcome the deficiencies in the prior art, there is provided this is random most using simulated annealing for one kind
Optimized algorithm is selected cluster feature, can be selected and be estimated most beneficial feature to topology, and topology is estimated caused by excluding
The Estimating topology of networks method based on simulated annealing of the feature of meter error increase.
The purpose of the present invention is achieved through the following technical solutions:Estimating topology of networks side based on simulated annealing
Method, comprises the following steps:
S1, the network to known network topological structure carry out network tomography detection;
S2, using wavelet packet decomposition algorithm process result of detection obtain cluster feature;
S3, carry out feature selection using simulated annealing;
The feature used when S4, estimation that the feature chosen is topological as other all unknown networks, using cohesion level
Clustering method estimates network topology.
Further, the step S2 concrete methods of realizing is:WAVELET PACKET DECOMPOSITION each time iteration all by last letter
Two parts of low frequency and high frequency number are subdivided into, wherein S is primary signal, and A represents approximate filtration, is considered as the low frequency of front first order signal
Part;D represents details filtration, is considered as the HFS of front first order signal;
Wherein y [n] is signal to be decomposed, and g [n] is low pass filter, and h [n] is high pass filter, meets h [n]=(- 1
)ng[1-n];G [n] correspond to scaling function ψ (t) of wavelet transformation, referred to as scaling vector;H [n] correspond to the little of wavelet transformation
Wave function
Calculated as g [n] using the scaling vector corresponding to db4 small echos:Signal is carried out respectively low pass, high pass filter
After ripple, carry out down-sampled, respectively obtain the WAVELET PACKET DECOMPOSITION coefficient of low frequency and high frequency:
In the same manner, to Y1,0[n] and Y1,1[n] carries out same treatment, 4 groups of WAVELET PACKET DECOMPOSITION coefficients of the second layer is obtained, with this
Analogize, obtain the WAVELET PACKET DECOMPOSITION coefficient of any number of plies;Its recurrence formula is:
Using primary signal X of the 4 layers of WAVELET PACKET DECOMPOSITION to each destination nodeiProcessed, 16 groups little are most obtained at last
Ripple bag decomposition coefficient, is designated as YI, j, wherein i=0,1 ..., N-1 is purpose node serial number, and j=0,1 ..., 15 is wavelet packet point
Solution coefficient number, XiWith YI, jIt is all vector.
Further, step S3 includes following sub-step:
Some groups are randomly choosed in S31, the 16 groups of WAVELET PACKET DECOMPOSITION coefficients obtained from S2 as the feature estimated first, will
Feature selection state vector F=(f1, f2..., f16) represent, wherein f1, f2..., f16∈ { 0,1 }, represents when value is 0
This feature is chosen in unselected this feature, 1 expression;
R is made to be results set, for storage feature selection result and the key-value pair of its error rate composition, R={ } when initial;
Initialization temperature T=T0=2m, m are purpose node number;
Iterationses upper limit L=20 under same temperature is set, iteration count l=0 is initialized, initialization new explanation does not receive
Number of times t=0, lower the temperature frequency n=0;
S32, sample is carried out according to F estimating based on the topology of Agglomerative Hierarchical Clustering, and calculated with tree edit distance and open up
The error rate e (F) of estimated result is flutterred, R=R ∪ { F are updated:e(F)};
S33, as l < L execution step S34, otherwise skip to step S39;
S34, change F in t-th feature selection state, keep the selection state of other features constant, obtain one group it is new
Solution F 'i=(f 'I, 1, f 'I, 2..., f 'I, 16), its satisfaction:
Make i take all over all values, obtain 16 groups of new explanations:F′1, F '2..., F '16;
S35, judge that new explanation whether there is in R, for the solution that is not present in R is performed and step S32 identical method,
Obtain e (F 'i), update R=R ∪ { F 'i:e(F′i), and find e*=min (e (F 'i))=e (F 'm) with corresponding F 'm;
S36, l=l+1 is made, calculate Δ e=e*- e (F), new explanation F ' is received according to metropolis criterions with Probability pm, p's
Calculation is as follows:
If receiving new explanation, F=F ' is madem, e (F)=e*, t=0 skips to step S33;Otherwise t=t+1, carries out S37;
S37, as l < L execution step S38, otherwise skip to step S39;
S38, for each feature in F changes existing selection state with 50% probability, obtain one group it is new
Solution F 'm, perform and step S32 identical method, obtain e*=e (F 'm), update R=R ∪ { F 'm:e(F′m), skip to step
S36;
S39, n=n+1 is made,L=0;As t < 10 and n < 10, step S33 is skipped to;Otherwise select most from R
Excellent solution F*As the result of feature selection, terminate algorithm.
