CN103310275A - Novel codebook design method based on ant colony clustering and genetic algorithm - Google Patents

Abstract

The invention provides a novel codebook design method based on an ant colony clustering and genetic algorithm. The method includes the following steps: step one, training data are distributed in two-dimensional spaces of different dimensions randomly by using ants of different numbers, and an LF algorithm is adopted to perform generation of initial clustering; step two, clustering correction performed by the initial clustering according to the dimension of a codebook guarantees that the clustering number is the same as initial setup codebook dimensions; step three, under the premise that initial population is obtained successfully, individual selection, intersection and mutation operation are performed according to the basic procedure of a genetic algorithm until iteration is stopped and an optimal individual meeting requirements is obtained. The novel codebook design method based on the ant colony clustering and genetic algorithm overcomes the defect that correlation of initial selection and final design results is strong in an LBG algorithm, meanwhile, prevents similar LBG algorithm from getting into the inferior position of local optimum, and is suitable for the fields of large quantity processing, voice communication, mode recognition, internet protocol (IP) telephony and the like.

Description

Novel code book method for designing based on ant colony clustering and genetic algorithm

Technical field

The invention provides a kind of novel code book method for designing based on ant colony clustering and genetic algorithm, what it related to is a kind of novel code book method for designing for the used code book of vector quantization in the signal processing, what specially refer to is the code book design that quantizes for the line spectrum pair frequency parameter in the MELP (Mixed Excitation Linear Prediction) encoding scheme, the method can effectively weaken designing institute and get correlativity between code book and the training sequence, simultaneously effectively prevent from being absorbed in local optimum in the code book design process, be applicable at data dimension larger, distribute in the less situation of the bit number be used for transmitting and realize that the high precision of data quantizes.

Technical background

Progressively development along with society, the amount of communication data of mobile communication just progressively strengthens, traditional communication resource such as frequency spectrum can't satisfy daily communication requirement, the concept of vector quantization has been proposed for this kind present situation researchist, vector quantization can effectively improve data-handling efficiency, reduce the communication resource that it takies, thereby be widely used in the fields such as satellite communication, pattern-recognition, IP phone and large data compression.The concrete grammar of vector quantization is that the data of a multidimensional of transmitting terminal are processed, Euclidean distance by (being referred to as code vector) between the same dimension data of calculating these data and some, and then pick out between the two apart from minimum vector, and it is used as substituting of pending data, what data were transmitted when transmitting is the order label of minor increment code vector in all code vectors, and data receiver data-driven label after receive data carries out search index, thereby the optimum that obtains data that transmitting terminal is processed substitutes.The design that this shows code book is most important for the effect of vector quantization, effectively, code book can guarantee at data receiver raw data as far as possible accurately to be recovered accurately, so how in the situation that limited training sequence obtains the study hotspot that the code book of effect optimum just becomes the signal process field.

The algorithm of code book design occur the earliest be the LBG algorithm, proposed in 1980 by Linde, Buzo and Gray, this method does not need to know the probability distribution of training sequence, approaches optimum code book by training sequence being taked certain iterative algorithm.The key step of LBG algorithm is: 1, from training sequence, choose randomly meet the code book size number vector as initial code vector; 2, with vectors all in the training sequence according to the principle of minimum Eustachian distance be assigned to initial code vector in the first step around, be referred to as clustering processing; 3, the vector of all training sequences that comprise for each cluster calculates the centre of form, as new code vector; 4, calculate the distortion situation that adopts novel all vectors of codebook quantification training sequence, if distortion reaches the thresholding of initial setting then stops, finishing the code book training, otherwise re-cover step 2 and 3.The advantage of LBG algorithm is that operation steps is simple, and the code book training time is shorter, can satisfy most code book designing requirement, but the limitation of this algorithm is obvious equally: the random system of selection of taking in the time of at first initialized causes the code book of final gained when initial selected is different, its quantification effect for training sequence differs larger, and namely the correlativity of optimum code book and initial code vector is stronger; Secondly, the design phase easily is absorbed in local optimum, causes the convergence effect that can't finally be needed.For these two shortcomings of LBG algorithm, many experts and scholars both domestic and external have carried out long research, and have proposed multiple effective code book method for designing, have effectively improved the effect of code book design.

