CN103279793A - Task allocation method for formation of unmanned aerial vehicles in certain environment - Google Patents

Abstract

The invention discloses a task allocation method for formation of unmanned aerial vehicles in a certain environment, belonging to the technical field of unmanned aerial vehicles. The task allocation method comprises the following steps of determining a coding sequence of a task allocation algorithm; determining a preponderant function of the unmanned aerial vehicles formed to execute a task; determining a speed update formula and a position update formula of a discrete particle swarm optimization; determining the flow of a tabu search; and determining the flow of hybrid optimization. According to the task allocation method for the formation of the unmanned aerial vehicles in the certain environment, the continuous particle swarm optimization is discretized, the algorithm is simply and conveniently operated on the premise that optimizing property can be guaranteed, and the effectiveness of the discrete particle swarm method is indicated through simulation. According to the task allocation method for the formation of the unmanned aerial vehicles in the certain environment, a supplement strategy of the tabu search algorithm is provided, and the local optimizing capacity of the algorithm is enhanced when the inertia weight [omega] of the particle swarm optimization is larger, i.e. the particle swarm embodies stronger variety, so that the original two algorithms realize complementing each other's advantages, the searching performance can be improved, and the judgment can be verified in multiple groups of simulated tests.

Description

Unmanned vehicle formation method for allocating tasks under a kind of definite environment

Technical field

The invention belongs to the unmanned vehicle technical field, relate to the flight formation task allocation technique, specifically, refer to the unmanned vehicle formation method for allocating tasks under a kind of definite environment.

Background technology

Current had nearly that more than 30 countries drop into research and the productions that a large amount of manpower and financial resources are engaged in unmanned plane.Through vicennial development, this technology is comparative maturity, bringing into play in the army and the people's every field and to act on, however, single frame unmanned plane exists some problems when carrying out task, for example single frame unmanned plane may be subjected to the restricted number of sensor, can not observe the target area from multi-angle is omnibearing, when facing the wide area search task, can not effectively cover whole region of search; If what carry out is rescue task, single frame unmanned plane is restricted aspect load, often influences the usefulness of whole rescue, brings bigger loss, and in addition, in case single frame unmanned plane breaks down, interrupt task is returned immediately, may incur loss through delay rescue opportunity.

In recent years because the complicacy of environment and the multiplicity of task, single frame unmanned plane can't satisfy the needs of attacking over the ground with rescue, above-mentioned shortcoming at single frame unmanned plane, abroad proposed in recent years multiple no-manned plane carry out jointly task concept and obtained certain achievement in research, carry out the jointly mode of task of multiple no-manned plane can significantly improve operation and the rescue usefulness of unmanned plane integral body, can increase the success ratio of executing the task and the ability of anti-accident, based on above-mentioned advantage, the multiple no-manned plane aerial mission that carries out jointly will become an important directions of the development of unmanned plane from now on.Wherein multiple no-manned plane cotasking assignment problem is carry out the jointly key issue of task of multiple no-manned plane.

This is a multiple constraint to the distribution of multiple no-manned plane cotasking in essence, cross coupling complicated multiple goal integer optimization problem, its obvious characteristic is that the value of decision variable disperses, if there is noninferior solution in a multi-objective optimization question, often there are a lot of, form noninferior set, the decision maker generally selects one or several satisfied non-bad optimum solution to separate as final, so generally there is not single optimum solution in multi-objective optimization question, but optimum collection of Pareto, must between the optimum solution of the optimum collection of Pareto, weigh compromise, and then obtain to make the optimum solution of decision maker's satisfaction.

The concept that the multiple no-manned plane cotasking distributes is: before multiple no-manned plane carries out task, based on certain environment knowledge and mission requirements, quantity or load according to unmanned plane, for each unmanned plane distributes one or one group sequence task (target or locus point) is arranged, so that in the task of finishing most probable number simultaneously, the usefulness of aircraft integral body reaches optimum.Because the decision variable of multiple goal integer optimization problem disperses, therefore the optimum theory in the classical mathematics continuously and basic skills generally can't be directly used in and find the solution the integer optimization problem, have formed following three kinds of main solutions through the development of decades:

First kind: the solution of supposing discrete case is limited, so compares after can finding out all feasible solutions again, is referred to as the method for exhaustion or enumerative technique, and this method generally can only be used for problem on a small scale;

Second kind: namely ignore the integer requirement earlier, find the solution according to continuous situation, then solution is carried out integer and handle.This method comprises branch and bound method and cutting-plane method.The calculated amount of branch and bound method and the integer programming of cutting-plane method solution is less than enumerative technique, but the problem that its requirement is found the solution is linearity, and its computation complexity is exponential increase with problem scale, therefore also is rarely used in extensive problem;

The third: bionic intelligence calculates, with natural biotic population feature or phenomenon abstract be mathematic(al) representation, the biosystem of simulating nature circle, rely on biosome self instinct fully, optimize the individual needs (environmental suitability generally by fitness function represented) of survival condition to conform by unconscious optimizing behavior.The bionic intelligence computing method have (1) probability global optimization; (2) do not rely on the strict mathematical property of optimization problem; (3) essential concurrency; (4) characteristics such as self-organization and evolution.