Further, estimate that concrete methods of realizing is based on the topology of Agglomerative Hierarchical Clustering in step S32:
S321, each cluster the distance between sample is defined, obtain distance matrix;For each destination node i, if
The feature chosen isWherein 0≤k1< k2< ... < kn≤ 15 and k1, k2... kn∈N;Feature is spelled
It is connected in a new vector
For any two destination node i, j, its characteristic vector Y is soughtiWith YjBetween correlation coefficient:
WhereinWithY is represented respectivelyiWith YjMathematic expectaion be average:
Node i is defined, the correlation distance between j is:
dI, j=1- | ρI, j|
When two destination nodes i, when the characteristic vector dependency of j is stronger, the absolute value ρ of its correlation coefficientI, jJust closer to
1, then correlation distance d between the twoI, jJust closer to 0;
Obtain the correlation distance matrix between all purposes node:
dI, j=dJ, i, (i, j=1,2 ..., n) and d1,1=d2,2=...=dN, n=0;So matrix D is a diagonal
For 0 symmetrical matrix;
S322, hierarchical clustering is carried out, obtain topology, comprised the following steps:
Each sample is each classified as into a class when S3221, initialization, common n classes are designated as G1, G2..., Gn, make it distinguish
N leaf node of clustering tree is constituted, N is designated as1, N2..., Nn;Weight w is assigned for each node1, w2..., wn, make w1=w2
=...=wn=0, between class distance matrix is G=D;Any two class Gi, GjThe distance between be gI, j=dI, j, initialization cluster time
Number l=1;
S3222, in G find minimum range gS, t=mini≠j(gI, j), by corresponding two classes Gs, GtAggregate into one
New class Gn+l, construct new node Nn+lAs Ns, NtFather node, make its weight wn+l=gS, t;
S3223, calculate new class G using formulan+lTo the distance of other classes, two classes Gp, GqThe distance between computing formula
For:
S3224, wherein np, nqRespectively class Gp, GqIn number of samples;
S3225, renewal between class distance matrix G:By in G with Gs, GtRelated row and column is eliminated, and is represented in last addition
New class Gn+lRanks, corresponding value is distance of the new class to other classes;
S3226, renewal cluster number of times l=l+1;
S3227, repetition S3222~S3225, until only remaining class G2n-1。
Further, in cluster process, two classes are all merged into a new class to step S322 by each wheel, directly
To an only surplus class, and generate the tax power binary tree that a nodes are 2n-1;According on following compatible rule merging binary tree
Node:One thresholding t=0.02 is set, from node Nn+1Start to N2n-2Till, if a node NiWith its father node NjPower
Value meets
Then by NiWith NjMerge, will NiThe father node of child node be changed to NjAfter delete Ni, the tree for finally giving is used as this
The Network traffic model of secondary estimation.
Further, in step S32 with tree edit distance calculate topological estimated result error rate it is concrete
Implementation method is:
Define the edit operation of three kinds of trees:
Change node label:The label for defining the leaf node of tree-shaped network topology is each self-corresponding numbering, remaining section
Point label is sky;If node label is changed to into cost r (a → b)=2m of b by a, wherein m is purpose section during network detection
Point number, that is, the leaf node number set;
Deletion of node:Non-root node v in tree T is deleted, if its father node is v ', the father node of the child node of v is changed to
v′;If cost r (a → Λ)=1, a of deletion of node is the label of node to be deleted, Λ represents empty node;
Increase node:The father node of the part child node of v ' and is changed to v by newly-increased node v as the child node of node v ';
If cost r (Λ → a)=1, the Λ for increasing node represents empty node, a is the label of node to be increased;
If E is from tree T1To tree T2Editing process, it comprises several tree edit operation e1, e2..., en;Remember T1
It is converted into T2Totle drilling cost r (E)=r (e1)+r(e2)+…+r(en);Tree edit distance is min (r (E)), will T1It is converted into T2
Minimum total cost;
To try to achieve tree edit distance, problem is converted to and seeks two tree T1With T2Between maximum match subtree problem:If altogether
It is M, T with subtree1With T2Matching relationship be (M, T1, T2), belong to T1But it is N to be not belonging to the node set of M1, belong to T2But no
The node set for belonging to M is N2, T1The label of interior joint i is designated as l1(i), T2The label of interior joint i is designated as l2(i);Then have:
By dynamic programming algorithm, r ((M, T are obtained1, T2))Minima, i.e. tree edit distance.
