Air combat confrontation game strategy optimization system based on multi-population adaptive correction evolutionary algorithm
By constructing an air combat game simulation module and a task constraint module, and combining a multi-population evolution optimization module and adaptive correction technology, the problems of speed, accuracy and stability of air combat game strategy optimization are solved, and the rational allocation of air combat tasks and the maximization of benefits are realized.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- ZHEJIANG UNIV
- Filing Date
- 2022-12-21
- Publication Date
- 2026-06-12
AI Technical Summary
Existing air combat game strategy optimization systems are unable to achieve rapid and accurate strategy optimization, resulting in difficulties in providing effective command and decision support on the battlefield.
An air combat game simulation module is constructed, which combines an air combat mission constraint module and a multi-swarm evolution optimization module. A hierarchical particle swarm algorithm with local center selection and adaptive correction technology are used to find the optimal strategy that satisfies the constraints and output it to the air combat elements.
It achieves rapid, accurate, and stable optimization of air combat game strategies, maximizes our gains, avoids premature convergence, and ensures the rationality and effectiveness of the strategy.
Smart Images
Figure CN115994484B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of air combat game strategies and optimization algorithms, and in particular to an air combat game strategy optimization system based on a multi-population adaptive correction evolution algorithm. Background Technology
[0002] The decision-making outcomes of air combat confrontation are one of the most important reference information for modern air battlefield command and decision-making. With the increasing intelligence and complexity of current air combat elements and the diversity of air combat missions, each air combat mission requires multiple friendly elements to complete, and sometimes a single element also needs to handle multiple air combat missions. However, air combat elements are finite, especially when the number of elements on both sides is similar, and friendly elements often cannot meet all air combat mission requests. Against this backdrop, how to rationally conduct confrontational game theory based on the combat situation of both sides to allocate tasks to friendly elements is crucial to battlefield confrontation.
[0003] Most current air combat game strategy optimization systems struggle to achieve rapid and accurate strategy optimization, thus failing to provide effective assistance for command and decision-making in real-world battlefields. Therefore, this invention, based on the simulation mathematical modeling of air combat game scenarios, utilizes multiple population evolution algorithms to efficiently find the optimal game strategy. This enables the maximization of game combat gains even with limited air combat resources, providing a fast, accurate, and stable air combat game strategy optimization system. Summary of the Invention
[0004] To address the problems of slow decision-making speed, low accuracy, and unstable optimization results in current air combat game theory, this invention aims to provide a fast, accurate, and stable air combat game theory strategy optimization system. The system comprehensively considers battlefield elements and tasks through an air combat game theory simulation module, constructing a mathematical simulation model of the air combat game. An air combat task constraint module further improves the simulation model to ensure the rationality of the final strategy. A multi-population evolutionary optimization module selects the optimal strategy that satisfies the constraints, maximizing the benefits of the air combat mission. Finally, an optimal strategy output module outputs the found optimal strategy to each air combat element to guide them in completing their respective tasks. This solves the problem of the difficulty in achieving strategy optimization in traditional air combat game theory, providing a fast, accurate, and stable air combat game theory strategy optimization system.
[0005] The objective of this invention is achieved through the following technical solution: an air combat adversarial game strategy optimization system based on a multi-population adaptive correction evolutionary algorithm, characterized by comprising the following modules: an air combat adversarial game simulation module, an air combat mission constraint module, a multi-population evolutionary optimization module, and an optimal strategy output module.
[0006] The air combat confrontation simulation module constructs a dynamic model of the opposing sides based on the motion laws of integrated air combat battlefield elements and combat mission confrontation:
[0007]
[0008] x(t0)=x0
[0009] in, The system equations are given, and the initial state of the system is x(t0) = x0. Represents n x The state variables of a dimensional game. Represents n u Control variables in the game. Since the players' game strategies are unknown, we consider the opponent's strategy under extreme intelligence and our own strategy under the worst-case scenario. That is, the payoff functions for the dynamic game between the two sides are as follows:
[0010]
[0011]
[0012] In the model, the variable subscripts A and B represent our side and the enemy side, respectively, and the payoff function J... A Including time periods [t0, t f Integral terms on ] With terminal time t f The final value term Φ A [x(t f Payment function J] B Including time periods [t0, t f Integral terms on ] With terminal time t f The final value term Φ B [x(t f )], and These are bilateral game strategies involving the control variables of the opposing side and the player's side.