Summary of the invention

The objective of the invention is to solve problems of the prior art, unmanned vehicle formation task under a kind of definite environment allocation strategy is provided, distribute related large-scale complex multiple goal integer optimization problem at task, designed a kind of hybrid optimization strategy based on the modern optimization algorithm, accelerate modern optimization convergence of algorithm speed and local search ability, the present invention has designed the particle of discrete particle cluster algorithm and has represented more new formula of mode and velocity location, discrete particle cluster algorithm is combined with tabu search algorithm, carry out coarse search with discrete particle cluster algorithm, carry out the essence search with tabu search algorithm, the advantage separately of discrete particle cluster algorithm and tabu search algorithm is carried out comprehensively, finished the unmanned vehicle formation task allocating task under definite environment.

Unmanned vehicle formation task allocation strategy under described definite environment comprises following step:

Step 1: the coded sequence of the allocation algorithm that sets the tasks;

Step 2: determine the advantage function that unmanned plane is formed into columns and executed the task;

Step 3: the speed of determining discrete particle cluster algorithm is new formula and position new formula more more;

Step 4: the flow process of determining tabu search;

Step 5: the flow process of determining hybrid optimization.

The invention has the advantages that:

(1) with continuous particle cluster algorithm discretize.Traditional continuous particle cluster algorithm other groups algorithm relatively has simple to operate, be convenient to understand, advantages such as fast convergence rate, shortcoming is that this algorithm is directed to continuous problem mostly, shortcoming at particle cluster algorithm, the present invention proposes a kind of discrete particle cluster method, with continuous particle cluster algorithm discretize, more new formula of the speed of discrete domain particle cluster algorithm and position has been proposed, under the prerequisite that guarantees its optimizing performance, make that simultaneously algorithm is simple, easy to operate, and shown the validity of this discrete particle cluster method by emulation.

(2) great deal of research results shows, discrete particle cluster algorithm is compared with other algorithms, the shared mechanism of information is different, be example with the genetic algorithm, in genetic algorithm, chromosomal information is shared after by the roulette method, that is to say that the speed that whole population is moved to optimal region is comparatively uniform, the optimal particle of particle cluster algorithm all can influence all particles in the population in for evolutionary process at each, so the easy precocity of algorithm is absorbed in local optimum.At this shortcoming, the present invention proposes the replenishment strategy of tabu search algorithm, be more greatly that population embodies the stronger multifarious moment at particle cluster algorithm inertia weight ω, the local optimal searching ability that adds strong algorithms, make original two kinds of algorithms realize mutual supplement with each other's advantages, the search performance that improves, and in several groups of l-G simulation tests, verified above-mentioned judgement.

Description of drawings

Fig. 1 is task advantage table;

Fig. 2 is the coded system synoptic diagram of task allocation algorithms;

Fig. 3 is S ⁱ(t) in the vector With Element exchanges synoptic diagram;

Fig. 4 is S ^{I '}(t) element and P between [a, b] in the vector ⁱ(t) corresponding element exchanges synoptic diagram in the vector;

Fig. 5 is S ^{I "}(t) element and P between [a, b] in the vector _g(t) corresponding element exchanges synoptic diagram in the vector;

Fig. 6 is the process flow diagram of discrete particle cluster algorithm;

Fig. 7 is the coded system synoptic diagram of tabu search structure neighborhood solution;

Fig. 8 is the tabu search process flow diagram;

Fig. 9 is unmanned vehicle formation task hybrid optimization distribution method process flow diagram provided by the invention;

Figure 10 is the formation overall advantage function curve that unmanned vehicle formation task is distributed;

Figure 11 is the formation overall advantage function simulation curve that unmanned vehicle formation task is distributed;

Figure 12 is the carry out advantage function tables of 10 tasks of 10 unmanned planes;

The formation overall advantage function curve that Figure 13 distributes for the unmanned vehicle formation task of carrying out three kinds of policing types among the embodiment;

Figure 14 is for to carry out the curve that Monte Carlo simulation obtains to three kinds of strategies among Figure 13.

Embodiment

The present invention is described in detail below in conjunction with drawings and Examples.