The invention has the beneficial effects as follows:
1st, using simulated annealing, this stochastic optimization algorithm is selected cluster feature the present invention, can be selected
Go out and most beneficial feature is estimated to topology, in excluding existing network chromatography imaging topological detecting method, gathered using cohesion level
Class method is estimated not select cluster feature before topology, so as to the feature of caused topological estimation difference increase;
2. during feature selection is carried out using simulated annealing, present invention uses the side of two kinds of acquirement new explanations
Method, the first can quickly find the locally optimal solution in a certain scope, can jump out the scope and find in the range of other for second
Optimal solution, this method of the present invention can shorten the algorithm time, and can obtain more accurately estimated result.
Description of the drawings
Fig. 1 is the flow chart of the present invention;
Fig. 2 is WAVELET PACKET DECOMPOSITION process schematic;
Fig. 3 is WAVELET PACKET DECOMPOSITION calculation flow chart.
Specific embodiment
The present invention is bent to the Delay Variation of original input signal, i.e. source node to each destination node using WAVELET PACKET DECOMPOSITION
Line is processed, and obtains signal time-frequency characteristics on different frequency bands.By the WAVELET PACKET DECOMPOSITION coefficient on relatively more different paths
Dependency between vector, is obtained between different destination nodes and shares portraying for length, and then by Agglomerative Hierarchical Clustering side
Method obtains Network traffic model.
In principle, to propagation delay time on each time-frequency domain change grasp it is more, just can more clearly depict from
Similarity degree of the source node to the path of different destination nodes.But, it is actually not all among WAVELET PACKET DECOMPOSITION coefficient
Coefficient is all conducive to obtaining correct result, wherein there is noise data, the DC component of such as low frequency is subject to actual physics chain
The impact of road length, if not overlapping trees length proportion is higher, then will have a strong impact on the judgement to overlapping trees length.For
Partial Feature is eliminated as far as possible for topology estimates the negative effect that produces, the accuracy that topology is estimated is lifted as much as possible, need
Feature selection is done to these coefficients.On the basis of topology estimation accuracy is improve, the reduction of characteristic number is also caused subsequently
The reduction of amount of calculation.
In order to assess whether selected feature can exactly estimate topology, a known network topology is needed as ginseng
According to.Accurate topological structure can be obtained using methods such as traceroute.After obtaining the topology, the network is carried out
Tomography is detected, and the topology that tomography is obtained compares with the topology that other modes are obtained, and finding makes two topologys most
Similar characteristic set, the characteristic set as after in practical application, for the network of unknown topologies to be detected and opened up
Flutter estimation.
Technical scheme is further illustrated below in conjunction with the accompanying drawings.
As shown in figure 1, the Estimating topology of networks method based on simulated annealing, comprises the following steps:
S1, the network to known network topological structure carry out network tomography detection;
S2, under real network environment, network traffics have very high sudden, so can cause to have in time lag curve
More high frequency components occur, and in order to more meticulously obtain the time-frequency characteristics in high fdrequency components, the present invention utilizes wavelet packet
Decomposition algorithm processes result of detection and obtains cluster feature;Concrete methods of realizing is:As shown in Fig. 2 WAVELET PACKET DECOMPOSITION changes each time
Generation is two parts of low frequency and high frequency all by last signal subdivision, and iterationses are more, then frequency range divides thinner, wherein
S is primary signal, and A represents that approximately (approximation) is filtered, and is considered as the low frequency part of front first order signal;D represents details
(detail) filter, be considered as the HFS of front first order signal;
As shown in figure 3, wherein y [n] is signal to be decomposed, g [n] is low pass filter, h for the calculation process of WAVELET PACKET DECOMPOSITION
[n] is high pass filter, meets h [n]=(- 1)ng[1-n];G [n] correspond to scaling function ψ (t) of wavelet transformation, referred to as chi
Degree vector;H [n] correspond to the wavelet function of wavelet transformation
Calculated as g [n] using the scaling vector corresponding to db4 small echos:Signal is carried out respectively low pass, high pass filter
After ripple, carry out down-sampled, respectively obtain the WAVELET PACKET DECOMPOSITION coefficient of low frequency and high frequency:
In the same manner, to Y1,0[n] and Y1,1[n] carries out same treatment, 4 groups of WAVELET PACKET DECOMPOSITION coefficients of the second layer is obtained, with this
Analogize, obtain the WAVELET PACKET DECOMPOSITION coefficient of any number of plies;Its recurrence formula is:
Using primary signal X of the 4 layers of WAVELET PACKET DECOMPOSITION to each destination nodeiProcessed, 16 groups little are most obtained at last
Ripple bag decomposition coefficient, is designated as YI, j, wherein i=0,1 ..., N-1 is purpose node serial number, and j=0,1 ..., 15 is wavelet packet point
Solution coefficient number, XiWith YI, jIt is all vector.Because this 16 groups of WAVELET PACKET DECOMPOSITION coefficients are not can correctly to reflect egress
Between dependency, therefore feature selection is carried out to this 16 groups of WAVELET PACKET DECOMPOSITION coefficients, obtain some of wherein best results
System number.
