CN111157002A - Aircraft 3D path planning method based on multi-agent evolutionary algorithm - Google Patents

Abstract

The invention provides an aircraft 3D path planning method based on multi-agent evolution, which comprises the following steps: setting the starting point coordinate, the end point coordinate and the multi-agent evolution parameter of the aircraft, and initializing the multi-agent grid L^tThen, the tth generation multi-agent grid L is calculated^tMiddle intelligent agent L_i,j ^tAccording to the energy of the intelligent agent, obtaining the global optimal intelligent agent Best^tThen to L^tMiddle intelligent agent L_i,j ^tPerforming neighborhood competition operators, evolving competition failure agents through crossover and mutation operators, and gathering competition winning agents into Part_winThe intelligent agent in the system carries out self-learning operator to complete the evolution of the intelligent agent grid, and finally, the maximum iteration number is controlled to calculate L^tAnd (5) acquiring the 3D path planning result of the aircraft by the medium-intelligent body energy. The method can effectively reduce the calculated amount of the 3D path planning of the aircraft and improve the convergence speed of the 3D path planning of the aircraft on the basis of ensuring the practicability of the 3D path planning result of the aircraft.

Description

Aircraft 3D path planning method based on multi-agent evolutionary algorithm

Technical Field

The invention belongs to the technical field of path planning, relates to a 3D path planning method of an aircraft, and particularly relates to a 3D path planning method of an aircraft based on a multi-agent evolutionary algorithm, which can be used in the fields of intelligent navigation, unmanned flight and the like of the aircraft.

Background

The aircraft path planning can enable the aircraft to automatically calculate the optimal path from the starting point to the end point, and is a key factor for realizing autonomous navigation and control of all aircraft. At present, a plurality of path planning methods such as a random tree method, an A star algorithm and the like are provided around the 2D path planning of a two-dimensional map, and a good planning effect is obtained. Compared with 2D path planning, the environment for 3D path planning is more complex, modeling difficulty is higher, the planned optimal path is not always the shortest path, and more importantly, complex terrain factors, dangerous areas in the environment and the like need to be considered. In recent years, with the wide application of aircraft in fields such as forest fire prevention, aerial survey, on-site rescue, express delivery, military reconnaissance. The complex environment brings great difficulty to the 3D path planning of the aircraft, and on the other hand, the aircraft can be subjected to efficient real-time path planning in a dynamic environment due to more demands. Therefore, the efficient path planning method in the 3D space is designed, and the method has great significance for automatic navigation control and unmanned flight of the aircraft. The 3D path planning of the aircraft is widely applied in practice, theoretically, modeling modes of the problem are rich, the problem belongs to a typical NP-hard problem, and the problem is difficult to solve, so that how to efficiently obtain a better path planning result is fundamentally a difficult problem.

At present, research methods for planning 3D paths of aircrafts in documents are mainly divided into traditional search algorithms represented by a random tree method and an A-star algorithm, and meta-heuristic algorithms represented by a genetic algorithm, a simulated annealing algorithm and a particle swarm optimization algorithm. In the traditional search algorithm, a 3D environment space is modeled into a network topology connected graph model, and then path planning is converted into a problem of solving the shortest path on the network topology connected graph, so that high-speed path planning can be rapidly carried out. However, the method for constructing the network topology connectivity graph has the problem of difficulty in modeling the environment when facing complex environments, and in the actual many applied 3D spaces, the optimal path is not the shortest path generally. Therefore, the practical application of the traditional search algorithm in 3D path planning is limited to a great extent. Compared with the traditional search algorithm, the meta-heuristic algorithm only needs to design the loss function value of the planned path in the complex environment, and the modeling complexity is high-efficiency and low in difficulty. Meta-heuristic algorithms are currently generally considered to be the best method after trade-offs in terms of performance, scalability, and ease of implementation. The genetic algorithm is used as an efficient meta-heuristic algorithm and is effectively applied to the field of path planning. For example, the paper "Fast Genetic Algorithm Path Planner for Fixed-Wing Military UAV Using GPU" (IEEE Transactions on Aerospace and Electronic Systems, Page No.: 99, 2018) published by Roberge et al proposes a 3D Path planning method Using Genetic Algorithm GA. The method comprises the steps of coding a planned path into chromosomes, randomizing and generating a large number of chromosomes on the basis of simulating biological evolution to form a chromosome population representing the path, calculating the fitness value of each path according to the environment, calculating the competitiveness of each individual in the whole population to ensure that the path represented by the outstanding chromosome can propagate and survive in the population, and realizing parallel calculation of the fitness values of the individuals in the population by using a graphic processor GPU when calculating the fitness value of the path, so that the calculation process of evaluating the whole population is accelerated, and finally aiming at the path planning problem, the method designs variation on the path represented by the chromosome and performs cross operation to ensure that the chromosome is evolved. The method realizes highly parallel and self-adaptive optimization of path planning, and has the advantages of low loss value of the optimized path result and strong practicability. However, this method is essentially a GA-based algorithm, and the individuals used for generating offspring are selected from the whole population according to fitness, so the fitness distribution of the whole population must be predetermined and then compared and competed globally, and therefore the planning of the trajectory is not promoted from the perspective of the algorithm. In fact, natural selection is itself a local phenomenon that only makes information globally shared by the gradual diffusion of local competition, in relation to the local environment in which the individual is located. In this method, the amount of calculation is doubled when the population size of the path represented by the designed chromosome increases. The path represented by each chromosome focuses on the fitness value of other paths in the whole population in the optimization process, and the path planning process needs the synchronous co-evolution of the whole population formed by all paths, which results in the defects of large calculation amount and slow convergence speed of the method to a great extent.