ZongboXie and JiuchaoFeng have proposed a kind of code book method for designing based on Mo Keer fuzzy clustering and gradient algorithm in paper " Codebook Design For Vector Quantization Based on a Kernel Fuzzy Learning Algorithm ", in the scheme Euclidean distance computing among the LBG is taked to examine the method for distance, greatly reduce computation complexity, simulated effect show code the design's effect is compared with LBG to some extent and is promoted simultaneously.H.B.Kekre and ChetanAgarwal successfully are applied to genetic algorithm and simulated annealing in the code book design in article " Codebook optimization using genetic algorithm and simulated annealing ", under the prerequisite that does not additionally increase code book size, code book has been carried out further optimization, and fully used the good ability of searching optimum of genetic algorithm, avoided being absorbed in local optimum, effect is remarkable.Chen Qian carries out the iteration cluster operation three times by all individualities before the individual choice in to the genetic algorithm process in paper " quantizing to improve algorithm based on the image vector of hereditary LBG ", when taking full advantage of LBG algorithm local search ability, improved the effect of code book design and accelerated convergence of algorithm speed.Most of code book method for designing that proposes now and used all can't improve design effect and weaken initial selected and net result between correlativity obtain to take into account, although the algorithm that has simultaneously because be easy to is absorbed in the thresholding that code book that local optimum causes calculating at last can satisfy initial setting, but the actual quantification effect of using is relatively poor, therefore propose a kind ofly can effectively avoid being absorbed in local optimum, and the less code book method for designing of final code book and initial selected correlativity is the problem that present urgent need will be researched and solved.

Summary of the invention

1, purpose:

In order to improve the quantification effect of code book, and then the degree of accuracy of lifting vector quantization in actual signal is processed, traditional code book method for designing has plenty of and solely utilizes ant colony clustering algorithm, take full advantage of the lower advantage of its cluster result and initial setting up correlativity, the code book that obtains like this can't satisfy application request on quantification effect, and within the limited training time, can't reach convergence state sometimes, some schemes are solely to utilize genetic algorithm or simulated annealing, although avoided being absorbed in local optimum, but scheme still needs initial individual selection in initial, correlativity is close between optimal result and the initial selected, the scheme that also has is that to sacrifice quantification effect be cost, calculated amount in the design of reduction code book and time complexity, most method all can't be in quantification effect and real time complexity, obtain between the initial correlativity and take into account, the purpose of this invention is to provide a kind of novel code book method for designing based on ant colony clustering and genetic algorithm, it is to propose a kind of codebook design schemes based on ant colony clustering and genetic algorithm, this scheme takes full advantage of the very low advantage of ant colony clustering algorithm and initial selected correlativity in the starting stage, in the two-dimensional space of different size, take the ant of different numbers to carry out the generation of initial clustering bunch, the centre of form of calculating each clustering cluster is used as it as individual chromosome in the initial population in the genetic manipulation, then according to genetic algorithm it is taked to select, intersect and mutation operation, in the practical operation for the feature of real application data, dynamic interlace operation and sudden change scheme have been adopted, until being quantification effect, the fitness of optimum individual reaches the thresholding of setting, when having avoided design process to be absorbed in local optimum, take full advantage of the outstanding ability of searching optimum of genetic algorithm, effectively improved the effect of code book design.

2, technical scheme

The present invention is characterized in: the code book design phase at first takes the LF algorithm in the ant colony clustering algorithm to carry out the generation of initial clustering, the cardinal principle of this algorithm is to copy occurring in nature ant heap to be directly proportional for the size that attractive force is with ant is piled of the worker ant of carrying ant corpse, the larger attractive force of ant heap is larger, otherwise less, thereby form a kind of positive feedback, number for the size of the two-dimensional space of taking in this stage and ant is constantly to change, and then obtains different Clustering Effects; Then, the number of the clustering cluster that may exist for the cluster that obtains in the first step and the different shortcoming of code book size of setting are carried out the correction of clustering cluster; At last, for the initial individuality that obtains before, it is used as the individuality of first generation population, take the basic procedure of genetic algorithm, carry out selection, intersection and the mutation operation of genetic algorithm, finally shut-down operation when the fitness of optimum individual surpasses the initial setting thresholding is used as optimum individual output as the result that code book designs.