S3, carry out feature selection using simulated annealing, with the error known to the topological sum for estimating between topology come
The quality that judging characteristic is selected;Including following sub-step:
Some groups are randomly choosed in S31, the 16 groups of WAVELET PACKET DECOMPOSITION coefficients obtained from S2 as the feature estimated first, will
Feature selection state vector F=(f1, f2..., f16) represent, wherein f1, f2..., f16∈ { 0,1 }, represents when value is 0
This feature is chosen in unselected this feature, 1 expression;
R is made to be results set, for storage feature selection result and the key-value pair of its error rate composition, R={ } when initial;
Initialization temperature T=T0=2m, m are purpose node number;
Iterationses upper limit L=20 under same temperature is set, iteration count l=0 is initialized, initialization new explanation does not receive
Number of times t=0, lower the temperature frequency n=0;
S32, sample is carried out according to F estimating based on the topology of Agglomerative Hierarchical Clustering, and calculated with tree edit distance and open up
The error rate e (F) of estimated result is flutterred, R=R ∪ { F are updated:e(F)};
Estimate that concrete methods of realizing is based on the topology of Agglomerative Hierarchical Clustering:
S321, each cluster the distance between sample is defined, obtain distance matrix;For each destination node i, if
The feature chosen isWherein 0≤k1< k2< ... < kn≤ 15 and k1, k2... kn∈N;Feature is spelled
It is connected in a new vector
For any two destination node i, j, its characteristic vector Y is soughtiWith YjBetween correlation coefficient:
WhereinWithY is represented respectivelyiWith YjMathematic expectaion be average:
Node i is defined, the correlation distance between j is:
dI, j=1- | ρI, j|
When two destination nodes i, when the characteristic vector dependency of j is stronger, the absolute value ρ of its correlation coefficientI, jJust closer to
1, then correlation distance d between the twoI, jJust closer to 0;
Obtain the correlation distance matrix between all purposes node:
dI, j=dJ, i, (i, j=1,2 ..., n) and d1,1=d2,2=...=dN, n=0;So matrix D is a diagonal
For 0 symmetrical matrix;
S322, hierarchical clustering is carried out, obtain topology, comprised the following steps:
Each sample is each classified as into a class when S3221, initialization, common n classes are designated as G1, G2..., Gn, make it distinguish
N leaf node of clustering tree is constituted, N is designated as1, N2..., Nn;Weight w is assigned for each node1, w2..., wn, make w1=w2
=...=wn=0, between class distance matrix is G=D;Any two class Gi, GjThe distance between be gI, j=dI, j, initialization cluster time
Number l=1;
S3222, in G find minimum range gS, t=mini≠j(gI, j), by corresponding two classes Gs, GtAggregate into one
New class Gn+l, construct new node Nn+lAs Ns, NtFather node, make its weight wn+l=gS, t;
S3223, calculate new class G using formulan+lTo the distance of other classes, two classes Gp, GqThe distance between computing formula
For:
S3224, wherein np, nqRespectively class Gp, GqIn number of samples;
S3225, renewal between class distance matrix G:By in G with Gs, GtRelated row and column is eliminated, and is represented in last addition
New class Gn+lRanks, corresponding value is distance of the new class to other classes;
S3226, renewal cluster number of times l=l+1;
S3227, repetition S3222~S3225, until only remaining class G2n-1;
In above-mentioned cluster process, two classes are all merged into a new class by each wheel, until only remaining a class and raw
Binary tree is weighed into tax of the nodes for 2n-1;According to the node on following compatible rule merging binary tree:One thresholding t is set
=0.02, from node Nn+1Start to N2n-2Till, if a node NiWith its father node NjWeights meet
Then by NiWith NjMerge, will NiThe father node of child node be changed to NjAfter delete Ni, the tree for finally giving is used as this
The Network traffic model of secondary estimation.