Disclosure of Invention

The invention aims to overcome the defects of the prior art, provides an aircraft 3D path planning method based on a multi-agent evolutionary algorithm, and aims to reduce the calculated amount of the aircraft 3D path planning and improve the convergence speed of the aircraft 3D path planning on the basis of ensuring the practicability of the aircraft 3D path planning result.

In order to achieve the purpose, the technical scheme adopted by the invention comprises the following steps:

(1) setting the starting point coordinates and the end point coordinates of the aircraft:

setting the starting point coordinate of the aircraft in the digital elevation map model map as st ═ (x)₀,y₀,z₀) The coordinate of the end point is end ═ x_N,y_N,z_N) Dividing a path from st to end into N sections, wherein N is more than 20;

(2) setting parameters of multi-agent evolution:

setting a multi-agent grid adopted by multi-agent evolution as L and a self-learning multi-agent grid as sL, wherein the size of L is L_size×L_size，L_sizeMore than 4, the evolution algebra is T, the maximum evolution algebra is T, T is more than 40, and the size of sL is sL_size×sL_size，3≤sL_sizeThe self-learning evolution algebra is ts, the self-learning maximum evolution algebra is sT, sT is more than 3 and less than 9, and t is equal to 0;

(3) initializing Multi-agent grid L^t：

(3a) The position coordinate set of N-1 path planning points of the random initial aircraft in the map is path_i,j＝{(x₁,y₁,z₁)_i,j,(x₂,y₂,z₂)_i,j,…,(x_N-1,y_N-1,z_N-1)_i,jAnd will path_i,jSt and end are combined to obtain the agent L of the ith row and the jth column in the L_i,j ^t,L_i,j ^t＝[st,path_i,j,end]；

(3b) Mixing L with_size×L_sizeInitialization multi-agent grid L formed by combining intelligent agents^t，

(4) Computing tth generation multi-agent grid L^tMiddle intelligent agent L_i,j ^tEnergy of (2):

through t-th generation multi-agent grid L^tMiddle intelligent agent L_i,j ^tLoss function value F on map_loss(L_i,j ^tMap), calculate L in Lt_i,j ^tEnergy En (L) of_i,j ^t)＝1/F_loss(L_i,j ^t,map)；

(5) Obtaining a globally optimal agent Best^t：

Judging whether t is equal to 0 or not, if yes, judging L^tAgent cBest with maximum medium energy^tBest as a globally optimal agent^tOtherwise, comparing the agent cBest^tAnd Best^t-1And the agent with large energy is taken as the global optimal agent Best^t；

(6) To L^tMiddle intelligent agent L_i,j ^tPerforming neighborhood competition operators:

(6a) definition of L^tMiddle intelligent agent L_i,j ^tThe four neighborhood agents of