The novel code book method for designing that is based on ant colony clustering and genetic algorithm that Fig. 1 provides, a kind of novel code book method for designing based on ant colony clustering and genetic algorithm of the present invention, its concrete step is as follows:

Step 1: utilize the ant of different numbers training data to be distributed in randomly on the two-dimensional space of different size, take the LF algorithm to carry out the generation of initial clustering according to Fig. 2.

Step 2: the correction of the clustering cluster that the initial clustering that carries out according to Fig. 3 carries out according to code book size,, guarantee that the number of cluster is identical with the code book size of initial setting.

Step 3: under the prerequisite that initial population successfully obtains, according to the basic procedure of Fig. 4 according to genetic algorithm, carry out individual choice, intersection and mutation operation, until iteration stopping is met the optimum individual of requirement.

Wherein, that uses in step 1 takes the LF algorithm to carry out the generation of initial clustering according to Fig. 2, and the parameter of using in the cluster process mainly comprises ant number m, two-dimensional space size Z, cluster radius s, distinctiveness ratio constant alpha, and constant value k _1,k ₂, the carrying probability P _pPut down probability P _d

In step 1, take the LF algorithm to carry out the basic step of generation of initial clustering as follows:

(1) setting initial ant number is m, and step-length is ma, and the two-dimensional space size is initially Z, and step-length is za, and the data with training sequence under current ant number and two-dimensional space are assigned on the two-dimensional space randomly, and all ants all are arranged to light condition.This moment is for the arbitrary element O in the training sequence _i, its position X _iY _iAll be to be in randomly in the Z*Z space.

(2) for each ant, on Cluster space, choose randomly a position, if being in, current ant have data to exist on light condition and the current location, calculate the similarity of all data elements in current data element and the cluster radius centered by current location, one (0 of random generation, 1) numerical value in the scope, if random number greater than the carrying probability then ant carries this data, data empty on the current location.

(3) if do not have data and ant to be in the carrying state on the current location, use the similarity that calculates in the step (2) to calculate current ant for the probability that puts down of current data, calculating is judged after finishing: generate at random one (0,1) numerical value in the scope, if random number surpasses and to put down probability then ant is put down data, ant is arranged to light condition.

(4) position of ant carries out randomly that fixed length moves, and namely X or Y coordinate plus-minus step-length repeats move operation until the current position that moves to is not occupied by other ant.

(5) repeating step (2)-(4) are until to setting the equal complete operation of ant of number, obtain the cluster result (code book size that the clusters number that obtain this moment may reach with setting is different) for raw data after the complete operation.

(6) according to the ant number step-length ma that sets in the step (1) and two-dimensional space step-length za change initial parameter, repeating step (2)-(5).

Wherein, the correction of the clustering cluster that the initial clustering that carries out according to Fig. 3 in step 2 carries out according to code book size is consistent with the number that guarantees clustering cluster with code book size.

The basic step of the correction of the clustering cluster that the initial clustering that carries out in step 2 carries out according to code book size is as follows:

(1) initial parameter is set: code book size M, the clusters number K that obtains among Fig. 2.

(2) if M＜K, the cluster data that obtains among Fig. 2 is carried out descending sort according to the number of training sequence vector in each clustering cluster, then maximum clustering cluster and min cluster bunch are merged, replace original maximum clustering cluster, repeat union operation until clusters number is identical with the code book size of initial setting, i.e. M=K.

(3) if M〉K, the cluster data that obtains among Fig. 2 is carried out descending sort according to the number of training sequence vector in each clustering cluster, then the data in the maximum clustering cluster are divided into two according to the size to this clustering cluster centre of form distance, generate two new clustering cluster, repeat fractured operation until clusters number is identical with the code book size of initial setting, i.e. M=K.

(4) carry out centroid calculation for the current M that an obtains clustering cluster, the centre of form is used as code vector in the initial codebook.

Wherein, the basic procedure of the genetic algorithm of in step 3, using, adopt genetic algorithm will before the initial population that obtain select, the step of intersection and gene mutation operation, the parameter of wherein using comprises Population Size N _{Pop_size}, mutation probability P _m, crossover probability P _cWith iteration stopping thresholding ε.