The concrete methods of realizing that the error rate of topological estimated result is calculated with tree edit distance is:
Define the edit operation of three kinds of trees:
Change node label:The label for defining the leaf node of tree-shaped network topology is each self-corresponding numbering, remaining section
Point label is sky;If node label is changed to into cost r (a → b)=2m of b by a, wherein m is purpose section during network detection
Point number, that is, the leaf node number set;
Deletion of node:Non-root node v in tree T is deleted, if its father node is v ', the father node of the child node of v is changed to
v′;If cost r (a → Λ)=1, a of deletion of node is the label of node to be deleted, Λ represents empty node;
Increase node:The father node of the part child node of v ' and is changed to v by newly-increased node v as the child node of node v ';
If cost r (Λ → a)=1, the Λ for increasing node represents empty node, a is the label of node to be increased;
If E is from tree T1To tree T2Editing process, it comprises several tree edit operation e1, e2..., en;Remember T1
It is converted into T2Totle drilling cost r (E)=r (e1)+r(e2)+…+r(en);Tree edit distance is min (r (E)), will T1It is converted into T2
Minimum total cost;
To try to achieve tree edit distance, problem is converted to and seeks two tree T1With T2Between maximum match subtree problem:If altogether
It is M, T with subtree1With T2Matching relationship be (M, T1, T2), belong to T1But it is N to be not belonging to the node set of M1, belong to T2But no
The node set for belonging to M is N2, T1The label of interior joint i is designated as l1(i), T2The label of interior joint i is designated as l2(i);Then have:
By dynamic programming algorithm, r ((M, T are obtained1, T2)) minima, i.e. tree edit distance.
S33, as l < L execution step S34, otherwise skip to step S39;
S34, change F in ith feature selection state, keep the selection state of other features constant, obtain one group it is new
Solution F 'i=(f 'I, 1, f 'I, 2..., f 'I, 16), its satisfaction:
Make i take all over all values, obtain 16 groups of new explanations:F′1, F '2..., F '16;
S35, judge that new explanation whether there is in R, for the solution that is not present in R is performed and step S32 identical method,
Obtain e (F 'i), update R=R ∪ { F 'i:e(F′i), and find e*=min (e (F 'i))=e (F 'm) with corresponding F 'm;
S36, l=l+1 is made, calculate Δ e=e*- e (F), new explanation F ' is received according to metropolis criterions with Probability pm, p's
Calculation is as follows:
If receiving new explanation, F=F ' is madem, e (F)=e*, t=0 skips to step S33;Otherwise t=t+1, carries out S37;
S37, as l < L execution step S38, otherwise skip to step S39;
S38, for each feature in F changes existing selection state with 50% probability, obtain one group it is new
Solution F 'm, perform and step S32 identical method, obtain e*=e (F 'm), update R=R ∪ { F 'm:e(F′m), skip to step
S36;
S39, n=n+1 is made,L=0;As t < 10 and n < 10, step S33 is skipped to;Otherwise select most from R
Excellent solution F*As the result of feature selection, terminate algorithm.
The feature used when S4, estimation that the feature chosen is topological as other all unknown networks, utilizes and step S32
Middle identical Agglomerative Hierarchical Clustering method estimates network topology.
One of ordinary skill in the art will be appreciated that embodiment described here is to aid in reader and understands this
Bright principle, it should be understood that protection scope of the present invention is not limited to such especially statement and embodiment.This area
It is each that those of ordinary skill can make various other without departing from essence of the invention according to these technologies enlightenment disclosed by the invention
Plant concrete deformation and combine, these deformations and combination are still within the scope of the present invention.
Claims (6)
1. the Estimating topology of networks method of simulated annealing is based on, it is characterised in that comprised the following steps:
S1, the network to known network topological structure carry out network tomography detection;
S2, using wavelet packet decomposition algorithm process result of detection obtain cluster feature;
S3, carry out feature selection using simulated annealing;
The feature used when S4, estimation that the feature chosen is topological as other all unknown networks, using Agglomerative Hierarchical Clustering
Method estimates network topology.
2. the Estimating topology of networks method based on simulated annealing according to claim 1, it is characterised in that step S2
Concrete methods of realizing is:Each time iteration is two parts of low frequency and high frequency all by last signal subdivision to WAVELET PACKET DECOMPOSITION,
Wherein S is primary signal, and A represents approximate filtration, is considered as the low frequency part of front first order signal;D represents details filtration, is considered as previous
The HFS of rank signal;
Wherein y [n] is signal to be decomposed, and g [n] is low pass filter, and h [n] is high pass filter, meets h [n]=(- 1)ng[1-
n];G [n] correspond to scaling function ψ (t) of wavelet transformation, referred to as scaling vector;H [n] correspond to the wavelet function of wavelet transformation
Calculated as g [n] using the scaling vector corresponding to db4 small echos:Signal is carried out respectively low pass, high-pass filtering
Afterwards, carry out down-sampled, respectively obtain the WAVELET PACKET DECOMPOSITION coefficient of low frequency and high frequency:
In the same manner, to Y1,0[n] and Y1,1[n] carries out same treatment, obtains 4 groups of WAVELET PACKET DECOMPOSITION coefficients of the second layer, by that analogy,
Obtain the WAVELET PACKET DECOMPOSITION coefficient of any number of plies;Its recurrence formula is:
Using primary signal X of the 4 layers of WAVELET PACKET DECOMPOSITION to each destination nodeiProcessed, 16 groups of wavelet packets point are most obtained at last
Solution coefficient, is designated as YI, j, wherein i=0,1 ..., N-1 is purpose node serial number, and j=0,1 ..., 15 is WAVELET PACKET DECOMPOSITION coefficient
Numbering, XiWith YI, jIt is all vector.