The agent with the highest energy among the four neighborhood agents is L_i,j ^maxWherein, the% is expressed as mathematical modulus operation, and the competition failure agent set is Part_loseThe competitive winning agent is set as Part_winAnd make an order

(6b) Judging En (L)_i,j ^t)＜En(L_i,j ^max) If yes, executing the step (6c), otherwise, executing the step (6 d);

(6c) with P_cIs a probability pair L_i,j ^maxAnd L_i,j ^tCrossing to obtain offspring agent L_i,j ^gAnd with P_mIs a probability pair L_i,j ^gPerforming mutation to obtain new agent L_i,j ^newThen add all new agents to the set Part_loseAfter (5), executing step (6 e);

(6d) will agent L_i,j ^tAs a winning agent L_i,j ^winAnd add all winning agents to the set Part_winAfter (5), executing step (6 e);

(6e) part is to be_loseEach new agent and Part_winAdding each winning intelligent agent to the corresponding position in the multi-intelligent-agent grid L to obtain L^tTo L^t+1Middle generation multi-agent grid L^t+1/2；

(7) Part of competitive winning intelligent agent set_winThe agent in (1) performs self-learning operator:

(7a) defining the set after the competition winning agent evolves as Part_{win_new}And make an order

(7b) Part is treated according to the method of step (6c)_winEach winning agent L in the series_i,j ^winCarrying out sL_size×sL_sizeMinor variation to give a sum of L_i,j ^winThe associated ts-th generation of self-learning multi-agent mesh, let ts equal 0,

(7c) calculating sL according to the method of step (4)^i,j,tsMiddle intelligent agent sL_m,n ^i,j,tsAnd according to the method of step (6) on sL^i,j,tsMiddle sL_m,n ^i,j,tsPerforming field competition to obtain an intermediate generation self-learning intelligent agent grid sL^i,j,ts+1/2；

(7d) Judging whether ts < sT is true, if yes, making sL^i,j,ts+1＝sL^i,j,s+1/2Ts +1 and step (7c) is performed, otherwise sL is calculated^i,j,tsThe energy of each agent, and will sL^i,j,tsSelf-learning agent L with highest energy in_i,j ^win_newAdd set Part_{win_new}After (5), executing step (7 e);

(7e) part is to be_loseEach new agent and Part_{win_new}Adding each self-learning agent to the corresponding position in the multi-agent grid L to obtain the t +1 th generation multi-agent grid L^t+1；

(8) Obtaining a 3D path planning result of the aircraft:

let T be T +1, and judge T < T and hold, if yes, carry out step (4), otherwise, calculate L^tThe energy of each agent in the system, and comparing L^tAgent cBest with maximum medium energy^tAnd Best^t-1The agent with large energy is taken as the global optimal agent Best^tAnd outputting to realize the 3D path planning of the aircraft.

Compared with the prior art, the invention has the following advantages:

the method comprises the steps that 1, 3D path planning of an aircraft is achieved based on a multi-agent evolution algorithm, each planned path is represented as an agent and forms a multi-agent grid, each agent can only sense local environment, information of one agent is transmitted to field agents after the agent interacts with the neighborhood of the agent, the information is gradually diffused to the whole agent grid, the size of a population formed by optimized paths is smaller through the method, the situation that each agent independently lives in the population in the prior art is avoided, the ability that local parts can communicate with each other is given to the agents in the algorithm, and the calculated amount of the 3D path planning of the aircraft is effectively reduced.

2. When the multi-agent evolution algorithm is applied to the aircraft 3D path planning problem, on the basis of representing each planned path as an agent, a neighborhood competition operator and a self-learning operator of the agent are designed, the neighborhood competition operator can extract a group formed by elite paths in the group, the paths represented by weak individuals are optimized, and the self-learning operator can ensure that the elite individuals are optimized more deeply, so that local search can be effectively carried out in the evolution process.

3. According to the invention, by deeply excavating and utilizing the intelligent evolution resources of organisms and using a multi-agent evolution mode based on the perception and the reaction of the agents to the environment, the model of the multi-agent grid is closer to a real natural evolution mechanism, compared with the prior art, the loss value of the obtained 3D path is equivalent, and the practicability of the planning result of the 3D path of the aircraft is ensured.

Drawings

FIG. 1 is a flow chart of an implementation of the present invention.

FIG. 2 is a schematic diagram of the structure of a multi-agent grid in accordance with the present invention.