Wherein, the basic procedure of the genetic algorithm that adopts in the step 3, its concrete step is as follows:

(1) for all individuality calculating fitness value separately, then individuality is carried out descending sort according to fitness, whether judges the fitness of optimum individual greater than the iteration stopping thresholding, and if greater than the iteration stopping thresholding iteration work finish, enter step (4), on the contrary the step of entering (2).

(2) according to crossover probability P _cP before selecting in the ordering individual sequence _c* N _{Pop_size}Individuality directly enters the next generation, then carry out the individual generation work of residue of future generation: to ordering individual sequence, according to begin first and pairing last, second principle of matching with penultimate in front from the front, two individualities of pairing are used as parent, then take the single-point interlace operation to generate two new individualities, be placed into and be used as filial generation among the next generation.Repeating crossover is until the total individual number order satisfies Population Size.

(3) according to the gene mutation probability P _mChoose P _m* N _{Pop_size}Individuality carries out mutation operation, and operating process is: for each individual random number that generates in (0, a 1) scope, if random number is greater than P _mThen current individuality enters mutation operation, takes interior among a small circle sudden change for the gene on the chromosome of current individuality, and sudden change numerical value is the function of the fitness of individuality; If at random less than P _mThen current individual the preservation do not enter mutation operation.All finish selection and mutation operation for all individualities, enter step (1).

(4) individuality of output fitness maximum is as the result of final code book design.

3, advantage and effect:

This novel code book method for designing based on ant colony clustering and genetic algorithm provided by the invention has not only overcome the stronger shortcoming of the results relevance of initial selected and final design in the LBG algorithm, has avoided simultaneously similar LBG algorithm to be absorbed in the inferior position of local optimum.In the code book design process, at first by using the LF algorithm in the ant colony clustering algorithm, use the ant of different numbers on different two-dimensional spaces, training data to be carried out cluster, then may the characteristics different from code book size for cluster gained clustering cluster number, adopt the amendment scheme of introducing among Fig. 2, thereby the number of vector is identical with the size of code book in the individuality that guarantees to obtain, it is individual that the individuality that then will obtain is used as the first generation initial in the operatings of genetic algorithm, carrying out the ideal adaptation degree calculates, individual choice, intersect and the gene mutation operation, final fitness when optimum individual stops greater than the iteration stopping thresholding time, output optimum individual has at this moment guaranteed the carrying out of global optimization as the result of code book design.

The present invention is applicable to the fields such as large quantity processing, voice communication, pattern-recognition, IP phone, and the present invention mainly contains following advantage:

(1) use the LF algorithm to carry out the generation of initial clustering, it is irrelevant that the advantage that takes full advantage of the LF algorithm is that the initial position of last Clustering Effect and ant is selected, and so just fully having overcome different initial individual selections in the LBG algorithm affects for the strong correlation of final code book design result.

(2) use genetic algorithm to finish individual selection, intersection and mutation operation, finally obtain qualified individuality by continuous iteration, take full advantage of the stronger ability of searching optimum of genetic algorithm, overcome the shortcoming that the LBG algorithm easily is absorbed in local optimum, guaranteed that final design codebooks is based on whole training sequence and obtains.

(3) there is not the problem that is similar to the empty bag of LBG algorithm generation chamber in genetic algorithm in operating process, and the Da Bao chamber Split Method that does not propose for sky bag chamber has improved Clustering Effect in design process.

Description of drawings

Fig. 1 the method for the invention overview flow chart

Fig. 2 adopts the LF algorithm to carry out the process flow diagram of the generation of initial clustering

The process flow diagram of the correction of the clustering cluster that Fig. 3 initial clustering carries out according to code book size

The basic flow sheet of Fig. 4 genetic algorithm

M: code book size

O _i: data element

Z*Z: the two-dimensional space of distribution

S: cluster radius

α: distinctiveness ratio constant

P _p: the loading data probability

P _d: put down the data probability

N _{Pop_size}: Population Size

P _c: crossover probability

P _m: the gene mutation probability

ε: iteration stopping thresholding

Embodiment

Below in conjunction with accompanying drawing 1,2,3,4, the line spectrum pair frequency parameter of using in the MELP (Mixed Excitation Linear Prediction) encoding scheme is specifically introduced the given novel code book method for designing of the present invention as example, and wherein the line spectrum pair frequency parameter is 10 dimensions.