3. the Estimating topology of networks method based on simulated annealing according to claim 2, it is characterised in that step S3
Including following sub-step:
Some groups are randomly choosed as the feature estimated first, in S31, the 16 groups of WAVELET PACKET DECOMPOSITION coefficients obtained from S2 by feature
Selection state vector F=(f1, f2..., f16) represent, wherein f1, f2..., f16∈ { 0,1 }, represents unselected when value is 0
This feature is chosen in middle this feature, 1 expression;
R is made to be results set, for storage feature selection result and the key-value pair of its error rate composition, R={ } when initial;Initially
Change temperature T=T0=2m, m are purpose node number;
Iterationses upper limit L=20 under same temperature is set, iteration count l=0 is initialized, initialization new explanation does not receive number of times
T=0, lower the temperature frequency n=0;
S32, sample is carried out according to F estimating based on the topology of Agglomerative Hierarchical Clustering, and calculate topology with tree edit distance and estimate
The error rate e (F) of meter result, updates R=R ∪ { F:e(F)};
S33, as l < L execution step S34, otherwise skip to step S39;
S34, the selection state for changing ith feature in F, keep the selection state of other features constant, obtain one group of new solution
F′i=(f 'I, 1, f 'I, 2..., f 'I, 16), its satisfaction:
Make i take all over all values, obtain 16 groups of new explanations:F′1, F '2..., F '16;
S35, judge that new explanation whether there is in R, for the solution that is not present in R is performed and step S32 identical method, obtain
e(F′i), update R=R ∪ { F 'i:e(F′i), and find e*=min (e (F 'i))=e (F 'm) with corresponding F 'm;
S36, l=l+1 is made, calculate Δ e=e*- e (F), new explanation F ' is received according to metropolis criterions with Probability pm, the calculating of p
Mode is as follows:
If receiving new explanation, F=F ' is madem, e (F)=e*, t=0 skips to step S33;Otherwise t=t+1, carries out S37;
S37, as l < L execution step S38, otherwise skip to step S39;
S38, for each feature in F changes existing selection state with 50% probability, obtain one group of new solution F 'm,
Perform and step S32 identical method, obtain e*=e (F 'm), update R=R ∪ { F 'm:e(F′m), skip to step S36;
S39, n=n+1 is made,L=0;As t < 10 and n < 10, step S33 is skipped to;Optimal solution is otherwise selected from R
F*As the result of feature selection, terminate algorithm.
4. the Estimating topology of networks method based on simulated annealing according to claim 3, it is characterised in that the step
Estimate that concrete methods of realizing is based on the topology of Agglomerative Hierarchical Clustering in S32:
S321, each cluster the distance between sample is defined, obtain distance matrix;For each destination node i, if choosing
Feature beWherein 0≤k1< k2< ... < kn≤ 15 and k1, k2... kn∈N;It is by merging features
One new vector
For any two destination node i, j, its characteristic vector Y is soughtiWith YjBetween correlation coefficient:
WhereinWithY is represented respectivelyiWith YjMathematic expectaion be average:
Node i is defined, the correlation distance between j is:
dI, j=1- | ρI, j|
When two destination nodes i, when the characteristic vector dependency of j is stronger, the absolute value ρ of its correlation coefficientI, jJust closer to 1, then
Correlation distance d between the twoI, jJust closer to 0;
Obtain the correlation distance matrix between all purposes node:
dI, j=dJ, i, (i, j=1,2 ..., n) and d1,1=d2,2=...=dN, n=0;So it is 0 that matrix D is a diagonal
Symmetrical matrix;
S322, hierarchical clustering is carried out, obtain topology, comprised the following steps:
Each sample is each classified as into a class when S3221, initialization, common n classes are designated as G1, G2..., Gn, make it respectively constitute
N leaf node of clustering tree, is designated as N1, N2..., Nn;Weight w is assigned for each node1, w2..., wn, make w1=w2=...