[0013] Based on this, the air combat mission constraint module constructs complex constraints for battlefield adversarial game based on the characteristics of coupled and rapidly changing battlefield background environment:
[0014]
[0015]
[0016] u l ≤u(t)≤u u
[0017] t∈[t0,t f ]
[0018] Where E represents the battlefield confrontation normative equality constraint function. The set of, I, denotes the orthogonal inequality constraint function. The set of τ i Φ represents the final time of the integral in the constraint function. i (x(τ i )) represents τ i The final value term at time. Game time domain [t0, τ] i The integral term of ], u l and u u This represents the upper and lower bounds of the control variables in a game.
[0019] Furthermore, the multi-swarm evolution optimization module employs a novel hierarchical particle swarm optimization algorithm based on local center selection to find the optimal air combat game strategy, which can satisfy all battlefield environment and mission constraints while maximizing our own gains. The process is as follows:
[0020] 1) Randomly generate the first generation of initial particles to obtain the initial solution's velocity and position;
[0021] v op (1)=U×(v max -v min )+v min
[0022] r op (1) = int(U×(r max -r min )+r min )
[0023] Where o = 1, 2, ..., s, s is the group size, p corresponds to the parameters to be optimized, and v jp (1) and r jp (1) Let v represent the velocity and position of the p-th component of the o-th particle in the first generation, respectively. max v min Represents the maximum and minimum speeds, r max ,r min The maximum and minimum values of the position are represented by `int()`, which means rounding down. U is a random number uniformly distributed between [0,1]. Each dimension of the position is the task to be selected, so the dimensions of both speed and position are m.
[0024] 2) Calculate the fitness of the o-th particle. o :
[0025]
[0026] Among them, fit o This represents the particle fitness value; all other meanings are the same as described above. In maximizing the fitness value, a larger fitness value is preferable.
[0027] 3) Divide into multiple population topologies:
[0028] SP start =argmax((fit o |i∈dataset))
[0029] SP i+1 =S i *{i}∪SP i
[0030]
[0031] Among them nei i SP is the k-nearest point of the particle. start SP represents the initial state of population 1. i+1 Represents population 1 after the i-th update, t i This is used to determine whether point i is a particle in population 1. Selecting population 1 in this way has two advantages: 1. Particles with high local fitness can be selected; 2. The selected particles are far apart, ensuring the diversity of the entire population 1 and avoiding premature convergence. Particles other than those in population 1 belong to population 2.
[0032] 4) Update the inertia weight coefficient μ(t):
[0033]
[0034] Where μ max =0.9 is the upper limit of μ, μ min =0.4 is the lower limit of μ, and e represents the current iteration number. max This represents the maximum number of iterations.
[0035] 5) Update the velocity and position of the particles to generate a new swarm;
[0036]
[0037]
[0038]
[0039] r op (t+1) = int(r) op (t)+v op (t+1))
[0040] Among them, P ciThis represents the learning rate of the particles, where rand1 and rand2 are two random numbers between 0 and 1, and a = 0.05. That is, the improved, fully-learning individual pbest o G best These are the historical best solution for the o-th particle and the best solution for the entire swarm, respectively. op s1 For population 1, the speed update strategy is v op s2 The velocity update strategy for population 2. Updating the velocities of particles in both populations using this formula effectively prevents particles from falling into premature convergence and enhances their global search capabilities.
[0041] 6) Adaptive correction of comprehensive learning factors
[0042] For a particle population X, its particle fitness is f, the population size is N, and the problem dimension is Dim. The goal is to find its comprehensive learning factor pbest. f The steps are as follows:
[0043] Step 1: Assign a linearly increasing learning probability P to each particle in the population according to the formula above. ci .
[0044] Step 2: For each particle i, for each dimension, randomly select two particles a and b in the population, and take the dimension value of the particle with higher fitness as the candidate value.
[0045] Step 3: For each dimension of particle i, take a random number rand(0,1) and compare it with the P of that dimension. ci In comparison, if it is less than P ci Then the pbest of this dimension f The candidate value from the previous step, otherwise pbest f This represents the historical best value for particle i.
[0046] Step 4: If all dimension values are the historical best values of particle i, then randomly select two dimensions as candidate values in step 2.
[0047] By employing the novel adaptive correction technique described above, the algorithm can ensure that it purposefully corrects the learning factor based on the evolutionary position during the optimization process, thus avoiding premature convergence.
[0048] Finally, the optimal strategy output module outputs the global optimal solution, i.e. the optimal battle strategy, obtained from the multi-population evolution optimization module, and executes it on the battlefield elements.
[0049] The technical concept of this invention is as follows: A mathematical simulation model of air combat game theory is constructed by comprehensively considering battlefield elements and tasks through an air combat game theory simulation module; the simulation model is further improved through an air combat task constraint module to ensure the rationality of the final strategy; an optimal strategy that satisfies the constraints is selected through a multi-population evolution optimization module to maximize the benefits of air combat tasks; finally, the optimal strategy output module outputs the found optimal strategy to each air combat element to guide it in completing its corresponding task. This solves the problem of the difficulty in achieving strategy optimization in traditional air combat game theory, thus establishing a fast, accurate, and stable air combat game theory strategy optimization system.
[0050] The beneficial effects of this invention are mainly reflected in the following aspects: 1. The air combat confrontation game simulation module and the air combat mission constraint module model the mathematical simulation and constraints of the air combat confrontation game, realizing a near-realistic air combat confrontation game scenario. 2. The improved multi-population evolution optimization module allows particles with high local fitness to be selected, while the selected particles are far apart, ensuring population diversity. Then, the novel adaptive correction technology described above is used to avoid premature convergence, select game strategies that satisfy the constraints, and ensure strategy optimization, thereby maximizing the gains in the air combat confrontation game. Attached Figure Description
[0051] Figure 1 A connection diagram of an air combat adversarial game strategy optimization system based on a multi-population adaptive correction evolutionary algorithm. Detailed Implementation
[0052] The present invention will be further described below with reference to the accompanying drawings and embodiments:
[0053] refer to Figure 1 The air combat adversarial game strategy optimization system based on multi-population adaptive correction evolution algorithm includes an air combat adversarial game simulation module 1, an air combat mission constraint module 2, a multi-population evolution optimization module 3, and an optimal strategy output module 4, which are connected in sequence.
[0054] The air combat confrontation simulation module 1 constructs a dynamic model of the opposing sides based on the motion laws of integrated air combat battlefield elements and combat mission confrontation:
[0055]
[0056] x(t0)=x0
[0057] in, The system equations are given, and the initial state of the system is x(t0) = x0. Represents n x The state variables of a dimensional game. Represents n u Control variables in the game. Since the players' game strategies are unknown, we consider the opponent's strategy under extreme intelligence and our own strategy under the worst-case scenario. That is, the payoff functions for the dynamic game between the two sides are as follows:
[0058]
[0059]
[0060] In the model, the variable subscripts A and B represent our side and the enemy side, respectively, and the payoff function J... A Including time periods [t0, t f Integral terms on ] With terminal time t f The final value term Φ A [x(t f Payment function J] B Including time periods [t0, t f Integral terms on ] With terminal time t f The final value term Φ B [x(t f )], and These are bilateral game strategies involving the control variables of the opposing side and the player's side.
[0061] Based on this, the air combat mission constraint module 2 constructs complex constraints for battlefield adversarial game based on the characteristics of coupled and agile changes in the battlefield background environment:
[0062]
[0063]
[0064] u l ≤u(t)≤u u
[0065] t∈[t0,t f ]
[0066] Where E represents the battlefield confrontation normative equality constraint function. The set of, I, denotes the orthogonal inequality constraint function. The set of τ i Φ represents the final time of the integral in the constraint function. i (x(τ i )) represents τ i The final value term at time. Game time domain [t0, τ] i The integral term of ], ul and u u This represents the upper and lower bounds of the control variables in a game.
[0067] Furthermore, the multi-population evolutionary optimization module 3 employs a novel hierarchical particle swarm optimization algorithm based on local center selection to find the optimal air combat game strategy, which can satisfy all battlefield environment and mission constraints while maximizing our own gains. The process is as follows:
[0068] 1) Randomly generate the first generation of initial particles to obtain the initial solution's velocity and position;
[0069] v op (1)=U×(v max -v min )+v min
[0070] r op (1) = int(U×(r max -r min )+r min )
[0071] Where o = 1, 2, ..., s, s is the group size, p corresponds to the parameters to be optimized, and v jp (1) and r jp (1) Let v represent the velocity and position of the p-th component of the o-th particle in the first generation, respectively. max v min Represents the maximum and minimum speeds, r max ,r min The maximum and minimum values of the position are represented by `int()`, which means rounding down. U is a random number uniformly distributed between [0,1]. Each dimension of the position is the task to be selected, so the dimensions of both speed and position are m.
[0072] 2) Calculate the fitness of the o-th particle. o :
[0073]
[0074] Among them, fit o This represents the particle fitness value; all other meanings are the same as described above. In maximizing the fitness value, a larger fitness value is preferable.
[0075] 3) Divide into multiple population topologies:
[0076] SP start =argmax((fit o |i∈dataset))
[0077] SP i+1 =S i*{i}∪SP i
[0078]
[0079] Among them nei i SP is the k-nearest point of the particle. start SP represents the initial state of population 1. i+1 Represents population 1 after the i-th update, t i This is used to determine whether point i is a particle in population 1. The novel method described above has the following two advantages in selecting population 1: 1. Particles with high local fitness can be selected; 2. The selected particles are far apart, ensuring the diversity of the entire population 1 and avoiding premature convergence. Besides population 1, the other particles belong to population 2.
[0080] 4) Update the inertia weight coefficient μ(t):
[0081]
[0082] Where μ max =0.9 is the upper limit of μ, μ min =0.4 is the lower limit of μ, and e represents the current iteration number. max This represents the maximum number of iterations.
[0083] 5) Update the velocity and position of the particles to generate a new swarm;
[0084]
[0085]
[0086]
[0087] r op (t+1) = int(r) op (t)+v op (t+1))
[0088] Among them, P ci This represents the learning rate of the particles, where rand1 and rand2 are two random numbers between 0 and 1, and a = 0.05. That is, the improved, fully-learning individual pbest o G best These are the historical best solution for the o-th particle and the best solution for the entire swarm, respectively. op s1 For population 1, the speed update strategy is v op s2The velocity update strategy for population 2. Updating the velocities of particles in both populations using this formula effectively prevents particles from falling into premature convergence and enhances their global search capabilities.
[0089] 6) Adaptive correction of comprehensive learning factors
[0090] For a particle population X, its particle fitness is f, the population size is N, and the problem dimension is Dim. The goal is to find its comprehensive learning factor pbest. f The steps are as follows:
[0091] Step 1: Assign a linearly increasing learning probability P to each particle in the population according to the formula above. ci .
[0092] Step 2: For each particle i, for each dimension, randomly select two particles a and b in the population, and take the dimension value of the particle with higher fitness as the candidate value.
[0093] Step 3: For each dimension of particle i, take a random number rand(0,1) and compare it with the P of that dimension. ci In comparison, if it is less than P ci Then the pbest of this dimension f The candidate value from the previous step, otherwise pbest f This represents the historical best value for particle i.
[0094] Step 4: If all dimension values are the historical best values of particle i, then randomly select two dimensions as candidate values in step 2.
[0095] By employing the novel adaptive correction technique described above, the algorithm can ensure that it purposefully corrects the learning factor based on the evolutionary position during the optimization process, thus avoiding premature convergence.
[0096] Finally, the optimal strategy output module 4 outputs the global optimal solution, i.e. the optimal battle strategy, obtained from the multi-population evolution optimization module, to the battlefield elements for execution.
[0097] The above embodiments are used to explain and illustrate the present invention, but not to limit the present invention. Any modifications and changes made to the present invention within the spirit and scope of the claims shall fall within the protection scope of the present invention.
Claims
1. A strategy optimization system for air combat adversarial game based on a multi-population adaptive corrective evolutionary algorithm, characterized in that, The system includes the following modules: an air combat game simulation module, an air combat mission constraint module, a multi-population evolutionary optimization module, and an optimal strategy output module. The multi-population evolutionary optimization module employs a novel hierarchical particle swarm optimization algorithm based on local center selection to find the optimal air combat game strategy, which can satisfy all battlefield environments and mission constraints while maximizing our own gains. The process is as follows: 1) Randomly generate the first generation of initial particles to obtain the initial solution's velocity and position; ； ； in , For group size, These correspond to the parameters to be optimized. and They represent the first The first particle The components in the first generation are velocity and position. , Indicates the maximum and minimum speeds. Indicates the maximum and minimum values at a given position; int() represents integer division. The random number is uniformly distributed between [0,1], where each dimension of the position is the task to be selected, so the dimensions of both velocity and position are m; 2) Calculate the first Fitness of individual particles : ； in, This represents the particle fitness value; all other meanings are the same as above. In the maximization problem, a larger fitness value is preferred. 3) Divide into multiple population topologies: ； ； ； in It is the k-nearest point of the particle. This represents the initial state of population 1. Represents population 1 after the i-th update. Used to determine whether point i is a particle in population 1; other particles besides those in population 1 belong to population 2; 4) Update the inertia weight coefficient : ； in for The upper limit, for The lower bound value, where e represents the current iteration number. This represents the maximum number of iterations. 5) Update the velocity and position of the particles to generate a new swarm; ； ； ； ； in, It is the learning rate of the particles. , , For an improved, fully-fledged learning individual; , They are the first The historical optimal solution for each particle and the optimal solution for the entire swarm; For population 1, the speed update strategy is as follows: The velocity update strategy for population 2 is as follows: updating the velocity of particles in both populations using this formula effectively avoids premature convergence and enhances their global search capability. 6) Adaptive correction of comprehensive learning factors For a particle population X, its particle fitness is Population size N, problem dimension Dim; The goal is to find its comprehensive learning factors. The steps are as follows: Step 1: Assign a linearly increasing learning probability to each particle in the population according to the formula above. ; Step 2: For each particle i, for each dimension, randomly select two particles a and b in the population, and take the dimension value of the particle with higher fitness as the candidate value. Step 3: For each dimension of particle i, take a random number rand(0,1) and multiply it by the value of that dimension. In comparison, if less than Then this dimension The candidate value from the previous step, and vice versa. This represents the historical best value for particle i. Step 4: If all dimension values are the historical best values of particle i, then randomly select two dimensions as candidate values in step 2. The optimal strategy output module outputs the global optimal solution, i.e. the optimal battle strategy, obtained from the multi-population evolution optimization module, and executes it on the battlefield elements.
2. The air combat adversarial game strategy optimization system based on a multi-population adaptive correction evolutionary algorithm according to claim 1, characterized in that: The air combat simulation module constructs a dynamic model of the opposing sides based on the motion laws of integrated air combat battlefield elements and combat mission confrontation: in, Here are the system equations, and the initial state of the system is... , express The state variables of a dimensional game. express The game involves controlling variables; since both sides are unaware of each other's game strategies, we consider the opponent's strategy under extreme intelligence and our own strategy under the worst-case scenario. Therefore, the payoff functions for the dynamic game between the two sides are as follows: In this model, the variable subscripts A and B represent our side and the enemy side, respectively, and the payoff function... Including time period Integral terms , with terminal time Final value term Payment function Including time period Integral terms , with terminal time Final value term , and These are bilateral game strategies where the control variables are the opposing side and the enemy side, respectively.
3. The air combat adversarial game strategy optimization system based on a multi-population adaptive correction evolutionary algorithm according to claim 1, characterized in that: The air combat mission constraint module constructs complex constraints for battlefield adversarial game based on the characteristics of coupled and rapidly changing battlefield background environment: Where E represents the battlefield confrontation normative equality constraint function. The set of, I, denotes the orthogonal inequality constraint function. The set, This represents the final time of the integral in the constraint function. express The final value term at time; Game Time Domain The integral term, and This represents the upper and lower bounds of the control variables in a game.