A kind of novel code book method for designing based on ant colony clustering and genetic algorithm of the present invention, its concrete step is as follows:

Step 1: default code book is of a size of 64, i.e. Population Size N _{Pop_size}Be 64.According to the scheme that Fig. 2 provides, the initial number of at first setting ant is 100, and step-length is 5, and two-dimensional space is of a size of 100, and step-length is 10, and cluster radius s is 5, and the distinctiveness ratio constant alpha is 2, constant k _1,k ₂Be respectively 0.4,0.6, all ants all are arranged to light condition.The calculating of data similarity is carried out according to formula (1), and ant carries the data probability and puts down the calculating of data probability according to formula (2), (3).Concrete formula is as follows:

$f\left(O_i\right)=\max \{0,\frac{1}{s^2}{\mathrm{\Σ}}_{O_j,\mathrm{Neigh}(s*s)}[1-\frac{d(o_i,o_j)}{\mathrm{\α}}\left]\right\}...\left(1\right)$

$P_p\left(O_i\right)={\left(\frac{k_1}{k_1+f\left(o_i\right)}\right)}^2...\left(2\right)$

$P_d\left(O_i\right)=\left\{\begin{array}{c}2f\left(o_i\right),\mathrm{whenf}\left(o_i\right)<k_2\\ 1,f\left(o_i\right)\&GreaterEqual;k_2\end{array}\right....\left(3\right)$

Under the prerequisite that ant number and two-dimensional space size are determined, cluster process is:

(1) for i ant, random is distributed on the two-dimensional space, has data to exist on light condition and the position if ant is in, and calculates respectively similarity and carrying probability according to formula (1), (2).Generate the random number in (0, a 1) scope, if random number greater than the carrying probability then ant carrying data, data empty on the current point, ant is arranged to the carrying state.Do not have data to exist on carrying state and the current location if ant is in, then calculate according to formula (3) and put down probability, generate one (0,1) random number in the scope, if random number is greater than putting down probability then ant is put down data, store data on the current point, ant becomes light condition.

(2) ant carries out position movement, movement rule is not X ± 1 then to the border of X-axis, the words on the border that X-axis arrives are Y coordinate ± 1 then, which kind of form of taking to add and subtract depends on the generation of random number, number in random [1,1] scope that generates is if random number is for to add on the occasion of then getting, random number be negative value then for subtracting, keep mobile until do not occupied by other ant on the position.

(3) repeat (1) (2) according to iterations, obtain the cluster result under current ant number and the two-dimensional space size.

(4) according to step-size change ant number and the two-dimensional space size set, repeat (1) (2) (3), until obtain the cluster result of enough number of times.

Step 2: carry out the correction of clustering cluster for the initial clustering that obtains in the step 1 according to code book size: the centre of form of directly calculating each clustering cluster if the clustering cluster number is identical with code book size obtains code book; If the clustering cluster number how much carries out descending sort with clustering cluster according to the number of training sequence vector in affiliated bunch greater than code book size, then maximum clustering cluster and min cluster bunch are merged, replace original maximum clustering cluster, repeat union operation until the clustering cluster number is identical with the code book size of initial setting; If the clustering cluster number is less than code book size then the clustering cluster that obtains is carried out descending sort according to the number of training sequence vector in each clustering cluster, then the data in the maximum clustering cluster are divided into two according to the size to this bunch centre of form distance, generate two new clustering cluster, repeat fractured operation until clusters number is identical with the code book size of initial setting.For all take steps two operation of the cluster result under all ant numbers and the two-dimensional space size.

Step 3: the basic procedure according to genetic algorithm after the completing steps two operates, and has obtained the individuality of the population first generation this moment, sets the initial value of using parameter in the genetic algorithm: P _cBe 0.6, P _mBe that 0.5, ε is 0.02.For each individuality, the distortion summation when all vectors adopt this individuality to quantize in the calculation training sequence, i.e. formula (4), the ideal adaptation degree is the function of distortion summation, formula (5).