FIG. 3 is a schematic diagram of crossover and variation implementations of the present invention.

FIG. 4 is a graph comparing simulation of traces obtained under different sizes of populations according to the present invention and the prior art.

FIG. 5 is a graph comparing simulations of convergence speed of the present invention and the prior art.

Fig. 6 is a simulation comparison graph of the stability of the 3D path planning results of the present invention and the prior art.

Detailed Description

The invention is described in further detail below with reference to the figures and the specific embodiments.

Referring to fig. 1, the present invention includes the steps of:

step 1) setting a starting point coordinate and an end point coordinate of the aircraft:

setting the coordinates of a starting point of the aircraft in the digital elevation map model map as st ═ 0,0 and the coordinates of an end point as end ═ 400,0, and dividing a path between the st and the end into N sections, wherein N is 27;

step 2) setting parameters of multi-agent evolution:

setting a multi-agent grid adopted by multi-agent evolution as shown in fig. 2 as L, a self-learning multi-agent grid as sL, and an agent living in each grid position, wherein the grids have direct connection relations, namely, are neighbors, each agent can only interact with the neighborhood, and the size of L is L_size×L_size，L_size5, T is evolution algebra, T is maximum evolution algebra, T is 200, and sL is of size sL_size×sL_size，sL_sizeThe self-learning evolution algebra is ts, the self-learning maximum evolution algebra is sT, the sT is 6, and t is 0;

step 3) initializing multi-agent grid L^tThe method comprises the following steps that a multi-agent grid is initialized according to a task target of the 3D path planning, each agent in the grid represents a planned path, all agents exist in a grid environment, the agents transmit information to field agents after acting with neighborhoods of the agents, the information is gradually diffused to the whole agent grid, the population scale can be smaller, the condition that each individual in the prior art lives in a population independently is avoided, the individual in an algorithm is endowed with the capacity of local mutual communication, and the calculated amount of the 3D path planning of the aircraft can be effectively reduced:

(3a) the position coordinate set of N-1 path planning points of the random initial aircraft in the map is path_i,j＝{(x₁,y₁,z₁)_i,j,(x₂,y₂,z₂)_i,j,…,(x_N-1,y_N-1,z_N-1)_i,jAnd will path_i,jSt and end are combined to obtain the agent L of the ith row and the jth column in the L_i,j ^t,L_i,j ^t＝[st,path_i,j,end]；

(3b) Mixing L with_size×L_sizeInitialization multi-agent grid L formed by combining intelligent agents^t，

Step 4) calculating the tth generation multi-agent grid L^tMiddle intelligent agent L_i,j ^tEnergy of (2):

through t-th generation multi-agent grid L^tMiddle intelligent agent L_i,j ^tLoss function value F on map_loss(L_i,j ^tMap), calculate L^tMiddle L_i,j ^tEnergy En (L) of_i,j ^t)：

En(L_i,j ^t)＝1/F_loss(L_i,j ^t,map)；

F_loss(L_i,j ^t,map)＝w₁*Loss_path(L_i,j ^t)+w₂*Loss_altitude(L_i,j ^t,map)；

Wherein Loss_path(L_i,j ^t) For path length Loss, Loss_altitude(L_i,j ^tMap) is the loss in altitude, w, of the path projected onto the map₁And w₂Are respectively a pair of Loss_path(L_i,j ^t) And Loss_altitude(L_i,j ^tPenalty weight of (x) map_k,y_k,z_k) Represents L_i,jThe value of the k-th point coordinate in the agent,

representing the summation of the N sample calculated values taken, | · | | non-woven phosphor₂Expression solution2 norm of vector, [ (x)_k-1,y_k-1,z_k-1)→(x_k,y_k,z_k)]_MIs shown at L_i,jM point coordinates obtained by performing M times of mean value sampling on a road section formed by connecting the k-1 point to the k point, map ∩ [ (x)_k-1,y_k-1,z_k-1)→(x_k,y_k,z_k)]_MThe method comprises the following steps of (1) solving M altitude values of positions where M point coordinates obtained by sampling are projected to map, wherein mean {. cndot } represents the average of element values in a set;

step 5) obtaining a global optimal agent Best^t：

Judging whether t is equal to 0 or not, if yes, judging L^tAgent cBest with maximum medium energy^tBest as a globally optimal agent^tOtherwise, comparing the agent cBest^tAnd Best^t-1And the agent with large energy is taken as the global optimal agent Best^t；