At first problem to be solved is described, task advantage table as shown in Figure 1, how many abilities normalization digitized representation unmanned plane (also claiming unmanned vehicle) has and carries out this task in the table, wherein the numeric representation unmanned plane i of the capable j of i row how many abilities j(that executes the task is arranged is target j in the task advantage table), be defined as C _{I, j}, i=1 among Fig. 1,2 ..., 20; J=1,2 ..., 10.The purpose that unmanned vehicle formation task is distributed is the combination of determining that unmanned plane is executed the task, makes the execution realizes maximal efficiency that unmanned plane is formed into columns and executed the task.

The invention provides the unmanned vehicle formation method for allocating tasks under a kind of definite environment, comprise the steps:

Step 1: according to the set the tasks coded sequence of allocation algorithm of task advantage table;

The coded sequence of task allocation algorithms as shown in Figure 2, coded system adopts the integer coding sequence, coded sequence S length equals unmanned plane sum N _S, its element s _i, 1≤i≤N _sExpression, each element s _iIn positive random number represent the mission number that this unmanned plane is carried out, suppose that task quantity is N _T, 1≤s then _i≤ N _TAccording to the data in the task advantage table of Fig. 1, with unmanned plane sum N _S=20, task quantity N _T=10 is example, and the coded sequence array has 20 elements among Fig. 2, so the coded sequence of Fig. 2 represents that unmanned plane 1 executes the task 2, unmanned plane 2 executes the task 1, and unmanned plane 3 executes the task 10, and unmanned plane 4 executes the task 3 ... the rest may be inferred, is that unmanned plane 20 executes the task 7 at last.

Step 2: determine the advantage function that unmanned plane is formed into columns and executed the task;

The advantage function that the unmanned plane formation is executed the task is determined jointly by task advantage table and coded sequence, establishes i element s in the coded sequence _iThe positive integer of middle storage is j, i.e. s _i=j supposes to exist target sequence T, and the element number among the T is N _T, its element t _i(i=1,2 ..., N _T) expression, then the computing formula of advantage function C is:

$C=\underset{i=1}{\overset{N_T}{\mathrm{\Σ}}}{t}_i---\left(1\right)$

Wherein,

$t_i=\left\{\begin{array}{cc}\underset{j=1}{\overset{N_S}{\mathrm{\&Sigma;}}}{C}_{j,i}& \mathrm{if}(s_j=i,\underset{j=1}{\overset{N_S}{\mathrm{\&Sigma;}}}{C}_{j,i}<1)\\ 1& \mathrm{if}(s_j=i,\underset{j=1}{\overset{N_S}{\mathrm{\&Sigma;}}}{C}_{j,i}\&GreaterEqual;1)\end{array}\right.---\left(2\right)$

t _i=1 expression unmanned plane is formed into columns and has been carried out task t fully _iSo the advantage function is taken as 1 also no longer to be increased.

Step 3: the speed of determining discrete particle cluster algorithm is new formula and position new formula more more;

Being described below of standard particle group algorithm: be located in the target search space of M dimension a colony that is formed by n particle.Wherein, the position of particle i is expressed as

Namely i particle is S in the position in the target search space of M dimension ⁱEach particle position S ⁱBe exactly a potential feasible solution, can obtain position S to the computing formula (1) of its substitution advantage function ⁱAdvantage functional value C (S ⁱ), be used for passing judgment on the good and bad degree of particle i.The flying speed of particle i response is expressed as

The desired positions that particle i experiences is designated as the individual extreme value of particle i, is expressed as

The desired positions that all particles of whole colony live through is designated as global extremum, is expressed as P _g=(p _G1, p _G2..., p _GM).Each particle upgrades self speed and position according to individual extreme value and global extremum, is shown below:

$\left\{\begin{array}{c}{v}_j^i(t+1)=\mathrm{\&omega;}\&CenterDot;v_j^i\left(t\right)+{\mathrm{<figure-callout\; id="1"\; label="c"\; filenames="HDA00003100894900021.png,HDA00003100894900022.png"\; state="\{\{state\}\}">c</figure-callout>}}_1\&CenterDot;\mathrm{<figure-callout\; id="1"\; label="rand"\; filenames="HDA00003100894900021.png,HDA00003100894900022.png"\; state="\{\{state\}\}">rand</figure-callout>}1\&CenterDot;(p_j^i\left(t\right)-s_j^i\left(t\right))+{\mathrm{<figure-callout\; id="2"\; label="c"\; filenames="HDA00003100894900021.png,HDA00003100894900022.png"\; state="\{\{state\}\}">c</figure-callout>}}_2\&CenterDot;\mathrm{<figure-callout\; id="2"\; label="rand"\; filenames="HDA00003100894900021.png,HDA00003100894900022.png"\; state="\{\{state\}\}">rand</figure-callout>}2\&CenterDot;(p_{\mathrm{gj}}\left(t\right)-s_j^i\left(t\right))\\ s_j^i(t+1)=s_j^i\left(t\right)+v_j^i(t+1)\end{array}\right.,j=\mathrm{1,2},...,M---\left(3\right)$