$\mathrm{dis}=\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{(x_i-{\hat{x}}_i)}^2...\left(4\right)$

$\mathrm{fitness}\left(i\right)=\frac{1}{{\left({\mathrm{dis}}_i\right)}^2}...\left(5\right)$

(1) carries out descending sort for all individualities according to fitness, then select and interlace operation, according to crossover probability P _cSelect that front 38 individualities directly enter the next generation in the ordering individual sequence, then carry out the individual generation work of residue of future generation: to ordering individual sequence, according to begin first and pairing last, second principle of matching with penultimate in front from the front, two individualities of pairing are used as parent, then take interlace operation to generate two new individualities, take the weighting interlace operation according to fitness in the intersection process, namely according to formula (6):

$X_1=a*X+(1-a)*\hat{X}...\left(6\right)$

$X_2=(1-a)*X+a*\hat{X}$

Wherein the calculating of a weight is carried out according to formula (7).

$a=\frac{f_A^2}{{(f_A+f_{\underset{A}{^}})}^2}...\left(7\right)$

Newly-generated individuality is placed into is used as filial generation among the next generation.Repeating crossover is until the total individual number order satisfies Population Size.

(2) all individualities are taked the gene mutation operation, according to the gene mutation probability, choose 32 individualities and carry out mutation operation, operating process is: for each individual random number that generates in (0, a 1) scope, if random number is greater than 0.5 then current individuality enters mutation operation, sudden change in taking among a small circle for the gene on the chromosome of current individuality, five different random positions of sudden change position, sampling point place [1,10], sudden change numerical value is the random number in (1,1) scope; If at random less than 0.5, then current individual the preservation do not enter mutation operation.All finish selection and mutation operation for all individualities.

(3) calculate all individual fitness according to formula (4), (5), if the fitness of optimum individual then stops iteration greater than the thresholding that changes, on the contrary repetitive process (1) (2).

(4) the output optimum individual is as the net result of code book design.

In sum, a kind of novel code book method for designing based on ant colony clustering and genetic algorithm that the present invention proposes, innovation is 2 points: utilized the LF algorithm of ant colony clustering to carry out the generation of initial clustering on the one hand, take full advantage of the final Clustering Effect of LF and initial ant set positions without the advantage of Close relation, this has just overcome traditional LBG algorithm initial selected to the impact of net result; Utilize on the other hand genetic algorithm to carry out generation and the selection of optimum individual, take full advantage of the outstanding ability of searching optimum of genetic algorithm, overcome simultaneously the shortcoming that the LBG algorithm very easily is absorbed in local optimum, can guarantee that design codebooks is close to global optimum, the gene mutation link has adopted interlace operation based on fitness according to practical operation to the line spectrum pair frequency parameter, change artificially population development direction, accelerated speed of convergence.Therefore, the present invention is on the basis of traditional ant colony clustering algorithm and genetic algorithm, fully taken into account LF algorithm possibly shortcoming that can't arrive convergence state in limiting time in the practical operation, and take full advantage of the powerful global optimum's ability of genetic algorithm, be fit to be applied to the vector quantization code book design process in early stage, for the research in the fields such as voice coding, data compression, pattern-recognition very large reference value arranged.

Claims (4)

1. novel code book method for designing based on ant colony clustering and genetic algorithm, it is characterized in that: the step of the method is as follows:

Step 1: utilize the ant of different numbers that training data is distributed on the two-dimensional space of different size randomly, take the LF algorithm to carry out the generation of initial clustering;

Step 2: the correction of the clustering cluster that initial clustering carries out according to code book size guarantees that the number of cluster is identical with the code book size of initial setting;

Step 3: under the prerequisite that initial population successfully obtains, according to the basic procedure of genetic algorithm, carry out individual choice, intersection and mutation operation, until iteration stopping is met the optimum individual of requirement.

2. a kind of novel code book method for designing based on ant colony clustering and genetic algorithm according to claim 1, it is characterized in that: the LF algorithm of taking of using in step 1 carries out the generation of initial clustering, the parameter of using in the cluster process mainly comprises ant number m, two-dimensional space size Z, cluster radius s, distinctiveness ratio constant alpha, and constant value k _1,k ₂, the carrying probability P _pPut down probability P _d

In step 1, take the LF algorithm to carry out the basic step of generation of initial clustering as follows:

(1) setting initial ant number is m, step-length is ma, the two-dimensional space size is initially Z, step-length is za, data with training sequence under current ant number and two-dimensional space are assigned on the two-dimensional space randomly, all ants all are arranged to light condition, and this moment is for the arbitrary element O in the training sequence _i, its position X _iY _iAll be to be in randomly in the Z*Z space;