Step 6) for L^tMiddle intelligent agent L_i,j ^tPerforming neighborhood competition operators, optimizing disadvantaged individuals in the population through the domain competition, extracting elite populations in the population, and finally obtaining a semi-evolved multi-agent grid:

(6a) definition of L^tMiddle intelligent agent L_i,j ^tThe four neighborhood agents of

The agent with the highest energy among the four neighborhood agents is L_i,j ^maxWherein, the% is expressed as mathematical modulus operation, and the competition failure agent set is Part_loseThe competitive winning agent is set as Part_winAnd make an order

(6b) Judging En (L)_i,j ^t)＜En(L_i,j ^max) If yes, executing the step (6c), otherwise, executing the step (6 d);

(6c) with P_cIs a probability pair L_i,j ^maxAnd L_i,j ^tCrossing to obtain offspring agent L_i,j ^gAnd with P_mIs a probability pair L_i,j ^gPerforming mutation to obtain new agent L_i,j ^newThen add all new agents to the set Part_loseAfter (5), step (6e) is performed, wherein the new agent L_i,j ^newThe acquisition process comprises the following steps:

(6c1) randomly generated real numbers r between 0 and 1₁If r is₁＞P_cLet the offspring agent L_i,j ^g＝L_i,j ^maxExecuting the step (6c3), otherwise, executing the step (6c 2);

(6c2) to L_i,j ^maxAnd L_i,j ^tPerforming crossover operator, and randomly selecting an integer p in a range of (1, N-1)_nThen from p_nThe position divides two intelligent agents into a front section and a rear section, and an intelligent agent L_i,j ^maxAnd L_i,j ^tThe posterior portions are connected to form a child individual, as shown in fig. 3 (a):

randomly generating a random integer p between 1 and N-1_nAe of No. 1, N-1, and taking L_i,j ^maxMiddle 0 th position coordinate to p th position coordinate_nCoordinate set path of path segment formed by position coordinates₁Taking L_i,j ^tMiddle (p)_n+1 position coordinate to Nth position coordinate to form the coordinate set path of the path segment₂Will path₁、path₂Combined to form a descendant agent L_i,j ^g＝[path₁,path₂]；

(6c3) Randomly generated real numbers r between 0 and 1₂If r is₂＞P_mLet L_i,j ^new＝L_i,j ^gExecuting the step (6c5), otherwise, executing the step (6c 4);

(6c4) to L_i,j ^gPerforming mutation operator, randomly selecting an integer N in a range of (1, N-1), and then changing the coordinate of the nth position into a value between the coordinates of the (N-1) th position and the (N + 1) th position to obtain a new agent, as shown in fig. 3 (b):

randomly generating a random integer N ← random (1, N-1) between 1 and N-1, obtaining L_i,j ^gMiddle (n-1) position coordinate (x)_n-1,y_n-1,z_n-1) And n +1 position coordinates (x)_n+1,y_n+1,z_n+1) Randomly generating 3 real numbers r between 0 and 1_x、r_y、r_zAnd calculating the coordinate of the variation position as (x)_r,y_r,z_r)：

x_r＝x_n-1+r_x(x_n+1-x_n-1)；

y_r＝y_n-1+r_y(y_n+1-y_n-1)；

z_r＝z_n-1+r_z(z_n+1-z_n-1)；

Mixing L with_i,j ^gThe nth position coordinate is reassigned to (x)_r,y_r,z_r) And make L_i,j ^new＝L_i,j ^g；

(6c5) Will new agent L_i,j ^newAdd to set Part_losePerforming the following steps;

(6d) will agent L_i,j ^tAs a winning agent L_i,j ^winAnd add all winning agents to the set Part_winAfter (5), executing step (6 e);

(6e) part is to be_loseEach new agent and Part_winAdding each winning intelligent agent to the corresponding position in the multi-intelligent-agent grid L to obtain L^tTo L^t+1Middle generation multi-agent grid L^t+1/2；