T is evolutionary generation in the formula, and ω is inertia weight, the size of inertia weight has determined the particle present speed is inherited what, suitable selection inertia weight can make particle have balanced exploring ability and development ability; c ₁, c ₂Be positive constant, be called the study factor, the study factor makes self to be had to the close ability of the historical optimum point of self historical optimum point and colony, make particle can the oneself sum up and in the colony excellent individual learn c ₁, c ₂Be generally two random numbers that are evenly distributed between [0,1].This shows standard particle group algorithm towards problem be continuous, is not suitable for and solves the dispersed problem that integer is optimized, so the present invention proposes a kind of discrete particle cluster algorithm, the process flow diagram of discrete particle cluster algorithm as shown in Figure 6.

The essence of discrete particle cluster algorithm is that particle constantly adjusts position and speed according to own and companion's flying experience, thereby to optimal location flight, the reposition of particle is particle rapidity, individual extreme value and global extremum results of interaction.Therefore to the position of standard particle group algorithm and speed more new formula redefine, obtain in the discrete particle cluster algorithm i particle at the position S in the target search space that M ties up ⁱ:

S ⁱ(t+1)＝f ₃(c ₂,P _g(t),f ₂(c ₁,P ⁱ(t),f ₁(ω,S ⁱ(t))))(4)

S ⁱ(t+1) for evolutionary generation be the position of the particle i behind the t+1, S ⁱ(t) be that particle i behind the t is at the position in M dimension target search space, P for evolutionary generation _g(t) be the global extremum of population behind the evolutionary generation t, P ⁱ(t) be the individual extreme value of particle i behind the evolutionary generation t, f ₁, f ₂, f ₃Represent that respectively inertia weight upgrades operator, discrete particle cluster algorithm autognosis operator and social recognition operator.c ₁, c ₂Be the study factor, ω is inertia weight.

Initialization ω, c ₁, c ₂, calculate weight decline coefficient

Initialization n primary S ⁱ(t); Calculate the advantage function P of population according to formula (1), formula (2) ⁱ(t)=S ⁱ(t), ask P _g(t);

Definition S ^{I '}, S ^{I "}Be intermediate variable, then:

$S^{i^{\&prime;}}\left(t\right)=f_1(\mathrm{\&omega;},S^i\left(t\right))=\left\{\begin{array}{c}{g}_1\left(S^i\left(t\right)\right),\mathrm{rand}\left(\right)\&le;\mathrm{\&omega;}\\ S^i\left(t\right),\mathrm{rand}\left(\right)>\mathrm{\&omega;}\end{array}\right.---\left(5\right)$

This formula is discrete particle cluster algorithm inertia part, specifically describes as followsly, generates equally distributed random number between one [0,1] by rand () function, is designated as r ₁If this random number is greater than ω, particle position S then ^{I '}(t)=S ⁱ(t) remain unchanged, if this random number smaller or equal to ω, S then ^{I '}(t)=g ₁(S ⁱ(t)), g wherein ₁(S ⁱ(t)) expression produces two [1, N _S] between different random count a and b, then with S ⁱ(t) element of a and b correspondence in the vector

With

Exchange, as shown in Figure 3, random number is respectively 7 and 14, and corresponding element is respectively 4 and 9, exchanges 4 and 9.

For intermediate variable S ^{I "}, after evolutionary generation is t, have:

$S^{i^{\&prime;\&prime;}}\left(t\right)=f_2({\mathrm{<figure-callout\; id="1"\; label="c"\; filenames="HDA00003100894900021.png,HDA00003100894900022.png"\; state="\{\{state\}\}">c</figure-callout>}}_1,P^i\left(t\right),S^{i^{\&prime;}}\left(t\right))=\left\{\begin{array}{c}{g}_2(P^i\left(t\right),S^{i^{\&prime;}}\left(t\right)),\mathrm{rand}\left(\right)\&le;c_1\\ S^{i^{\&prime;}}\left(t\right),\mathrm{rand}\left(\right)>c_1\end{array}\right.---\left(6\right)$

This formula (6) is discrete particle cluster algorithm autognosis part, specifically describes as followsly, generates equally distributed random number between one [0,1] by rand () function, is designated as r ₂If this random number is greater than c ₁, the position S after t the iteration of particle i then ^{I "}(t)=S ^{I '}(t) remain unchanged, if this random number is smaller or equal to c ₁, S then ^{I "}(t)=g ₂(P ⁱ(t), S ^{I '}(t)), g wherein ₂(P ⁱ(t), S ^{I '}(t)) expression produces two [1, N _S] between different random count c and d, then with S ^{I '}(t) element and P between [c, d] in the vector ⁱ(t) corresponding element exchanges in the vector, as shown in Figure 4.