(2) for each ant, on Cluster space, choose randomly a position, if being in, current ant have data to exist on light condition and the current location, calculate the similarity of all data elements in current data element and the cluster radius centered by current location, one (0 of random generation, 1) numerical value in the scope, if random number greater than the carrying probability then ant carries this data, data empty on the current location;

(3) if do not have data and ant to be in the carrying state on the current location, use the similarity that calculates in the step (2) to calculate current ant for the probability that puts down of current data, calculating is judged after finishing: generate at random one (0,1) numerical value in the scope, if random number surpasses and to put down probability then ant is put down data, ant is arranged to light condition;

(4) position of ant carries out randomly that fixed length moves, and namely X or Y coordinate plus-minus step-length repeats move operation until the current position that moves to is not occupied by other ant;

(5) repeating step (2)-(4) are until to setting the equal complete operation of ant of number, complete operation obtains the cluster result for raw data afterwards, and the code book size that the clusters number that obtain this moment may reach with setting is different;

(6) according to the ant number step-length ma that sets in the step (1) and two-dimensional space step-length za change initial parameter, repeating step (2)-(5).

3. a kind of novel code book method for designing based on ant colony clustering and genetic algorithm according to claim 1, it is characterized in that: the correction of the clustering cluster that the initial clustering that carries out in step 2 carries out according to code book size is consistent with the number that guarantees clustering cluster with code book size;

The basic step of the correction of the clustering cluster that the initial clustering that carries out in step 2 carries out according to code book size is as follows:

(1) initial parameter is set: code book size M, the clusters number K that obtains among Fig. 2;

(2) if M＜K, the cluster data that obtains is carried out descending sort according to the number of training sequence vector in each clustering cluster, then maximum clustering cluster and min cluster bunch are merged, replace original maximum clustering cluster, repeat union operation until clusters number is identical with the code book size of initial setting, i.e. M=K;

(3) if M〉K, the cluster data that obtains is carried out descending sort according to the number of training sequence vector in each clustering cluster, then the data in the maximum clustering cluster are divided into two according to the size to this clustering cluster centre of form distance, generate two new clustering cluster, repeat fractured operation until clusters number is identical with the code book size of initial setting, i.e. M=K;

(4) carry out centroid calculation for the current M that an obtains clustering cluster, the centre of form is used as code vector in the initial codebook.

4. a kind of novel code book method for designing based on ant colony clustering and genetic algorithm according to claim 1, it is characterized in that: the basic procedure of the genetic algorithm of in step 3, using, adopt genetic algorithm will before the initial population that obtain select, the step of intersection and gene mutation operation, the parameter of wherein using comprises Population Size N _{Pop_size}, mutation probability P _m, crossover probability P _cWith iteration stopping thresholding ε;

Wherein, the basic procedure of the genetic algorithm that adopts in the step 3, its concrete step is as follows:

(1) for all individuality calculating fitness value separately, then individuality is carried out descending sort according to fitness, judge that whether the fitness of optimum individual is greater than the iteration stopping thresholding, if greater than the iteration stopping thresholding iteration work finish, enter step (4), on the contrary the step of entering (2);

(2) according to crossover probability P _cP before selecting in the ordering individual sequence _c* N _{Pop_size}Individuality directly enters the next generation, then carry out the individual generation work of residue of future generation: to ordering individual sequence, according to begin first and pairing last, second principle of matching with penultimate in front from the front, two individualities of pairing are used as parent, then take the single-point interlace operation to generate two new individualities, be placed into and be used as filial generation among the next generation, repeating crossover is until the total individual number order satisfies Population Size;

(3) according to the gene mutation probability P _mChoose P _m* N _{Pop_size}Individuality carries out mutation operation, and operating process is: for each individual random number that generates in (0, a 1) scope, if random number is greater than P _mThen current individuality enters mutation operation, takes interior among a small circle sudden change for the gene on the chromosome of current individuality, and sudden change numerical value is the function of the fitness of individuality; If at random less than P _mThen current individual the preservation do not enter mutation operation, all finishes for all individualities and selects and mutation operation, enters step (1);

(4) individuality of output fitness maximum is as the result of final code book design.

Images