Step 7) set Part of competitive winning intelligent agent_winThe intelligent agent in the system carries out self-learning operator, the self-learning operator can ensure that the elite individual in the population is deeply optimized, and the population is optimizedIn the process, local search is carried out near the high-quality individuals, and the aim of accelerating the convergence speed of the whole group is achieved:

(7a) defining the set after the competition winning agent evolves as Part_{win_new}And make an order

(7b) Part is treated according to the method of step (6c)_winEach winning agent L in the series_i,j ^winCarrying out sL_size×sL_sizeMinor variation to give a sum of L_i,j ^winRelated ts generation self-learning multi-agent grid, let ts equal 0

(7c) Calculating sL according to the method of step (4)^i,j,tsMiddle intelligent agent sL_m,n ^i,j,tsAnd according to the method of step (6) on sL^i,j,tsMiddle sL_m,n ^i,j,tsPerforming field competition to obtain an intermediate generation self-learning intelligent agent grid sL^i,j,ts+1/2；

(7d) Judging whether ts < sT is true, if yes, making sL^i,j,ts+1＝sL^i,j,s+1/2Ts +1 and step (7c) is performed, otherwise sL is calculated^i,j,tsThe energy of each agent, and will sL^i,j,tsSelf-learning agent L with highest energy in_i,j ^win_newAdd set Part_{win_new}After (5), executing step (7 e);

(7e) part is to be_loseEach new agent and Part_{win_new}Adding each self-learning agent to the corresponding position in the multi-agent grid L to obtain the t +1 th generation multi-agent grid L^t+1Let t be t + 1;

step 8) obtaining a 3D path planning result of the aircraft:

judging whether T is more than T, if so, executing the step (4), otherwise, calculating L^tThe energy of each agent in the system, and comparing L^tAgent cBest with maximum medium energy^tAnd Best^t-1Energy ofThe size of the agent is larger, and the agent with larger energy is taken as the global optimal agent Best^tAnd outputting to realize the 3D path planning of the aircraft.

The technical effects of the present invention will be further described with reference to simulation experiments.

1. Simulation conditions and contents:

environment of experimental run: the operating system is Microsoft windows 10, and the programming simulation language is matlab. The experiment uses four digital elevation map model data, and the convergence rate curve and the loss function value are averaged results after 40 independent experiments.

Simulation 1: the results of comparative simulation of the trajectories obtained by the present invention and the prior GA techniques under different scale populations are shown in fig. 4.

Simulation 2: the results of comparative simulation of the convergence rate of the present invention and the prior art GA are shown in fig. 5.

Simulation 3: the results of the comparison and simulation of the stability of the 3D path planning results of the invention and the prior GA technology are shown in FIG. 6.

2. And (3) simulation result analysis:

referring to FIG. 4, FIG. 4(a) compares the trace results obtained from 3D path planning performed on a real elevation data map model according to the prior art GA, and FIG. 4(b) compares the trace results obtained from 3D path planning performed on a synthetic virtual elevation data map model according to the prior art GA. The altitude represented by the circular area in the figure is infinitely high. In different map environments, the size of the population scale in the prior art is adjusted, so that the prior art can obtain a path planning result similar to that of the method. The invention sets the number of agents in the multi-agent grid to be 25, and when the finally used prior art population scale is 100, the track result of the 3D path planning similar to that of the invention in the figure 4 can be obtained. The invention uses smaller-scale population, obtains the result similar to the prior art, and shows that the invention can effectively reduce the calculated amount of the 3D path planning of the aircraft.

Referring to fig. 5, the abscissa is the number of iterations and the ordinate is the normalized loss function value of the optimal individual in the population during the iteration process, the present invention uses the neighborhood competition and the self-learning process, thereby accelerating the evolution process of outstanding individuals. The method has absolute advantage in convergence speed, and can tend to converge in 40 iterations, while the algebraic number required for achieving convergence in the prior art is 70. Experimental results show that the method can effectively improve the convergence speed of the 3D path planning of the aircraft

Referring to fig. 6, the abscissa represents four different 3D maps, and the ordinate represents the average loss function value of the population, and the loss function values of the trajectories optimized by the present invention are comparable to the results obtained by the prior art in four environments. The method can ensure the practicability of the aircraft 3D path planning result.