Based on formula (5) and formula (6), obtaining evolutionary generation is the position S of the particle i behind the t+1 ⁱ(t+1) be:

$S^i(t+1)=f_3({\mathrm{<figure-callout\; id="2"\; label="c"\; filenames="HDA00003100894900021.png,HDA00003100894900022.png"\; state="\{\{state\}\}">c</figure-callout>}}_2,P_g\left(t\right),S^{i^{\&prime;\&prime;}}\left(t\right))=\left\{\begin{array}{c}{g}_3(P_g\left(t\right),S^{i^{\&prime;\&prime;}}\left(t\right)),\mathrm{rand}\left(\right)\&le;c_2\\ S^{i^{\&prime;\&prime;}}\left(t\right),\mathrm{rand}\left(\right)>c_2\end{array}\right.---\left(7\right)$

This formula is discrete particle cluster algorithm social recognition part, specifically describes as followsly, generates equally distributed random number between one [0,1] by rand () function, is designated as r ₃If this random number is greater than c ₂, particle position S then ⁱ(t+1)=S ^{I "}(t) remain unchanged, if this random number is smaller or equal to c ₂, S then ⁱ(t+1)=g ₃(P _g(t), S ^{I "}(t)), g wherein ₃(P _g(t), S ^{I "}(t)) expression produces two [1, N _S] between different random count A and B, then with S ^{I "}(t) element and P between [A, B] in the vector _g(t) corresponding element exchanges in the vector, as shown in Figure 5.

Upgrade individual extreme value and global extremum, if S ⁱ(t+1) advantage function is greater than P ⁱ(t) advantage function, then P ⁱ(t+1)=S ⁱOtherwise P (t+1), ⁱ(t+1)=P ⁱ(t); If S ⁱ(t+1) advantage function is greater than P _g(t) advantage function, then P _g(t+1)=S ⁱOtherwise P (t+1), _g(t+1)=P _g(t).

Inertia weight ω initial value ω among the present invention _StartBe taken as 0.9, increase end value ω with evolutionary generation _EndDrop to ω _End=0.4, for example evolutionary generation n is 100, and then t is for inertia weight ω value ω _tBe ω _t=ω _End+ (ω _Start-ω _End) (t/n), weight decline coefficient lambda _ω=(ω _Start-ω _End) (t/n), similar c ₁, c ₂Initial value c _1start, c _2startAll be taken as 0.6, increase end value c with evolutionary generation _1end, c _2endBe 0.2, the t for c ₁Value c _1tBe c _1t=c _1end+ (c _1start-c _1end) (t/n), weight decline coefficient

T is for c ₂Value c _2tBe c _2t=c _2end+ (c _2start-c _2end) (t/n), weight decline coefficient ${\mathrm{\&lambda;}}_{{\mathrm{<figure-callout\; id="2"\; label="c"\; filenames="HDA00003100894900021.png,HDA00003100894900022.png"\; state="\{\{state\}\}">c</figure-callout>}}_2}=({\mathrm{<figure-callout\; id="2"\; label="c"\; filenames="HDA00003100894900021.png,HDA00003100894900022.png"\; state="\{\{state\}\}">c</figure-callout>}}_{2\mathrm{start}}-{\mathrm{<figure-callout\; id="2"\; label="c"\; filenames="HDA00003100894900021.png,HDA00003100894900022.png"\; state="\{\{state\}\}">c</figure-callout>}}_{2\mathrm{end}})\&CenterDot;(t/n).$

Step 4: the flow process of determining tabu search;

Tabu search algorithm has fast convergence rate, stronger advantages such as local search ability, but that the last convergence result of tabu search algorithm is influenced by initial solution is bigger, and the bad tabu search algorithm that very easily makes of the selection of initial solution is absorbed in local optimum.Tabu search algorithm is a kind of embodiment of artificial intelligence, is a kind of expansion of local neighborhood searching algorithm.The basic thought of tabu search algorithm is: a given current solution and a neighborhood, in the neighborhood of current solution, determine some candidate solutions then.If the advantage function of optimal candidate solution correspondence is better than " best so far " state, then ignore its taboo characteristic, substitute current solution and " best so far " state with it, and corresponding object is added the taboo table, revise the term of office of each object in the taboo table simultaneously; If there is not an above-mentioned candidate solution, then selecting the optimum condition of non-taboo in candidate solution is new current solution, and ignores it and the quality of current solution, simultaneously response object is added the taboo table, and revise taboo show in term of office of each object.So repeat said process, until satisfying stop condition.

The tabu search flow process that the present invention adopts as shown in Figure 8, concrete steps are as follows:

(4.1) the tabu search algorithm parameter is set, initialization taboo table.