=wn=0, between class distance matrix is G=D;Any two class Gi, GjThe distance between be gI, j=dI, j, initialization cluster number of times l=
1;
S3222, in G find minimum range gS, t=mini≠j(gI, j), by corresponding two classes Gs, GtAggregate into a new class
Gn+l, construct new node Nn+lAs Ns, NtFather node, make its weight wn+l=gS, t;
S3223, calculate new class G using formulan+lTo the distance of other classes, two classes Gp, GqThe distance between computing formula be:
S3224, wherein np, nqRespectively class Gp, GqIn number of samples;
S3225, renewal between class distance matrix G:By in G with Gs, GtRelated row and column is eliminated, and represents new class in last addition
Gn+lRanks, corresponding value is distance of the new class to other classes;
S3226, renewal cluster number of times l=l+1;
S3227, repetition S3222~S3225, until only remaining class G2n-1。
5. the Estimating topology of networks method based on simulated annealing according to claim 4, it is characterised in that the step
In cluster process, two classes are all merged into a new class to S322 by each wheel, until only remaining a class, and generate one
Nodes weigh binary tree for the tax of 2n-1;According to the node on following compatible rule merging binary tree:One thresholding t=0.02 is set,
From node Nn+1Start to N2n-2Till, if a node NiWith its father node NjWeights meet
Then by NiWith NjMerge, will NiThe father node of child node be changed to NjAfter delete Ni, the tree for finally giving estimates as this
The Network traffic model of meter.
6. the Estimating topology of networks method based on simulated annealing according to claim 4, it is characterised in that the step
The concrete methods of realizing of the error rate for calculating topological estimated result with tree edit distance in S32 is:
Define the edit operation of three kinds of trees:
Change node label:Define tree-shaped network topology leaf node label be each self-corresponding numbering, remaining node mark
Sign as sky;If node label is changed to into cost r (a → b)=2m of b by a, wherein m is destination node during network detection
Number, that is, the leaf node number set;
Deletion of node:Non-root node v in tree T is deleted, if its father node is v ', the father node of the child node of v v ' is changed to into;If
Cost r (a → Λ)=1, a of deletion of node is the label of node to be deleted, and Λ represents empty node;
Increase node:The father node of the part child node of v ' and is changed to v by newly-increased node v as the child node of node v ';If increasing
Plus cost r (Λ → a)=1, the Λ of node represents empty node, a is the label of node to be increased;
If E is from tree T1To tree T2Editing process, it comprises several tree edit operation e1, e2..., en;Remember T1Conversion
To T2Totle drilling cost r (E)=r (e1)+r(e2)+…+r(en);Tree edit distance is min (r (E)), will T1It is converted into T2Most
Little totle drilling cost;
To try to achieve tree edit distance, problem is converted to and seeks two tree T1With T2Between maximum match subtree problem:If common son
Set as M, T1With T2Matching relationship be (M, T1, T2), belong to T1But it is N to be not belonging to the node set of M1, belong to T2But it is not belonging to M
Node set be N2, T1The label of interior joint i is designated as l1(i), T2The label of interior joint i is designated as l2(i);Then have:
By dynamic programming algorithm, r ((M, T are obtained1, T2)) minima, i.e. tree edit distance.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610817914.6A CN106685686B (en) | 2016-09-12 | 2016-09-12 | Network topology estimation method based on simulated annealing |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610817914.6A CN106685686B (en) | 2016-09-12 | 2016-09-12 | Network topology estimation method based on simulated annealing |
Publications (2)
Publication Number | Publication Date |
---|---|
CN106685686A true CN106685686A (en) | 2017-05-17 |
CN106685686B CN106685686B (en) | 2020-09-18 |
Family
ID=58840010
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201610817914.6A Active CN106685686B (en) | 2016-09-12 | 2016-09-12 | Network topology estimation method based on simulated annealing |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN106685686B (en) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107426207A (en) * | 2017-07-21 | 2017-12-01 | 哈尔滨工程大学 | A kind of network intrusions method for detecting abnormality based on SA iForest |
CN111478807A (en) * | 2020-04-02 | 2020-07-31 | 山东省计算中心(国家超级计算济南中心) | Construction method of minimum feedback node set of directed multilayer network |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101540061A (en) * | 2009-04-10 | 2009-09-23 | 西北工业大学 | Topological and ordering matching method for disordered images based on simulated annealing |
CN102711206A (en) * | 2012-05-14 | 2012-10-03 | 南京邮电大学 | Simulated annealing-based wireless sensor network (WSN) hierarchical routing method |