Claims (3)

1. An aircraft 3D path planning method based on a multi-agent evolution algorithm is characterized by comprising the following steps:

(1) setting the starting point coordinates and the end point coordinates of the aircraft:

setting the starting point coordinate of the aircraft in the digital elevation map model map as st ═ (x)₀,y₀,z₀) The coordinate of the end point is end ═ x_N,y_N,z_N) Dividing a path from st to end into N sections, wherein N is more than 20;

(2) setting parameters of multi-agent evolution:

setting a multi-agent grid adopted by multi-agent evolution as L and a self-learning multi-agent grid as sL, wherein the size of L is L_size×L_size，L_sizeMore than 4, the evolution algebra is T, the maximum evolution algebra is T, T is more than 40, and the size of sL is sL_size×sL_size，3≤sL_sizeThe self-learning evolution algebra is ts, the self-learning maximum evolution algebra is sT, sT is more than 3 and less than 9, and t is equal to 0;

(3) initializing Multi-agent grid L^t：

(3a) The position coordinate set of N-1 path planning points of the random initial aircraft in the map is path_i,j＝{(x₁,y₁,z₁)_i,j,(x₂,y₂,z₂)_i,j,...,(x_N-1,y_N-1,z_N-1)_i,jAnd will path_i,jSt and end are combined to obtain the agent L of the ith row and the jth column in the L_i,j ^t,L_i,j ^t＝[st,path_i,j,end]；

(3b) Mixing L with_size×L_sizeInitialization multi-agent grid L formed by combining intelligent agents^t，

(4) Computing tth generation multi-agent grid L^tMiddle intelligent agent L_i,j ^tEnergy of (2):

through t-th generation multi-agent grid L^tMiddle intelligent agent L_i,j ^tLoss function value F on map_loss(L_i,j ^tMap), calculate L^tMiddle L_i,j ^tEnergy En (L) of_i,j ^t)＝1/F_loss(L_i,j ^t,map)；

(5) Obtaining a globally optimal agent Best^t：

Judging whether t is equal to 0 or not, if yes, judging L^tAgent cBest with maximum medium energy^tBest as a globally optimal agent^tOtherwise, comparing the agent cBest^tAnd Best^t-1And the agent with large energy is taken as the global optimal agent Best^t；

(6) To L^tMiddle intelligent agent L_i,j ^tPerforming neighborhood competition operators:

(6a) definition of L^tMiddle intelligent agent L_i,j ^tThe four neighborhood agents of

The agent with the highest energy among the four neighborhood agents is L_i,j ^maxWhereinThe% is expressed as mathematical modular operation, and the competition failure agent set is Part_loseThe competitive winning agent is set as Part_winAnd make an order

(6b) Judging En (L)_i,j ^t)＜En(L_i,j ^max) If yes, executing the step (6c), otherwise, executing the step (6 d);

(6c) with P_cIs a probability pair L_i,j ^maxAnd L_i,j ^tCrossing to obtain offspring agent L_i,j ^gAnd with P_mIs a probability pair L_i,j ^gPerforming mutation to obtain new agent L_i,j ^newThen add all new agents to the set Part_loseAfter (5), executing step (6 e);

(6d) will agent L_i,j ^tAs a winning agent L_i,j ^winAnd add all winning agents to the set Part_winAfter (5), executing step (6 e);

(6e) part is to be_loseEach new agent and Part_winAdding each winning intelligent agent to the corresponding position in the multi-intelligent-agent grid L to obtain L^tTo L^t+1Middle generation multi-agent grid L^t+1/2；

(7) Part of competitive winning intelligent agent set_winThe agent in (1) performs self-learning operator:

(7a) defining the set after the competition winning agent evolves as Part_{win_new}And make an order

(7b) Part is treated according to the method of step (6c)_winEach winning agent L in the series_i,j ^winCarrying out sL_size×sL_sizeMinor variation to give a sum of L_i,j ^winThe associated ts-th generation of self-learning multi-agent mesh, let ts equal 0,

(7c) calculating sL according to the method of step (4)^i,j,tsMiddle intelligent agent sL_m,n ^i,j,tsAnd according to the method of step (6) on sL^i,j,tsMiddle sL_m,n ^i,j,tsPerforming field competition to obtain an intermediate generation self-learning intelligent agent grid sL^i,j,ts+1/2；