The taboo table is the formation of a first in first out, is defined as L _Tabu, queue length is 10, if coded sequence continues algebraically n in the taboo table _Tabu＞5, with the deletion from the taboo table of this coded sequence, this coded sequence is forgotten in expression.Produce initial solution S at random ⁱ(t), t=0, current solution

(4.2) at initial solution S ⁱ(t) produce some not neighborhood solutions in the taboo table near, neighborhood solution number is n _S=10, the method for structure neighborhood solution divide following two kinds as shown in Figure 7, first kind is to produce two [1, N _S] between different random count x and y, then with S ⁱ(t) element corresponding with random number x and y in the vector

With

Exchange; Second kind is to produce [1 a, N _T] between random number z and [1, N _S] between random number u, replace the element of u position with z, ask the solution of advantage function maximum in the set of neighborhood solution

(4.3) if The advantage function Greater than S ⁱ(t) advantage function C (S ⁱ(t)), then order

Will

Add the taboo table, the duration of upgrading the taboo table, will continue algebraically n _Tabu＞5 coded sequence deletion; If

The advantage function

Smaller or equal to S ⁱ(t) advantage function C (S ⁱ(t)), then keep

Constant, will

Add the taboo table, the duration of upgrading the taboo table, will continue algebraically n _Tabu＞5 coded sequence deletion.

(4.4) judge whether iteration is stipulated algebraically to tabu search algorithm, optimize the result if then finish tabu search algorithm output, otherwise transfer step (4.2) to;

By the description of above-mentioned tabu search flow process neighborhood function, taboo table and continue the key that algebraically is tabu search as can be seen, wherein, the neighborhood function has been continued to use the thought of Local Search; The taboo table has embodied tabu search algorithm and has avoided circuitous shortcoming of searching for; Lasting algebraically is to avoiding a kind of the loosening of strategy.

Step 5: the flow process of determining hybrid optimization;

Discrete particle cluster algorithm has simple to operate being easy to be realized, the advantage that ability of searching optimum is strong, but discrete particle cluster algorithm late convergence and solving precision are not high.Discrete particle cluster algorithm relies on the cooperation and competition between the initial random colony to instruct optimizing, in case certain particle is found a current optimal location, this position can be passed to other particles with discrete particle cluster algorithm social recognition form partly, make other particles rapidly near this optimal location, if this optimal location is local optimum, then other particles can't be searched for other zones in the population, cause discrete particle cluster algorithm to be absorbed in local optimum; Secondly owing to there is inertia weight, make that each particle probably misses optimum solution with certain coasting flight in the colony in flight course.

And for the tabu search algorithm in the step 4, also have the following disadvantages:

(1) initial solution there is stronger dependence, the solution that good initial solution can make tabu search search in solution space, the initial solution of difference then can reduce the speed of convergence of tabu search;

(2) iterative search is serial, only is the movement of a single state, but not parallel search.

For overcoming the shortcoming separately of above-mentioned discrete particle cluster algorithm and tabu search algorithm, introduce tabu search algorithm on the basis of discrete particle cluster algorithm.Discrete particle cluster algorithm is combined with tabu search algorithm, take full advantage of the global optimization ability of discrete particle cluster algorithm and the local search ability of tabu search algorithm, make two kinds of algorithms realize having complementary advantages, obtain than the better effect of single algorithm.

As shown in Figure 9, the overall framework of unmanned vehicle formation task hybrid optimization allocation strategy still is main frame with the discrete particle cluster algorithm, after difference is that the discrete particle cluster algorithm iteration is once, not the next iteration process that directly enters, but to the global extremum P of population _g(t) Xiang Liang certain neighborhood carries out local tabu search, if at P _g(t) find better global optimum position (global extremum) in Xiang Liang certain neighborhood, then replace P with this position vector _g(t) if vector is at P _g(t) do not find better global optimum position in Xiang Liang certain neighborhood, then not to P _g(t) vector upgrades, and that is to say after the iteration each time of discrete particle cluster algorithm has increased by one to global extremum P _g(t) tabu search of vector is operated, and increases the local search ability of algorithm, and concrete steps are as follows

(5.1) the tabu search algorithm parameter is set, initialization taboo table.

The taboo table is the formation of a first in first out, is defined as L _Tabu, queue length is 10, if coded sequence continues algebraically n in the taboo table _Tabu＞5, with the deletion from the taboo table of this coded sequence, this coded sequence is forgotten in expression, produces initial solution P at random _g(t), t=0, current solution P _Gcur(0)=P _g(0);

(5.2) at initial solution P _g(t) produce some not neighborhood solutions in the taboo table near, neighborhood solution number is n _S=10, the method for structure neighborhood solution is divided consistent with Fig. 7, asks the solution P of advantage function maximum in the set of neighborhood solution _Gbest(t);

(5.3) if P _Gbest(t) advantage function C (P _Gbest(t)) greater than P _Gcur(t) advantage function C (P _Gcur(t)), then make P _Gcur(t)=P _Gbest(t), with P _Gcur(t) add the taboo table, the duration of upgrading the taboo table, will continue algebraically n _Tabu＞5 coded sequence deletion; If P _Gbest(t) advantage function C (P _Gbest(t)) smaller or equal to P _Gcur(t) advantage function C (P _Gcur(t)), then keep P _Gcur(t) constant, with P _Gbest(t) add the taboo table, the duration of upgrading the taboo table, will continue algebraically n _Tabu＞5 coded sequence deletion.

(5.4) judge whether iteration is stipulated algebraically to tabu search algorithm, optimize P as a result if then finish tabu search algorithm output _GcurOtherwise transfer step (5.2) to (t).

Figure 10 has provided and has utilized discrete particle cluster algorithm to carry out the distribution of unmanned vehicle formation task separately, utilizing tabu search algorithm to carry out unmanned vehicle formation task separately distributes and utilizes the mixed strategy of discrete particle cluster algorithm and tabu search algorithm to carry out the formation overall advantage function curve that unmanned vehicle formation task is distributed, the advantage function table that task is distributed as shown in Figure 1, the mixed strategy of utilizing discrete particle cluster algorithm and tabu search algorithm as can be seen just can converge on the optimum solution after 10 generations of evolving, and still was all to be better than preceding two kinds of methods qualitatively in the optimum solution of acquisition in speed of convergence.Figure 11 has carried out Monte Carlo simulation to above three kinds of strategies, simulation times is 50 times, each method evolutionary generation was 50 generations, and the quality of as seen utilizing the mixed strategy of discrete particle cluster algorithm and tabu search algorithm to separate is the quality that is better than utilizing separately solution that this method obtains.The optimal solution vector that obtains is [7,6,2,6,4,8,7,5,10,1,3,9,5,2,9,1,6,3,4,8].

In order to verify the versatility of this mixed strategy, Figure 12 provides the carry out advantage function table of 10 tasks of 10 unmanned planes, Figure 13 has provided and has utilized discrete particle cluster algorithm to carry out the distribution of unmanned vehicle formation task separately, utilizing tabu search algorithm to carry out unmanned vehicle formation task separately distributes and utilizes the mixed strategy of discrete particle cluster algorithm and tabu search algorithm to carry out the formation overall advantage function curve that unmanned vehicle formation task is distributed, the mixed strategy of utilizing discrete particle cluster algorithm and tabu search algorithm as can be seen just can converge on the optimum solution after 5 generations of evolving, and still was all to be better than preceding two kinds of methods qualitatively in the optimum solution of acquisition in speed of convergence.Figure 14 has carried out Monte Carlo simulation to above three kinds of strategies, simulation times is 50 times, each method evolutionary generation was 50 generations, and the quality of as seen utilizing the mixed strategy of discrete particle cluster algorithm and tabu search algorithm to separate is the quality that is better than utilizing separately solution that this method obtains.The optimal solution vector that obtains is [9,3,8,10,3,7,6,7,10,6].

Claims (6)

1. the unmanned vehicle formation method for allocating tasks under the definite environment is characterized in that, may further comprise the steps:

Step 1: the coded sequence of the distribution method that sets the tasks;

Step 2: determine the advantage function that unmanned plane is formed into columns and executed the task;

Step 3: the speed of determining discrete particle cluster algorithm is new formula and position new formula more more, obtains in the discrete particle cluster algorithm i particle at the position S in the target search space of M dimension ⁱ

Step 4: the flow process of determining tabu search;

Step 5: the flow process of determining hybrid optimization.

2. the unmanned vehicle formation method for allocating tasks under a kind of definite environment according to claim 1 is characterized in that: described coded sequence adopts the integer coding sequence, and according to task advantage table, coded sequence S length equals unmanned plane sum N _S, its element s _i, 1≤i≤N _sExpression, each element s _iIn positive random number represent the mission number that this unmanned plane is carried out, suppose that task quantity is N _T, 1≤s then _i≤ N _T

3. the unmanned vehicle formation method for allocating tasks under a kind of definite environment according to claim 1, it is characterized in that: described advantage function is determined jointly by task advantage table and coded sequence, establishes i element s in the coded sequence _iThe positive integer of middle storage is j, i.e. s _i=j supposes to exist target sequence T, and the element number among the T is N _T, its element t _iExpression, then the computing formula of advantage function C is:

$C=\underset{i=1}{\overset{N_T}{\mathrm{\&Sigma;}}}{t}_i---\left(1\right)$

Wherein,

$t_i=\left\{\begin{array}{cc}\underset{j=1}{\overset{N_S}{\mathrm{\&Sigma;}}}{C}_{j,i}& \mathrm{if}(s_j=i,\underset{j=1}{\overset{N_S}{\mathrm{\&Sigma;}}}{C}_{j,i}<1)\\ 1& \mathrm{if}(s_j=i,\underset{j=1}{\overset{N_S}{\mathrm{\&Sigma;}}}{C}_{j,i}\&GreaterEqual;1)\end{array}\right.---\left(2\right)$

t _i=1 expression unmanned plane is formed into columns and has been carried out task t fully _iSo the advantage function is taken as 1 also no longer to be increased, i=1,2 ..., N _T

