A vehicle model configuration method based on an evolutionary computing framework

Abstract

The application relates to the field of computer technology application, in particular to a vehicle model configuration method based on an evolution calculation framework, which comprises the following steps: acquiring a take-off and landing task set and vehicle formation data; constructing an initial entity set of a generation evolution framework, each entity corresponding to a configuration relationship between all take-off and landing tasks and vehicle models; performing one round of evolution operation on the current entity set to obtain a new generation entity set; repeating the iteration until a termination condition is met; and outputting a vehicle model configuration result corresponding to an entity with the highest fitness. The method can dynamically balance the global exploration and local development capabilities of the generation evolution framework through adaptive selection of Thompson sampling to destroy and reconstruct operation elements, can improve the resource saving amount of the vehicle model configuration scheme, and can preferentially ensure full coverage of take-off and landing tasks in the iteration process.

Description

Technical Field

[0001] This invention relates to the field of computer technology applications, and in particular to a method for configuring vehicle models based on an evolutionary computing framework. Background Technology

[0002] Aircraft type (model) configuration is a crucial step in airline flight planning, used to assign the appropriate aircraft type to each flight. Different aircraft types differ in terms of seating capacity, operating costs, and cabin layout. The outcome of aircraft type configuration directly impacts the number of passengers a flight can carry and its operating costs, thereby affecting the airline's operating revenue.

[0003] The configuration of vehicle types must meet several constraints, including: each flight must have exactly one vehicle operating (flight coverage constraint); two consecutive flights operated by the same vehicle must meet airport connection conditions and minimum turnaround time requirements (flight connection constraint); the number of vehicles of each type in use must not exceed the number of vehicles of that type in operation (fleet size constraint); the expected number of passengers in each cabin class must not exceed the capacity of the cabin class corresponding to the vehicle type used (cabin capacity constraint); and the number of overnight vehicles of each type at each airport must not exceed the airport's maximum capacity (number of overnight vehicles constraint), etc.

[0004] Existing modeling methods for solving this problem include spatiotemporal network methods and connection network methods, which transform the problem into a network flow model and then use integer programming for exact solutions. However, exact algorithms have high computational complexity and are difficult to handle large-scale real-world scenarios. Therefore, researchers widely employ heuristic algorithms, including generational evolutionary frameworks (i.e., genetic algorithms), simulated annealing, and ant colony optimization. Among these, generational evolutionary frameworks iteratively optimize the entity set through operations such as selection, crossover, and mutation, and are applied to solving large-scale vehicle configuration problems. In the application of generational evolutionary frameworks, the quality of the initial entity set affects the convergence speed and solution quality of the algorithm. Randomly generated initial solutions may violate constraints on flight connections or the number of vehicle types. In the early stages of iteration, mutation operations with strong perturbation capabilities may be needed to increase the diversity of the entity set; in the later stages of iteration, mutation operations with strong local fine-tuning capabilities may be needed to improve the solution quality. In existing methods, the selection of mutation operations usually adopts fixed or random strategies, lacking a mechanism for dynamic adjustment based on the search stage, which makes the algorithm prone to getting trapped in local optima or slow convergence. In applying reinforcement learning methods to operator selection tasks within heuristic algorithms, related research employs methods such as Thompson sampling to achieve adaptive selection by maintaining the probability distribution of operator historical performance. However, the reward signal in these methods typically uses the change in the objective function directly. When the problem simultaneously contains multiple optimization sub-objectives (e.g., takeoff and landing mission coverage versus profit) and these sub-objectives differ by orders of magnitude, a single change in the objective function may cause the reward signal to be dominated by the sub-objective with the larger value, preventing the learning strategy from effectively reflecting the optimization needs of other sub-objectives. Furthermore, existing methods may suffer from adjustment lag when the adaptive decision-making strategy continuously produces inferior solutions, thus affecting the overall stability of the algorithm. Summary of the Invention

[0005] To address the aforementioned technical problems, the technical solution adopted by this invention is as follows:

[0006] This invention provides a method for configuring vehicle models based on an evolutionary computation framework, the method comprising:

[0007] Acquire takeoff and landing mission sets and vehicle formation data.

[0008] Based on the set of takeoff and landing missions and the vehicle formation data, an initial entity set for the generational evolution framework is constructed, wherein each entity in the initial entity set corresponds to the configuration relationship between all takeoff and landing missions and vehicle models.

[0009] An evolutionary operation is performed on the current entity set to obtain a new generation entity set. This evolutionary operation includes: selecting parent entities from the current entity set; performing a recombination operation on the selected parent entities to generate intermediate child entities; determining the deconstruction and reconstruction operators used in the current round using an adaptive operator selection strategy based on Thompson sampling; deconstructing and reconstructing the selected entities using the determined deconstruction and reconstruction operators to generate child entities; calculating the fitness of the child entities; and incorporating the child entities into the new generation entity set. The initial value of the current entity set is the initial entity set.

[0010] The new generation of entity set is used as the current entity set, and the evolution operation is repeated until the criteria for stopping iteration are met.

[0011] Output the vehicle model configuration result corresponding to the entity with the highest fitness value at the end of the iteration.

[0012] This invention is based on Thompson sampling adaptive selection of destruction and reconstruction operators, enabling the mutation operation of the generational evolution framework to dynamically adjust operator preferences according to historical rewards. It emphasizes global exploration in the early stages of iteration and local development in the later stages, thereby effectively improving the resource savings of the final vehicle configuration scheme while satisfying multiple constraints such as flight connections. At the same time, by setting a penalty term in the objective function to prioritize full coverage of take-off and landing tasks, and updating the first parameter of the beta distribution with a second increment according to the hierarchical reward rule when the resource savings increase, the problem of the resource savings signal being completely submerged by the coverage signal is avoided. It has the advantages of fast convergence speed, high solution quality, and strong robustness.

[0013] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of the present invention, nor is it intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 A flowchart illustrating a vehicle model configuration method based on an evolutionary computation framework, provided for an embodiment of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0018] It should be noted that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe the steps as sequential processes, many of these steps can be performed in parallel, concurrently, or simultaneously. Furthermore, the order of the steps can be rearranged. A process can be terminated when its operation is complete, but it may also have additional steps not included in the figures. A process can correspond to a method, function, procedure, subroutine, subroutine, etc.

[0019] This application provides a vehicle type configuration method based on an evolutionary computation framework for generating vehicle type configuration schemes in airline flight plans. A vehicle refers to an aircraft or equivalent aircraft that performs takeoff and landing missions. Vehicle type refers to the classification of a vehicle. Vehicle type configuration is the process of allocating vehicle types to each flight according to the flight schedule during the flight planning stage. The configuration process needs to consider factors such as the available quantity of different vehicle types, seat capacity, and resource consumption per unit flight time. Its objective is typically to minimize operational resource consumption or maximize resource savings. Figure 1 As shown, the method includes the following steps:

[0020] S100 acquires takeoff and landing mission sets and vehicle formation data.