(7d) Judging whether ts < sT is true, if yes, making sL^i,j,ts+1＝sL^i,j,s+1/2Ts +1 and step (7c) is performed, otherwise sL is calculated^i,j,tsThe energy of each agent, and will sL^i,j,tsSelf-learning agent L with highest energy in_i,j ^win_newAdd set Part_{win_new}After (5), executing step (7 e);

(7e) part is to be_loseEach new agent and Part_{win_new}Adding each self-learning agent to the corresponding position in the multi-agent grid L to obtain the t +1 th generation multi-agent grid L^t+1Let t be t + 1;

(8) obtaining a 3D path planning result of the aircraft:

and judging whether T is more than T, if so, executing the step (4), otherwise, calculating L^tThe energy of each agent in the system, and comparing L^tAgent cBest with maximum medium energy^tAnd Best^t-1The agent with large energy is taken as the global optimal agent Best^tAnd outputting to realize the 3D path planning of the aircraft.

2. The multi-agent evolution-based aircraft 3D path planning method according to claim 1, wherein the tth generation multi-agent mesh L in step (4)^tMiddle intelligent agent L_i,j ^tLoss function value F on map_loss(L_i,j ^tMap), the calculation formula is:

F_loss(L_i,j ^t,map)＝w₁*Loss_path(L_i,j ^t)+w₂*Loss_altitude(L_i,j ^t,map)；

wherein Loss_path(L_i,j ^t) For path length Loss, Loss_altitude(L_i,j ^tMap) is the loss in altitude, w, of the path projected onto the map₁And w₂Are respectively a pair of Loss_path(L_i,j ^t) And Loss_altitude(L_i,j ^tPenalty weight of (x) map_k,y_k,z_k) Represents L_i,jThe value of the k-th point coordinate in the agent,

representing the summation of the N sample calculated values taken, | · | | non-woven phosphor₂2 norm of the vector, [ (x)_k-1,y_k-1,z_k-1)→(x_k,y_k,z_k)]_MIs shown at L_i,jM point coordinates obtained by performing M times of mean value sampling on a road section formed by connecting the k-1 point to the k point, map ∩ [ (x)_k-1,y_k-1,z_k-1)→(x_k,y_k,z_k)]_MThe method comprises the following steps of calculating M altitude values of positions where M point coordinates obtained by sampling are projected to map, and mean {. DEG } represents the average of element values in a set.

3. The multi-agent evolution based aircraft 3D path planning method according to claim 1, characterized in that said new agent L in step (6c)_i,j ^newThe acquisition comprises the following steps:

(6c1) randomly generated 0 to 1Real number r of₁If r is₁＞P_cLet the offspring agent L_i,j ^g＝L_i,j ^maxExecuting the step (6c3), otherwise, executing the step (6c 2);

(6c2) to L_i,j ^maxAnd L_i,j ^tAnd (3) carrying out a crossover operator:

randomly generating a random integer p between 1 and N-1_nAe of No. 1, N-1, and taking L_i,j ^maxMiddle 0 th position coordinate to p th position coordinate_nCoordinate set path of path segment formed by position coordinates₁Taking L_i,j ^tMiddle (p)_n+1 position coordinate to Nth position coordinate to form the coordinate set path of the path segment₂Will path₁、path₂Combined to form a descendant agent L_i,j ^g＝[path₁,path₂]；

(6c3) Randomly generated real numbers r between 0 and 1₂If r is₂＞P_mLet L_i,j ^new＝L_i,j ^gOtherwise, performing step (6c 4);

(6c4) to L_i,j ^gCarrying out mutation operator:

randomly generating 1 to N_-Random integer N ← random (1, N-1) between 1, obtain L_i,j ^gMiddle (n-1) position coordinate (x)_n-1,y_n-1,z_n-1) And n +1 position coordinates (x)_n+1,y_n+1,z_n+1) Randomly generating 3 real numbers r between 0 and 1_x、r_y、r_zAnd calculating the coordinate of the variation position as (x)_r,y_r,z_r)：

x_r＝x_n-1+r_x(x_n+1-x_n-1)；

y_r＝y_n-1+r_y(y_n+1-y_n-1)；

z_r＝z_n-1+r_z(z_n+1-z_n-1)；

Mixing L with_i,j ^gThe nth position coordinate is reassigned to (x)_r,y_r,z_r) And make L_i,j ^new＝L_i,j ^g。

Images