4. the unmanned vehicle formation method for allocating tasks under a kind of definite environment according to claim 1 is characterized in that: i particle is at the position S in the target search space of M dimension in the described discrete particle cluster algorithm ⁱ:

S ⁱ(t+1)＝f ₃(c ₂,P _g(t),f ₂(c ₁,P ⁱ(t),f ₁(ω,S ⁱ(t))))(4)

S ⁱ(t+1) for evolutionary generation be the position of the particle i behind the t+1; S ⁱ(t) be that particle i behind the t is in the position in M dimension target search space for evolutionary generation; P _g(t) be the global extremum of population behind the evolutionary generation t; P ⁱ(t) be the individual extreme value of particle i behind the evolutionary generation t; f ₁, f ₂, f ₃Represent that respectively inertia weight upgrades operator, discrete particle cluster algorithm autognosis operator and social recognition operator; c ₁, c ₂Be the study factor, ω is inertia weight.

5. the unmanned vehicle formation method for allocating tasks under a kind of definite environment according to claim 1, it is characterized in that: the flow process of described tabu search is:

(4.1) the tabu search algorithm parameter is set, initialization taboo table;

The taboo table is the formation of a first in first out, is defined as L _Tabu, if coded sequence continues algebraically n in the taboo table _Tabu＞5, with the deletion from the taboo table of this coded sequence, this coded sequence is forgotten in expression, produces initial solution S at random ⁱ(t), t=0, current solution $S_{\mathrm{cur}}^i\left(0\right)=S^i\left(0\right);$

(4.2) at initial solution S ⁱ(t) produce some not neighborhood solutions in the taboo table near, neighborhood solution number is n _S=10, the method for structure neighborhood solution is divided following two kinds, and first kind is to produce two [1, N _S] between different random count x and y, then with S ⁱ(t) element corresponding with random number x and y in the vector

With

Exchange; Second kind is to produce [1 a, N _T] between random number z and [1, N _S] between random number u, replace the element of u position with z, ask the solution of advantage function maximum in the set of neighborhood solution

(4.3) if

The advantage function

Greater than S ⁱ(t) advantage function C (S ⁱ(t)), then order

Will

Add the taboo table, the duration of upgrading the taboo table, will continue algebraically n _Tabu＞5 coded sequence deletion; If

The advantage function

Smaller or equal to S ⁱ(t) advantage function C (S ⁱ(t)), then keep

Constant, will Add the taboo table, the duration of upgrading the taboo table, will continue algebraically n _Tabu＞5 coded sequence deletion;

(4.4) judge whether iteration is stipulated algebraically to tabu search algorithm, optimize the result if then finish tabu search algorithm output, otherwise transfer step (4.2) to.

6. the unmanned vehicle formation method for allocating tasks under a kind of definite environment according to claim 1 is characterized in that: the flow process of described hybrid optimization, and concrete steps are as follows:

(5.1) the tabu search algorithm parameter is set, initialization taboo table;

The taboo table is the formation of a first in first out, is defined as L _Tabu, queue length is 10, if coded sequence continues algebraically n in the taboo table _Tabu＞5, with the deletion from the taboo table of this coded sequence, this coded sequence is forgotten in expression, produces initial solution P at random _g(t), t=0, current solution P _Gcur(0)=P _g(0);

(5.2) at initial solution P _g(t) produce some not neighborhood solutions in the taboo table near, neighborhood solution number is n _S=10, construct the neighborhood solution, ask the solution P of advantage function maximum in the set of neighborhood solution _Gbest(t);

(5.3) if P _Gbest(t) advantage function C (P _Gbest(t)) greater than P _Gcur(t) advantage function C (P _Gcur(t)), then make P _Gcur(t)=P _Gbets(t), with P _Gcur(t) add the taboo table, the duration of upgrading the taboo table, will continue algebraically n _Tabu＞5 coded sequence deletion; If P _Gbest(t) advantage function C (P _Gbest(t)) smaller or equal to P _Gcur(t) advantage function C (P _Gcur(t)), then keep P _Gcur(t) constant, with P _Gbest(t) add the taboo table, the duration of upgrading the taboo table, will continue algebraically n _Tabu＞5 coded sequence deletion;

(5.4) judge whether iteration is stipulated algebraically to tabu search algorithm, optimize P as a result if then finish tabu search algorithm output _GcurOtherwise transfer step (5.2) to (t).

Images