[0021] Specifically, the set of takeoff and landing missions for the vehicle type to be configured can be obtained from the airline's flight planning system. Each takeoff and landing mission in the set includes time information, spatial information, and passenger demand information. The time information includes the scheduled departure and arrival times of the flight; the spatial information includes the departure and arrival airports; and the passenger demand information includes the expected number of passengers in each cabin class, which includes business class and economy class. In this application, takeoff and landing mission is synonymous with flight, referring to a complete flight mission that includes both takeoff and landing events.

[0022] Vehicle fleet data is obtained from the fleet management database. This data includes various vehicle models, the number of available vehicles for each model, seat capacity, and resource consumption per unit flight time. Seat capacity includes the number of business and economy class seats for each vehicle model. Resource consumption per unit flight time characterizes the amount of resources consumed by a particular vehicle model per unit of operating time. This resource consumption includes a comprehensive measure of fuel consumption and maintenance costs. For example, fuel consumption (kg / hour) and maintenance costs (e.g., maintenance cost equivalent resources / hour) are added together according to a preset weight to obtain the resource consumption per unit flight time. This consumption is a preset constant, with different values ​​for different vehicle models. The preset weights are set based on the relative importance of fuel and maintenance in total resource consumption.

[0023] In addition, the minimum turnaround time for each vehicle type at each airport is obtained, as well as the maximum number of overnight vehicles of each vehicle type that each airport can accommodate. The minimum turnaround time refers to the shortest time interval required for a vehicle to complete the landing of a previous flight and then take off for the next flight. This time interval is determined based on the vehicle type and the airport's capacity, for example, 45 minutes or 60 minutes. Each configuration cycle is one operating day. Within each configuration cycle, the actual number of overnight vehicles for each vehicle type and each airport is determined by the configuration plan and does not exceed the stated maximum number of overnight vehicles.

[0024] In this embodiment, the vehicle model configuration must meet the following constraints:

[0025] (1) Flight coverage constraint: Each flight is operated by one and only one vehicle, that is, each flight is assigned to exactly one vehicle in the configuration scheme.

[0026] (2) Flight connection constraints: For two consecutive flights operated by the same vehicle, the following two conditions must be met simultaneously:

[0027] Airport connection: The arrival airport of the previous flight must be the same as the departure airport of the next flight.

[0028] Minimum turnaround time: The scheduled arrival time of the preceding flight plus the minimum turnaround time of this vehicle type at the airport must be less than or equal to the scheduled departure time of the following flight.

[0029] (3) Fleet size constraint: The number of vehicles put into use for each vehicle model shall not exceed the number of available vehicles for that vehicle model.

[0030] (4) Cabin capacity constraints: For each flight, the expected number of passengers in each cabin shall not exceed the seat capacity of the cabin corresponding to the type of vehicle configured.

[0031] (5) Overnight vehicle quantity constraints: For each airport and each vehicle type, the actual number of overnight vehicles shall not exceed the maximum number of overnight vehicles set by the airport for that vehicle type.

[0032] (6) Vehicle balance constraint: During each configuration cycle, for each vehicle type at each airport, the number of vehicles arriving at the airport plus the initial number of vehicles at the airport should be equal to the number of vehicles departing from the airport plus the final number of vehicles at the airport.

[0033] The above constraints collectively define the feasible solution space for the vehicle configuration problem. The generational evolution framework and its various operational steps (initial entity set construction, mutation, repair, feasibility verification, etc.) described in this embodiment all aim to satisfy all of the above constraints.

[0034] S200, based on the set of takeoff and landing missions and the vehicle formation data, construct the initial entity set of the generational evolution framework.

[0035] In this embodiment, the generational evolution framework refers to a genetic algorithm. The entity set refers to the population in the genetic algorithm. An entity refers to an individual.

[0036] To ensure the convergence speed and solution quality of the generational evolution framework, the initial entity set should contain feasible solutions that satisfy constraints such as flight connections and fleet size. Randomly generated initial entity sets may produce a large number of infeasible solutions, which is detrimental to subsequent crossover and mutation operations. This embodiment uses a connection network method combined with a penalty iteration mechanism to generate high-quality elite solutions, which are then expanded into the initial entity set through gradient perturbation. Specifically, the steps include:

[0037] S201: Construct a connection network diagram for each vehicle model.

[0038] For each type of vehicle, a corresponding connection network graph is constructed. The connection network graph uses takeoff and landing tasks as nodes, and directed edges represent the connection relationship between two consecutive flights operated by the same vehicle. In this embodiment, each node in the connection network graph represents a flight, including the flight's departure airport, arrival airport, scheduled departure time, and scheduled arrival time. A directed edge points from the node representing the preceding flight to the node representing the following flight, indicating that the same vehicle is operating these two consecutive flights. If two takeoff and landing tasks satisfy the flight connection constraint, a directed edge is established between the preceding task node and the following task node.

[0039] For the current vehicle type being processed, if two flights meet the following conditions: the arrival airport of the previous flight is the same as the departure airport of the subsequent flight, and the difference between the planned departure time of the subsequent flight and the planned arrival time of the previous flight is greater than or equal to the minimum turnaround time of the vehicle type at the corresponding airport, then a directed edge is established between the two flight nodes.

[0040] The weight of each node is determined based on the amount of resource savings generated by the flight and the corresponding penalty weight (the method for determining the penalty weight will be introduced later).

[0041] The resource saving amount is used to characterize the degree of resource saving obtained by allocating the flight to the aforementioned vehicle type compared to not allocating the flight; that is, when no vehicle type is arranged to operate the flight, the resource saving amount is 0. In this embodiment, the resource saving amount of a single flight is determined based on the theoretical resource benefits and resource consumption of the flight.

[0042] Specifically, resource savings equal the theoretical resource benefits of the flight minus the actual resource consumption during operation. Theoretical resource benefits characterize the theoretical resource value that a flight can generate using the seating capacity of a specific vehicle type; its value equals the smaller of the expected number of passengers in each cabin class and the seating capacity of the assigned vehicle type in that cabin class, multiplied by the unit passenger resource weight corresponding to that cabin class, and then summed. Actual operational resource consumption equals the product of flight time and resource consumption per unit flight time.

[0043] In one example, the resource savings for a single flight are calculated as follows: For each cabin class of the flight, the smaller of the expected number of passengers and the seat capacity is taken. This smaller value is multiplied by the unit passenger resource weight corresponding to that cabin class to obtain the seat resource equivalent for that cabin class. The seat resource equivalents of all cabin classes are summed to obtain the theoretical resource benefit of the flight. The actual operating resource consumption is subtracted from the theoretical resource benefit, and the difference is the resource savings for the flight. Here, the unit passenger resource weight is a preset constant used to convert the seat capacity into a resource value with the same dimension as the corresponding unit flight time resource consumption; the unit flight time resource consumption is an inherent parameter of the vehicle model; and the flight time is determined by the difference between the flight's planned arrival time and planned departure time.

[0044] S202: A feasible solution is obtained by solving the connection network graph using a dynamic programming method based on the multiplier penalty method.

[0045] This step uses an iterative approach to extract a set of paths covering all flights from the connectivity graph. Each path corresponds to a flight sequence of one vehicle, and all paths together constitute a feasible solution. The dynamic programming method based on the multiplier penalty method sets initial penalty weights for all takeoffs and landings. During the iterative solution process, if any takeoff or landing task is not configured, its penalty weight is increased by a preset penalty step size until all takeoffs and landings are configured or the preset maximum number of iterations is reached. The specific iterative process is as follows:

[0046] (1) Initialization: Set an initial penalty weight μ for each flight f. f (0), in this embodiment, μ is taken. f (0) = 50000. This penalty weight is used to adjust the calculation of path weights in dynamic programming. That is, during the initialization phase, the penalty weights of all take-off and landing tasks are set to the same initial value to ensure that all take-off and landing tasks are treated fairly during the initial search of dynamic programming.

[0047] (2) For the current penalty weight set, for each vehicle (i.e., the starting point of each possible path), only consider the flights that are not currently covered by any path, and perform dynamic programming on the connection network graph to find the path with the largest weighted path sum. Let the set of nodes in the connection network graph be V, and each node v∈V corresponds to a take-off and landing task. The recursive method of dynamic programming is as follows: Let dp[v] represent the maximum cumulative weight when arriving at node v from the starting point. Then for each successor node w of node v, there exists a directed edge from node v to node w, indicating that flight v and flight w can be continuously operated by the same vehicle. The state transition equation is: dp[w]=max(dp[w],dp[v]+the comprehensive weight of node w).

[0048] Where dp[w] represents the maximum cumulative weight when reaching node w from the starting point. The overall weight of node w is equal to the penalty weight of node w plus the resource savings of flight w; the penalty weight of node w is the penalty weight μ set for flight w in the current iteration round. w (t); The resource savings of flight w are determined based on the theoretical resource benefits and resource consumption of that flight. By traversing all nodes in ascending order of flight departure time, the maximum cumulative weight of each node can be calculated sequentially, and the optimal predecessor node can be recorded, thus obtaining the optimal path from the starting point to each destination. Each optimal path represents a flight sequence operated by a vehicle.

[0049] (3) After solving the path for all vehicles, count the flights that are not covered by any path and form the set of uncovered flights Uncov(t), where t is the current iteration round.

[0050] (4) If Uncov(t) is empty or the number of iterations reaches the preset maximum number of iterations, the iteration ends and the solution with the most covered flights is output as the feasible solution.

[0051] (5) If Uncov(t) is not empty, then for each flight in the set of uncovered flights, increase the penalty weight of that flight by a preset penalty step size σ: μ f (t+1)=μ f (t)+σ, where σ=8000 in this embodiment. Then return to step (2) and proceed to the next iteration.

[0052] During the above iteration process, the continuous increase in penalty weights increases the path weighting and weight of uncovered flights in subsequent dynamic programming, thereby guiding the algorithm to prioritize covering these flights.

[0053] S203: Compare the resource savings of the feasible solution with those of the pre-trained solution. If the resource savings of the feasible solution are greater than those of the pre-trained solution, then the feasible solution is used as the baseline entity; otherwise, the pre-trained solution is used as the baseline entity. Apply gradient perturbation to the baseline entity to generate multiple different initial entities.

[0054] The pre-trained solution is obtained as follows: using the feasible solution as the initial solution, multiple rounds of random destruction and repair operations are performed, and a simulated annealing mechanism is used to accept new solutions. After iteration, the pre-trained solution is obtained. In this embodiment, the number of iteration rounds can be set to 50 rounds. In each round, a set of deconstruction and reconstruction operators are randomly selected from a variety of deconstruction and reconstruction operators to destroy and repair the current solution, generating a new solution. A simulated annealing mechanism is used to determine whether to accept the new solution: if the objective function value of the new solution is higher than that of the current solution, the new solution is accepted as the current solution for the next round of iteration; otherwise, the inferior solution is accepted with a preset probability, which decreases with the increase of the number of iteration rounds. For example, the initial acceptance probability is 0.8, and it is multiplied by a decay coefficient of 0.99 in each round. After pre-training, the pre-trained solution is obtained, and the number of times each operator is used and the accumulated first parameter value are recorded for setting the initial value of the first parameter in the beta distribution of the adaptive operator selection strategy based on Thompson sampling in the subsequent generational evolution framework. In this embodiment, the various deconstruction operators include at least one of the following:

[0055] (1) A random destructuring operation element, used to randomly and uniformly remove several flights from the parent entity;

[0056] (2) Worst resource saving amount deconstruction operation element, used to sort according to the resource saving amount of each flight and remove the preset number of flights with the lowest resource saving amount;

[0057] (3) Spatiotemporal correlation deconstruction operation element: Randomly select a flight as the reference flight, calculate the spatiotemporal distance between other flights and the reference flight, and remove several flights with the smallest spatiotemporal distance (i.e., the flights most similar to the reference flight). Let the two flights be flight i and flight j, then the formula for calculating the spatiotemporal distance D(i,j) is:

[0058] D(i,j)=|t i -t j | / 60+2×[(1-I(dep i =dep j ))+(1-I(arr i =arr j ))).

[0059] Among them, t i and t j These are the scheduled departure times (in minutes) for flights i and j, divided by 60 to convert to hours; dep i ,dep j These are the departure airports for flights i and j, respectively. i ,arr j Let i and j be the arrival airports for flights i and j, respectively. I(·) is an indicator function that is 1 when the condition is met and 0 otherwise.

[0060] In this formula, the time difference term reflects the proximity of departure times, and the airport matching term reflects the similarity between the departure and arrival airports. The smaller the value of D(i,j), the more similar the two flights are in time and space. Therefore, in the removal operation, several flights with the smallest time and space distance are selected as the removal targets.

[0061] (4) Time window deconstruction operation element, used to randomly select a time window and remove all flights within that time window;

[0062] (5) Air vehicle dismantling operation element, used to randomly clear all flights of a preset number of vehicles.

[0063] The plurality of reconstruction operation elements includes at least one of the following:

[0064] (1) Greedy reconstruction operation element, used to process each flight to be inserted in sequence. If there are multiple insertion positions that satisfy the flight connection constraints, insert the flight into the position with the highest resource saving.

[0065] (2) The second-order regret reconstruction operation element is used to calculate the difference in resource savings between the first and second best insertion positions for each flight to be inserted, and prioritizes the flight with the largest difference in resource savings; where the first best insertion position is the feasible insertion position that increases the resource savings of the current partial solution the most after inserting the flight; the second best insertion position is the feasible insertion position that increases the resource savings the second most.

[0066] (3) A third-order regret reconstruction operator is used to calculate the resource saving difference between the first and third best insertion positions for each flight to be inserted, and to prioritize the flight with the largest resource saving difference. If a flight does not have a third best insertion position, that is, there are fewer than 3 feasible insertion positions, then the resource saving difference is set to a very small preset constant, such as -10. 8 The third optimal insertion position refers to the feasible insertion position that maximizes resource savings.

[0067] To generate multiple differentiated initial entities from a baseline entity, a gradient perturbation strategy is employed. Specifically, a minimum destruction ratio r is set. min and maximum damage ratio r max For example, r min =0.05, r max =0.3, and will be from r min to r max The scope is divided into multiple levels, for example, 5 levels, each corresponding to an increasing damage ratio. For each damage ratio level, the following operations are performed sequentially:

[0068] (1) Randomly remove take-off and landing missions (flights) from the baseline entity at a proportion equal to the damage proportion, and the removed flights constitute the set to be restored;

[0069] (2) Randomly select one of the preset reconstruction operation elements, such as greedy reconstruction operation element, second-order regret reconstruction operation element, and third-order regret reconstruction operation element, and re-insert the flight in the set to be restored into the base entity at the position that satisfies the flight connection constraint to generate a new individual.

[0070] Using the above method, a series of new individuals can be generated that are structurally similar to the baseline entity but with progressively increasing levels of damage. When the damage ratio is low, individuals highly similar to the baseline entity are generated; when the damage ratio is high, individuals with significant structural differences are generated. This approach maintains the high-quality characteristics of the baseline entity while introducing rich diversity into the initial entity set.

[0071] The baseline entity, together with several different entities generated through gradient perturbation, constitutes the initial entity set of the generational evolution framework. The size of the initial entity set is a preset value, for example, set to 50 entities.

[0072] In this embodiment, each entity uses a coding method with the vehicle as the genetic unit. Each genetic bit corresponds to a vehicle, and the value of this genetic bit is the complete flight sequence operated by that vehicle, arranged in ascending order of scheduled departure time. The length of the chromosome is equal to the total number of available vehicles of all vehicle types. A complete chromosome contains the scheduling information of all vehicles, corresponding to a complete vehicle type configuration scheme. This coding method ensures that the number of available vehicles is not exceeded.

[0073] S300, Perform one round of evolution operation on the current entity set to obtain a new generation entity set. The one round of evolution operation includes:

[0074] S301, Select a parent entity from the current entity set, and perform a reorganization operation on the selected parent entity to generate an intermediate child entity.

[0075] The initial value of the current entity set is the initial entity set constructed by S200. This embodiment adopts a strategy that combines elite selection and tournament selection: the top a entities with the highest fitness in the current entity set (e.g., a=2) are retained and directly enter the next generation; tournament selection is performed: b entities are randomly selected (e.g., b=3), and the entity with the highest fitness among them is taken as one of the parents.

[0076] Uniform vehicle-level crossover is performed on the two parent entities, with the vehicle as the smallest genetic unit. A binary mask of the same length as the chromosome is randomly generated. Genes with a mask value of 1 are inherited from the first parent entity, and genes with a mask value of 0 are inherited from the second parent entity, thus generating an intermediate offspring entity. The recombination operation exchanges the entire flight sequence of the vehicle, but does not change the flight connections within the flight sequence. The flight connections of each vehicle are guaranteed by the parents. The recombination operation may result in the same flight being operated by multiple vehicles simultaneously or a flight not being operated by any vehicle. These issues will be addressed by subsequent feasibility guarantee mechanisms.

[0077] S302, employing an adaptive operator selection strategy based on Thompson sampling, determines the deconstruction and reconstruction operators to be used in the current round from a variety of deconstruction and reconstruction operators. Specifically, it includes the following sub-steps:

[0078] S3021 maintains a decision arm for each deconstruction operation element and a decision arm for each reconstruction operation element.

[0079] In this embodiment, there are 5 types of deconstruction operators and 3 types of reconstruction operators, for a total of 8 decision arms.

[0080] Each decision arm maintains a corresponding beta distribution parameter, which includes a first parameter α and a second parameter β.

[0081] The first parameter is used to characterize the cumulative intensity of the positive reward obtained by the decision arm, and the second parameter is a preset constant that is not updated with iteration; the initial values ​​of α and β for all decision arms are 1.

[0082] S3022 Before the current iteration number of the generational evolution framework reaches the preset delayed activation threshold (e.g., 10 generations), a random selection strategy is adopted to uniformly and randomly select from all deconstruction operation meta-decision arms and all reconstruction operation meta-decision arms respectively, that is, each decision arm of the same type has an equal probability of being selected.

[0083] S3023, when the sum of the number of times the same type of decision arm is selected reaches the preset minimum sampling threshold, sample independently from the beta distribution of the corresponding decision arm, select the decision arm with the largest sample value, and obtain the corresponding deconstruction operator and reconstruction operator; use the selected deconstruction operator and reconstruction operator as the operator pair used in the current round.

[0084] Specifically, when the sum of the selection counts of all deconstruction operator decision arms reaches a preset minimum sampling threshold (e.g., 20 times), and the sum of the selection counts of all reconstruction operator decision arms reaches a preset minimum sampling threshold, the Thompson sampling decision phase begins: Independent sampling is performed from the beta distribution of each deconstruction operator decision arm, and the deconstruction operator decision arm with the largest sample value is selected to obtain the corresponding deconstruction operator; independent sampling is performed from the beta distribution of each reconstruction operator decision arm, and the reconstruction operator decision arm with the largest sample value is selected to obtain the corresponding reconstruction operator; the selected deconstruction operator and reconstruction operator are used as the operator pair for the current round.

[0085] The beta distribution of each decision arm is determined by its current first and second parameters. The beta distribution is a continuous probability distribution defined on the interval [0,1], and its probability density function is determined by α and β. As the iteration process progresses, α is continuously updated based on historical rewards, while β remains constant. The mean of the beta distribution reflects the cumulative reward intensity of the decision arm under the current historical information.

[0086] S3024: After performing the mutation operation and calculating the fitness of the offspring entities, update the first parameter of the selected decision arm according to the reward signal. The selected decision arms include the decision arms of the deconstruction operator and the decision arms of the reconstruction operator.

[0087] The rules for generating the reward signal include:

[0088] Step 1: Obtain the take-off and landing mission coverage and resource savings of the child entity.

[0089] In each generation of evolutionary operations, after completing the mutation operation (S303) and generating a child entity, the takeoff and landing mission coverage and resource savings of the child entity are calculated. The takeoff and landing mission coverage is defined as the ratio of the number of flights with configured vehicle types in the current child entity to the total number of flights in the takeoff and landing mission set.

[0090] Step 2: If the take-off and landing task coverage of the offspring entity is greater than the take-off and landing task coverage of the offspring entity before this mutation operation, the first parameter of the selected decision arm will be increased by the first increment, that is, the first positive reward will be generated.

[0091] Step 3: If the resource savings of the offspring entity are greater than the resource savings of the current global optimum, then the first parameter of the selected decision arm is increased by a second increment, and the second increment is greater than the first increment, thus generating a second positive reward. The current global optimum refers to the configuration scheme corresponding to the individual with the highest resource savings discovered from the start of the algorithm to the current iteration. This global optimum is continuously updated during the evolution process: whenever a new individual's resource savings exceed the current recorded value, this new individual replaces the original individual to become the new global optimum.

[0092] In one illustrative embodiment, the ratio of the first increment to the second increment is configured as 1:10.

[0093] In the above rules, the first increment corresponds to a reward for improved coverage, and the second increment corresponds to a reward for improved resource savings, with the second increment being greater than the first. This is because coverage, i.e., whether the solution covers all takeoff and landing tasks, is a hard indicator at the constraint satisfaction level and must be achieved first, but its improvement space is limited (it cannot be further improved after reaching 100%). Resource savings, on the other hand, is a soft indicator at the optimization quality level, with greater improvement potential and a more profound impact on the overall system efficiency. Therefore, giving a higher reward weight to improvements in resource savings can effectively guide the algorithm to continuously evolve towards higher quality while ensuring the feasibility of the solution.

[0094] Through the above rules, the first increment (a smaller increment) drives the algorithm to improve takeoff and landing mission coverage, ensuring that the generated configuration scheme meets all coverage requirements at the constraint level. The second increment (a larger increment) guides the algorithm to continuously optimize resource savings after the coverage target is met, making the final scheme both feasible and economical. There is no conflict between the two levels of rewards: the coverage reward applies to the constraint satisfaction stage, and the resource savings reward applies to the quality optimization stage. Both jointly drive the algorithm to search for high-quality vehicle model configuration schemes within the feasible region.

[0095] S303, using the determined destructuring and reconstruction operators, the selected intermediate child entity is destructed and reconstructed to generate a child entity. The reconstruction operation re-inserts the removed take-off and landing tasks into positions that satisfy preset insertion constraints.

[0096] Specifically, firstly, based on the selected deconstruction operation element, some takeoff and landing tasks are removed from the intermediate child entity (the removal rules are defined in the previous section on each deconstruction operation element). The removed tasks constitute the set to be restored. Then, based on the selected reconstruction operation element, each takeoff and landing task in the set to be restored is re-inserted into a feasible position in the intermediate child entity (the insertion strategy is defined in the previous section on each reconstruction operation element). The feasible position refers to the position where, after inserting the takeoff and landing task, the current partial scheme still satisfies the flight connection constraints.

[0097] Once all tasks in the set to be restored have been successfully inserted, a complete final child entity is generated. If individual tasks cannot be inserted due to constraints during the insertion process, they will be handled by the subsequent feasibility guarantee mechanism. However, the design of the reconstruction operation element in this embodiment ensures that a feasible solution can be generated directly in most cases.

[0098] Through the above deconstruction and reconstruction operations, the structure of the intermediate offspring entities is purposefully disturbed and repaired, thereby introducing new flight combinations while retaining the high-quality characteristics of the parent generation, and realizing the mutation function of the genetic algorithm.

[0099] S304, calculate the fitness of the child entity and include the child entity in the new generation entity set.

[0100] In this embodiment, the fitness of an entity is calculated based on an objective function. The objective function includes a flight coverage penalty term and a resource saving term, specifically defined as follows:

[0101] The flight coverage penalty is equal to the number of flights without the configured vehicle type multiplied by the coverage penalty coefficient. The coverage penalty coefficient is a preset constant, for example, 10. 8 The value of this term is such that the minimum possible value of the flight coverage penalty term is greater than the maximum possible value of the individual flight resource saving term. With this setting, any entity that has not achieved full flight coverage will have a lower objective function value than entities that have achieved full coverage, thus ensuring that the algorithm prioritizes achieving full flight coverage. The resource saving term is equal to the sum of the resource savings of all configured flights.

[0102] The objective function value equals the resource saving term minus the flight coverage penalty term.

[0103] Fitness is positively correlated with the objective function value. In this embodiment, the objective function value is directly used as the fitness value, and the generational evolution framework iteratively optimizes the process with the goal of maximizing fitness.

[0104] After calculating the fitness of the child entity, the child entity is added to the new generation entity set. The new generation entity set has the same size as the current entity set.

[0105] Furthermore, the method provided in this embodiment also includes a three-stage adaptive adjustment mechanism, a parameter adaptive adjustment mechanism, a feasibility guarantee mechanism, and a robustness guarantee mechanism.

[0106] S305, a three-stage adaptive adjustment mechanism.

[0107] Building upon S3021's maintenance of beta distribution parameters for each decision arm, this embodiment also provides a three-stage adaptive adjustment mechanism. This mechanism sets different initial biases for the first parameters of each decision arm at different search stages, thereby guiding the generational evolution framework to dynamically adjust its emphasis on exploration and development during the iterative process. This mechanism complements the adaptive operator selection strategy of S302 and works in conjunction with the Thompson sampling decision and the discount decay operation described later in S3024.

[0108] Specifically, the entire iterative process of the generational evolution framework is divided into three stages according to the proportion of iteration rounds:

[0109] Phase 1 (first 30% of iterations): The initial first parameter value in this phase is inherited from the α value of each decision arm at the end of the early pre-training phase.

[0110] Phase 2 (30% to 70% of iterations): The initial first parameter value in this phase is inherited from the α value of each decision arm at the end of the pre-training phase.

[0111] The third stage (the last 30% of iterations): The initial first parameter value in this stage is inherited from the α value of each decision arm at the end of the later stage of pre-training.

[0112] At each stage transition, the accumulated first parameter value of the decision arm from the previous stage is saved. Meanwhile, the discount decay operation described later is executed independently within each stage to ensure appropriate forgetting of historical rewards and prevent high α values ​​from becoming entrenched in the long term.

[0113] Through the above three-stage adjustments, the algorithm can focus on global exploration in the early stage, balance exploration and development in the middle stage, and focus on local fine search in the later stage, thereby improving the quality of the final solution and the convergence efficiency.

[0114] During pre-training, the entire pre-training iteration is also divided into three phases proportionally to simulate the search process of the main iteration, making the beta distribution parameters obtained through pre-training more adaptive. Specifically, if the total number of pre-training rounds is Tpre (e.g., 50 rounds), then: at the beginning of each pre-training phase, the first parameter value of each decision arm is set to 1. After each pre-training phase, the final α value of each decision arm is recorded as the basis for the initial values ​​of each phase in the three-phase design of the main iteration.

[0115] S306, parameter adaptive adjustment mechanism.

[0116] Furthermore, the adaptive operator selection strategy also includes a parameter adaptive adjustment mechanism. This mechanism is used to prevent early positive reward preferences from becoming fixed, enabling the algorithm to dynamically adjust operator selection preferences as the search progresses. Specifically, it includes the following operations:

[0117] When the sum of the first parameters of the decision arms corresponding to all deconstruction operation elements in a certain search phase reaches a first preset threshold, or when the sum of the first parameters of the decision arms corresponding to all reconstruction operation elements in a certain search phase reaches a second preset threshold, a discount decay operation is performed on the first parameter of each decision arm. The second parameter of each decision arm is fixed as a preset constant, for example, 1, and is not updated with iteration. The specific steps of the discount decay operation are as follows: multiply the first parameter by a preset discount factor, for example, 0.88, to obtain a product value; compare the product value with a preset minimum value of the first parameter, for example, 1; if the product value is less than the minimum value of the first parameter, update the first parameter to the minimum value of the first parameter; otherwise, update the first parameter to the product value.

[0118] The first and second preset thresholds can be preset according to the number of operators and the desired discount decay trigger frequency. As an implementable approach, the first preset threshold is set to 150; the second preset threshold is equal to 90. When the sum of the first parameters of all deconstruction operators or all reconstruction operators reaches the corresponding threshold, the discount decay operation is triggered, multiplying the corresponding first parameter by a discount factor such as 0.88, thereby avoiding the operator preference from becoming fixed due to the infinite growth of parameters.

[0119] In another embodiment, the first preset threshold and the second preset threshold can be set to the same constant, such as 100. The specific values ​​can be determined through a small number of experiments based on the algorithm's convergence speed and the quality of the solution, or adaptively adjusted according to the total number of iterations. This application does not limit this.

[0120] Through the above operations, when the sum of the first parameters of the decision arms corresponding to all deconstruction or reconstruction operations at a certain stage reaches a preset threshold, discounting is applied to reduce the cumulative impact of historical rewards and prevent premature decision solidification. Simultaneously, the second parameter remains fixed as a constant, ensuring that the sampling probability of the decision arm depends only on the accumulation of positive rewards, avoiding excessive penalties due to inferior solutions and maintaining the algorithm's exploratory capabilities.

[0121] S307, Feasibility Assurance Mechanism.

[0122] In the entity set initialization, recombination, and mutation operations within the generational evolution framework, newly generated entities may violate flight connection constraints or flight coverage constraints. To ensure the feasibility of new entities, the following operations are automatically performed:

[0123] (1) Statistics on uncovered flights: Statistics on flights in the current entity that are not configured with vehicle models, forming a list to be repaired.

[0124] (2) Connection detection: Traverse the flight sequence operated by each vehicle in the current entity and check the connection feasibility between each pair of adjacent flights. If the adjacent flights do not meet the constraints such as flight connection constraints and cabin capacity constraints, the subsequent flight is removed from the sequence and added to the list to be repaired.

[0125] (3) Repair: All flights in the list to be repaired are re-inserted into the current entity through the third-order regret reconstruction operation element, so that the repaired entity satisfies all constraints.

[0126] (4) Feasibility verification: Verify whether the repaired entity meets all the constraints. If there are still flights that violate the constraints, such as flights that are not covered or do not meet the connection conditions, then the infeasible flights are extracted again and added to the list to be repaired, and the third-order regret reconstruction operation is called repeatedly to repair them until the current entity meets all the constraints.

[0127] S308, Robustness Guarantee Mechanism.

[0128] To prevent the adaptive operator selection strategy from continuously generating suboptimal solutions due to noise or model bias, this embodiment also provides a robustness guarantee mechanism to monitor the performance of the adaptive strategy and switch to a random selection strategy when necessary. Specifically, it includes the following steps:

[0129] (1) Every preset number of security check iterations, such as 25 generations, calculate the average fitness improvement rate obtained by using the adaptive operand selection strategy in the current period. The average fitness improvement rate is calculated by summing the fitness improvement of each generation of child entities relative to the parent entity in the current period, and then dividing by the number of times the adaptive operand selection strategy is used in the current period.

[0130] Simultaneously, the fitness improvement rate when the random selection strategy is executed is recorded. The random selection strategy is executed randomly with a preset probability, such as 10%, in each iteration, and the average fitness improvement rate of the random selection strategy is calculated for use in collecting reference data.

[0131] (2) If, within a series of preset periods, such as three consecutive periods, the average fitness improvement rate obtained by using the adaptive operator selection strategy is lower than the current reference value multiplied by a preset proportional coefficient, such as 0.75, then a circuit breaker operation is triggered. The circuit breaker operation includes: pausing the use of the adaptive operator selection strategy and switching to a random selection strategy. After switching to the random selection strategy, the average fitness improvement rate obtained by the currently fully executed random strategy is used as the reference value for recovery judgment.

[0132] (3) After switching to the random selection strategy, the recovery verification process of the adaptive operator selection strategy is initiated once every 25 generations at a preset recovery attempt interval: During the current recovery attempt period, the fitness improvement rate obtained by the random selection strategy and the fitness improvement rate obtained by the adaptive operator selection strategy are recorded in parallel. If the average fitness improvement rate obtained by the adaptive operator selection strategy is greater than or equal to the average fitness improvement rate obtained by the random selection strategy multiplied by the preset proportional coefficient, for example, 0.75, the adaptive operator selection strategy is reactivated, and the operator selection based on Thompson sampling is restored. Otherwise, the random selection strategy is maintained, and the process waits for the next recovery attempt period.

[0133] The recovery attempt interval and the security check interval are independent of each other. They can both be set to the same value, such as 25 generations, or they can be configured separately according to actual needs. When recovery verification fails, the algorithm continues to run with a random selection strategy, and in the next recovery attempt cycle, which starts from the end of the current recovery verification cycle, it will attempt recovery again after one recovery attempt interval.

[0134] Through the above mechanism, even if the adaptive strategy temporarily fails due to persistently low improvement, the algorithm can automatically switch to random exploration and verify the performance of the adaptive strategy in each recovery attempt cycle. The adaptive strategy is only reactivated when its performance meets the target; otherwise, the random selection strategy is maintained and the algorithm waits for the next verification, thereby achieving safe recovery of the adaptive strategy while ensuring the stability of the algorithm.

[0135] S400, take the new generation entity set as the current entity set, repeat the evolution operation until the judgment criteria for stopping iteration are met.

[0136] The criteria for stopping the iteration include at least one of the following:

[0137] The preset maximum number of iterations is reached, for example, 500 generations.

[0138] The take-off and landing task coverage of the best entity in the entity set reaches 100%, and the resource savings do not increase within a consecutive preset number of generations.

[0139] When any of the above judgment criteria are met, the iteration stops and step S500 is executed.

[0140] S500 outputs the vehicle model configuration result corresponding to the entity with the highest fitness in the entity set obtained at the end of the iteration as the optimal vehicle model configuration result.

[0141] The vehicle configuration results include the vehicle configuration information for each takeoff and landing mission (flight) and the flight sequence operated by each vehicle.

[0142] After obtaining the optimal vehicle configuration result, this result will be used as the basis for capacity allocation decisions during the airline's flight planning phase. Specifically, the configuration result can be structured and output as a capacity planning document in one or more of the following formats:

[0143] (1) Flight-Vehicle Type Assignment Table: Using the flight number as the row index, it records the vehicle type and its seat capacity configuration assigned to each flight, which is used to generate the capacity information of the flight timetable and reservation system;

[0144] (2) Vehicle schedule: Using the vehicle (aircraft) as the row index, it records the complete flight sequence operated by each vehicle within the planned period, including the take-off and landing times of each flight, the take-off and landing airports, and the transit time of connecting flights, which is used to guide the maintenance scheduling and flight crew matching.

[0145] (3) Resource occupancy summary table: Statistics on takeoffs and landings and overnight parking of each type of vehicle by airport and time period, used to match and verify with airport capacity and overnight parking space resources.

[0146] The capacity plan document is distributed to the airline's operations control center as a baseline scheme for the flight schedule execution phase. Based on this, the operations control center can further carry out subsequent operational processes such as crew scheduling, passenger reservations, and allocation of ground support resources.

[0147] In this embodiment, the above-mentioned optimal configuration results can also be converted into a visual aeronautical chart or data report, and presented to the capacity scheduling personnel through the airline's flight plan management interface for manual review or fine-tuning. In the automated capacity scheduling mode without manual intervention, the configuration results are directly used as the capacity configuration data of the final flight plan and submitted to the flight information system to start the subsequent flight execution process.

[0148] Steps S100 to S500 described in this embodiment can be implemented using computer program instructions. Those skilled in the art can write corresponding computer programs using any applicable programming language, such as Python, C++, or Java, based on the above process description, and run them on general-purpose computer equipment or dedicated embedded equipment to execute the above-described vehicle model configuration method.

[0149] Through the above application methods, the vehicle model configuration results generated by the method described in this application can be directly embedded into the existing airline's flight planning and operation control workflow, and have plug-and-play engineering adaptability.

[0150] This application has at least the following beneficial effects:

[0151] 1. The initial entity set is of high quality, resulting in fast algorithm convergence.

[0152] By combining the connection network method with iterative penalized dynamic programming based on the multiplier penalty method, feasible solutions satisfying all constraints such as flight coverage, flight connections, and fleet size can be quickly generated as elite solutions. Based on this, gradient perturbation is used to expand the initial entity set into a diverse set. Compared with random initialization methods, the initial entity set generated by this method has a higher proportion of feasible solutions, effectively reducing the computational overhead of repairing infeasible solutions in early iterations of the generational evolution framework, thereby accelerating convergence.

[0153] 2. Mutation operators can adaptively select and balance exploration and exploitation capabilities.

[0154] Multiple deconstruction and reconstruction operators were designed for mutation operations within the generational evolution framework. An adaptive operator selection strategy based on Thompson sampling was adopted, dynamically adjusting the selection probability according to the historical rewards of each operator during iteration. This mechanism prioritizes strong deconstruction operators in the early stages of the algorithm to increase entity set diversity, and prioritizes fine reconstruction operators in the later stages to improve local search capabilities. This achieves an adaptive balance between global exploration and local development, avoiding the premature convergence or low search efficiency problems caused by traditional fixed operator strategies or random selection strategies.

[0155] 3. Decoupling of multi-objective reward signals, resulting in a clear learning direction.

[0156] To address the challenge of optimizing both takeoff and landing mission coverage and resource savings in vehicle configuration, a two-tiered reward system was designed: a smaller reward for improved takeoff and landing mission coverage and a larger reward for improved resource savings. This decoupling mechanism ensures that the Thompson sampling-based learning process considers both full flight coverage and resource savings. The method separates the resource savings signal from the objective function, providing a second incremental reward based on the global optimal improvement in resource savings. This avoids the objective function being dominated by the flight coverage penalty term (whose penalty coefficient is much larger than the range of individual flight resource savings terms), thus preventing the resource savings optimization signal from being overwhelmed. This ensures that the improvement direction chosen by the operator aligns with the actual problem requirements.

[0157] 4. Parameters are adaptively adjusted to suppress the solidification of historical preferences.

[0158] A discount decay operation is introduced into the beta parameter update of Thompson sampling. When a preset condition is met, the first parameter of each decision arm is multiplied by a discount factor, and a minimum value limit is set. This mechanism can reduce the influence of historical cumulative rewards on the current decision and prevent operator preferences from becoming fixed in the long term due to early accidental positive rewards. At the same time, fixing the second parameter to a constant avoids excessive penalty for inferior solutions, thereby maintaining the diversity of the algorithm in the later stages of the search and further improving the probability of convergence to the global optimum.

[0159] 5. Robustness guarantee mechanism to improve algorithm stability

[0160] A circuit breaker and recovery mechanism is designed to periodically monitor the average fitness improvement rate of the adaptive operator selection strategy. When the improvement rate falls below the average improvement rate of the random selection strategy multiplied by a preset scaling factor, the algorithm automatically switches to the random selection strategy and reactivates the adaptive strategy after performance recovers. This mechanism effectively addresses situations where the adaptive operator selection strategy temporarily fails due to persistently low improvement, ensuring the overall optimization performance of the algorithm under various operating conditions.

[0161] 6. The feasibility assurance mechanism is complete, ensuring that all solutions meet actual operational constraints.

[0162] After entity set initialization, crossover, mutation and other operations, the system automatically performs out-of-coverage flight statistics, connection detection, repair based on the third-order regret operator and feasibility verification. This ensures that the final output vehicle model configuration result meets all constraints such as flight coverage, flight connection, fleet size, cabin capacity, number of overnight vehicles and vehicle balance, and is practically operable.

[0163] 7. Wide range of applications, adaptable to flight schedules of different sizes.

[0164] By leveraging the universality of coding and operator design, this method can be directly applied to the configuration of vehicle models for different airlines and flight scales.

[0165] This invention also provides an electronic device, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being configured to perform the method described in this invention.

[0166] This invention also provides a computer-readable storage medium storing computer-executable instructions for performing the methods described in this invention.

[0167] It should be understood that the various forms of processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this invention can be achieved, and this is not limited herein.

[0168] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for configuring vehicle models based on an evolutionary computation framework, characterized in that, The method includes: Acquire takeoff and landing mission sets and vehicle formation data; Based on the set of takeoff and landing missions and the vehicle formation data, an initial entity set for the generational evolution framework is constructed, wherein each entity in the initial entity set corresponds to the configuration relationship between all takeoff and landing missions and vehicle models; An evolutionary operation is performed on the current entity set to obtain a new generation entity set. This evolutionary operation includes: selecting a parent entity from the current entity set; performing a recombination operation on the selected parent entity to generate intermediate child entities; determining the deconstruction and reconstruction operators used in the current round using an adaptive operator selection strategy based on Thompson sampling; deconstructing and reconstructing the selected intermediate child entities using the determined deconstruction and reconstruction operators to generate child entities; calculating the fitness of the child entities and incorporating them into the new generation entity set; using the new generation entity set as the current entity set, repeating the evolutionary operation until the criteria for stopping iteration are met; the initial value of the current entity set is the initial entity set. Output the vehicle model configuration result corresponding to the entity with the highest fitness value at the end of the iteration.

2. The method according to claim 1, characterized in that, Based on the aforementioned set of takeoff and landing missions and vehicle formation data, an initial entity set for the generational evolution framework is constructed, specifically including: A connection network diagram is constructed for each type of carrier vehicle. The connection network diagram uses take-off and landing missions as nodes and directed edges to represent the connection relationship between two consecutive take-off and landing missions performed by the same carrier vehicle. The connection network graph is solved using a dynamic programming method based on the multiplier penalty method to obtain a feasible solution. The dynamic programming method based on the multiplier penalty method sets an initial penalty weight for all take-off and landing tasks. If there are any unconfigured take-off and landing tasks during the iterative solution process, the penalty weight of the take-off and landing task is increased by a preset penalty step size until all take-off and landing tasks are configured or the preset maximum number of iterations is reached. Compare the resource savings of the feasible solution with those of the pre-trained solution. If the resource savings of the feasible solution are greater than those of the pre-trained solution, then the feasible solution is used as the baseline entity; otherwise, the pre-trained solution is used as the baseline entity. Gradient perturbation is applied to the reference entity to generate multiple different initial entities, and the reference entity and the multiple different initial entities together constitute the initial entity set.

3. The method according to claim 1, characterized in that, The method of employing an adaptive operator selection strategy based on Thompson sampling to determine the deconstruction and reconstruction operators used in the current round specifically includes: A decision arm is maintained for each deconstruction operation element, and a decision arm is maintained for each reconstruction operation element; each decision arm corresponds to a beta distribution parameter at each stage, and the beta distribution parameter includes a first parameter and a second parameter; wherein, the first parameter is used to characterize the cumulative intensity of the positive reward obtained by the decision arm, and the second parameter is a preset constant; Before the current iteration number of the generational evolution framework reaches the preset delayed activation threshold, a random selection strategy is adopted to uniformly and randomly select from all deconstruction operation meta-decision arms and all reconstruction operation meta-decision arms respectively. When the sum of the number of times the same type of decision arm is selected reaches the preset minimum sampling threshold, independent sampling is performed from the beta distribution of the corresponding decision arm, and the decision arm with the largest sample value is selected to obtain the corresponding deconstruction operator and reconstruction operator; the selected deconstruction operator and reconstruction operator are used as the operator pair used in the current round. After performing the mutation operation and calculating the fitness of the offspring entities, the first parameter of the selected decision arm is updated based on the reward signal.

4. The method according to claim 3, characterized in that, The rules for generating the reward signal include: Obtain the take-off and landing mission coverage and resource savings of the child entity; the take-off and landing mission coverage is the ratio of the number of take-off and landing missions of the configured vehicle model to the total number of take-off and landing missions in the take-off and landing mission set; the resource savings are determined based on the theoretical resource benefits and resource consumption of all take-off and landing missions in the child entity. If the take-off and landing mission coverage of the child entity is greater than the take-off and landing mission coverage of the child entity before this mutation operation, then the first parameter of the selected decision arm is increased by the first increment. If the resource savings of the offspring entity are greater than the resource savings of the current global optimal solution, then the first parameter of the selected decision arm is increased by the second increment, and the second increment is greater than the first increment.

5. The method according to claim 3, characterized in that, The adaptive operand selection strategy also includes: In the search phase corresponding to the iterative process of the generational evolution framework, when the sum of the first parameters of the decision arms corresponding to all deconstruction operation elements in a certain search phase reaches a first preset threshold, a discount decay operation is performed on the first parameter of the decision arm of each deconstruction operation element; when the sum of the first parameters of the decision arms corresponding to all reconstruction operation elements in a certain search phase reaches a second preset threshold, a discount decay operation is performed on the first parameter of the decision arm of each reconstruction operation element. The discount attenuation operation is as follows: multiply the first parameter by a preset discount factor; if the product is less than the preset minimum value of the first parameter, then set the first parameter to the minimum value of the first parameter; otherwise, set it to the product.

6. The method according to claim 1, characterized in that, The fitness is calculated based on an objective function, which includes a takeoff and landing mission coverage penalty term and a resource saving term. The takeoff and landing mission coverage penalty term is the number of takeoff and landing missions without a configured vehicle type multiplied by a coverage penalty coefficient. The value of the coverage penalty coefficient is such that the minimum possible value of the takeoff and landing mission coverage penalty term is greater than the maximum possible value of the resource saving term for a single flight. The resource saving term is determined based on the theoretical resource benefits and resource consumption of each flight.

7. The method according to claim 1, characterized in that, The method further includes: Every preset number of security check iterations, calculate the average fitness improvement rate obtained by using the adaptive operand selection strategy in the current period; In each iteration, a random selection strategy is executed with a preset probability, and the average fitness improvement rate of the random selection strategy is collected as reference data. If the average fitness improvement rate of the adaptive operand selection strategy is lower than the reference data multiplied by a preset ratio coefficient for multiple consecutive preset periods, a circuit breaker operation is triggered to suspend the use of the adaptive operand selection strategy and switch to a random selection strategy. After switching to the random selection strategy, the recovery verification process is started at a preset recovery attempt interval. During the current cycle of the recovery verification, the adaptive operand selection strategy is temporarily activated. If its average fitness improvement rate is not lower than the average fitness improvement rate of the random selection strategy multiplied by the preset proportional coefficient, the adaptive operand selection strategy is reactivated; otherwise, the random selection strategy is maintained.

Images