CN103281256A (en) * | 2013-04-26 | 2013-09-04 | 北京邮电大学 | Network tomography-based end-to-end path packet loss rate detection method |
CN103678917A (en) * | 2013-12-13 | 2014-03-26 | 杭州易和网络有限公司 | Bus real-time arrival time predicting method based on simulated annealing algorithm |
WO2015161872A1 (en) * | 2014-04-23 | 2015-10-29 | Telefonaktiebolaget L M Ericsson (Publ) | Network tomography through selection of probing paths |
-
2016
- 2016-09-12 CN CN201610817914.6A patent/CN106685686B/en active Active
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101540061A (en) * | 2009-04-10 | 2009-09-23 | 西北工业大学 | Topological and ordering matching method for disordered images based on simulated annealing |
CN102711206A (en) * | 2012-05-14 | 2012-10-03 | 南京邮电大学 | Simulated annealing-based wireless sensor network (WSN) hierarchical routing method |
CN103281256A (en) * | 2013-04-26 | 2013-09-04 | 北京邮电大学 | Network tomography-based end-to-end path packet loss rate detection method |
CN103678917A (en) * | 2013-12-13 | 2014-03-26 | 杭州易和网络有限公司 | Bus real-time arrival time predicting method based on simulated annealing algorithm |
WO2015161872A1 (en) * | 2014-04-23 | 2015-10-29 | Telefonaktiebolaget L M Ericsson (Publ) | Network tomography through selection of probing paths |
Non-Patent Citations (2)
Title |
---|
汪马翔: "基于PMU测量数据的暂态稳定预测分类特征选择", 《中国优秀博硕士学位论文全文数据库 (硕士)(工程科技Ⅱ辑)》 * |
王鑫: "基于小波变换的非平稳网络拓扑估计方法研究", 《中国优秀硕士学位论文全文数据库(信息科技辑)》 * |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107426207A (en) * | 2017-07-21 | 2017-12-01 | 哈尔滨工程大学 | A kind of network intrusions method for detecting abnormality based on SA iForest |
CN107426207B (en) * | 2017-07-21 | 2019-09-27 | 哈尔滨工程大学 | A kind of network intrusions method for detecting abnormality based on SA-iForest |
CN111478807A (en) * | 2020-04-02 | 2020-07-31 | 山东省计算中心(国家超级计算济南中心) | Construction method of minimum feedback node set of directed multilayer network |
CN111478807B (en) * | 2020-04-02 | 2023-03-24 | 山东省计算中心(国家超级计算济南中心) | Construction method of minimum feedback node set of directed multilayer network |
Also Published As
Publication number | Publication date |
---|---|
CN106685686B (en) | 2020-09-18 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Tang et al. | Exploring dynamic property of traffic flow time series in multi-states based on complex networks: Phase space reconstruction versus visibility graph | |
CN104657744B (en) | A kind of multi-categorizer training method and sorting technique based on non-determined Active Learning | |
Wang et al. | Deep learning-enabled sparse industrial crowdsensing and prediction | |
CN107578061A (en) | Based on the imbalanced data classification issue method for minimizing loss study | |
CN107704877A (en) | A kind of image privacy cognitive method based on deep learning | |
Wang et al. | Cell selection with deep reinforcement learning in sparse mobile crowdsensing | |
CN108090510A (en) | A kind of integrated learning approach and device based on interval optimization | |
Chen et al. | Faster fundamental graph algorithms via learned predictions | |
Hussain et al. | Intelligent traffic flow prediction using optimized GRU model | |
Song et al. | Online traffic flow prediction for edge computing-enhanced autonomous and connected vehicles | |
Tomani et al. | Towards trustworthy predictions from deep neural networks with fast adversarial calibration | |
Shi et al. | Temporal dynamic matrix factorization for missing data prediction in large scale coevolving time series | |
CN116340524B (en) | Method for supplementing small sample temporal knowledge graph based on relational adaptive network | |
CN108596204B (en) | Improved SCDAE-based semi-supervised modulation mode classification model method | |
CN115545350A (en) | Comprehensive deep neural network and reinforcement learning vehicle path problem solving method | |
Nettasinghe et al. | Influence maximization over Markovian graphs: A stochastic optimization approach | |
CN106685686A (en) | Network topology estimation method based on simulated annealing | |
Zhu et al. | Adversarial training of LSTM-ED based anomaly detection for complex time-series in cyber-physical-social systems | |
Liu et al. | RT-GCN: Gaussian-based spatiotemporal graph convolutional network for robust traffic prediction | |
Dash | DECPNN: A hybrid stock predictor model using Differential Evolution and Chebyshev Polynomial neural network | |
CN115526293A (en) | Knowledge graph reasoning method considering semantic and structural information | |
CN113806561A (en) | Knowledge graph fact complementing method based on entity attributes | |
Patel et al. | Metaheuristic enabled deep convolutional neural network for traffic flow prediction: Impact of improved lion algorithm | |
Bhaumik et al. | STLGRU: Spatio-Temporal Lightweight Graph GRU for Traffic Flow Prediction | |
Zhao et al. | ATR: automatic trajectory repairing with movement tendencies |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |