Evolution method based on memory annual ring learning

Abstract

The invention relates to the technical field of artificial intelligence and evolutionary computing, and particularly discloses a memory annual ring learning-based evolutionary method, which comprises the following steps of: locking elite groups, extracting core declusters, coding task constraint fingerprints, backtracking efficient operators, drawing performance operator portraits, constructing memory annual ring units and storing the memory annual ring units in a memory annual ring library. When facing a new task, analyzing and generating a to-be-matched task fingerprint, matching a historical memory annual ring unit, extracting a core declustering and efficiency operator portrait, remodeling a population basis, generating a dynamic operator weight table, and executing dual-channel collaborative guide evolution; according to the method, through collaborative optimization of the solution space and the strategy space, the convergence speed and the solving performance of the algorithm in a new environment are remarkably improved, conversion from blind exploration to experience guidance is achieved, and the challenges of knowledge migration and experience reuse in dynamic change or staged tasks of a traditional evolutionary algorithm are solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence and evolutionary computing, in particular to an evolutionary method based on memory ring learning. BACKGROUND

[0002] Evolutionary algorithms exhibit strong global search capability when solving complex optimization problems, however, when faced with dynamic changes or phased tasks, traditional evolutionary algorithms generally face the core challenges of knowledge transfer and experience reuse. Whenever the target or constraints of the task change, the algorithm often needs to start a new round of random search from scratch, and cannot effectively utilize the valuable experience accumulated in solving previous similar tasks, which leads to a large amount of computing resources being used for repetitive exploration, seriously affecting the convergence efficiency of the algorithm and the adaptability to environmental changes.

[0003] The current industry generally adopted solution or traditional method, mostly focuses on maintaining population diversity in the evolution process of a single task or introducing random individuals when the task changes to cope with environmental changes. Some more advanced methods may save the historical elite individuals and reintroduce them into the population in the new environment, but this is only a simple reuse of high-quality solutions. These methods lack a systematic mechanism to analyze and understand the deep relationship between task characteristics, the structure of high-quality solutions, and efficient search strategies, so they cannot accurately match and call the most relevant historical experience according to the characteristics of the new task.

[0004] The disadvantage of the traditional method is that the granularity of knowledge reuse is too rough and lacks pertinence. Simply reusing old good individuals cannot guarantee that these individuals will still be competitive in the new task environment, and may even mislead the search direction. More importantly, traditional methods ignore the learning and transfer of search strategies, that is, they fail to identify which type of evolutionary operator performs better on which type of problem. This mode of only remembering the answer but not the method makes the algorithm still have to blindly and equally try various evolutionary operators when facing new problems, greatly limiting its convergence speed and solving performance in the new stage. SUMMARY

[0005] In view of this, in order to solve the problems proposed in the background art, an evolutionary method based on memory ring learning is proposed.

[0006] The purpose of the present application can be achieved by the following technical solutions: The present application provides an evolutionary method based on memory ring learning, comprising the following steps: S1, locking the elite group in the current generation population and grouping calculation according to the solution structure characteristics, selecting representative individuals from each group to generate a core solution cluster.

[0007] S2. Extract the key constraint parameters and performance target weights corresponding to the task stage when the core decluster is generated, and serialize and perform identifier operations on them to encode the task constraint fingerprint.

[0008] S3. Backtrack the evolutionary history of the task stages associated with the task constraint fingerprint, count the number of times the evolutionary operator generates offspring that are superior to the parent individual to calculate the performance contribution, and screen out the high-efficiency operators and their contribution values ​​to draw a performance operator profile.

[0009] S4. Forcefully bind and encapsulate the core declustering, task constraint fingerprint, and performance operator profile to construct memory ring units, and store the memory ring units in the memory ring library.

[0010] S5. Obtain the new key constraint parameters and performance target weights for the new task stage, and use the same operation rules as the encoding task constraint fingerprint to parse and generate the task fingerprint to be matched.

[0011] S6. Compare the fingerprint of the task to be matched with the task constraint fingerprint of each memory ring unit in the memory ring library, identify and locate the historical memory ring unit with the highest resonance degree, so as to match and extract the resonance ring data package.

[0012] S7. Extract the core solution cluster encapsulated within the resonant tree ring data packet and use it to replace the individual with the lowest fitness ranking in the current population, so as to apply the core solution cluster to reshape and form the foundation of the reshaped population.

[0013] S8. Synchronously parse the performance operator profile encapsulated within the resonant tree ring data package, and weight the selection probability of the high-efficiency operators recorded in the performance operator profile to generate a dynamic operator weight table using the performance operator profile.

[0014] S9. Using the reshaped population base as the starting population, and selecting evolutionary operators for subsequent iterations based on the weighted probabilities defined in the dynamic operator weight table, to execute dual-channel collaborative guided evolution.

[0015] Compared with the prior art, the embodiments of the present invention have at least the following advantages or beneficial effects: (1) The present invention provides a high-quality starting point for the evolutionary process by directly injecting the core solution cluster extracted from similar historical tasks into the current population at the beginning of the new task phase, effectively avoiding the algorithm from wasting a lot of computing resources due to random exploration in the early stage of the search, thereby significantly accelerating the convergence speed of the algorithm in the new environment.

[0016] (2) By directly injecting the core solution clusters extracted from similar historical tasks into the current population at the beginning of a new task phase, this invention provides a high-quality starting point for the evolutionary process, effectively avoiding the waste of a lot of computing resources due to random exploration in the early stage of the search, thereby significantly accelerating the convergence speed of the algorithm in the new environment.

[0017] (3) By retrospectively analyzing historical data, this invention draws and reuses profiles of efficient operators, enabling the algorithm to dynamically adjust the selection probability of evolutionary operators when facing new tasks, and to give priority to search strategies that have been proven effective in history, thereby realizing the transformation from "blind exploration" to "experience guidance" and improving the efficiency and directionality of the search process.

[0018] (4) This invention achieves deep optimization of the evolutionary process through the dual-channel collaborative guidance of the solution space and the policy space. On the one hand, it directly improves the quality of the solution by reshaping the population base, and on the other hand, it optimizes the search behavior through the dynamic operator weight table. The two work together to form a powerful synergy, which jointly promotes the algorithm to quickly and efficiently approach the optimal solution in the new task stage, and improves the overall intelligence level and solution performance of the evolutionary method. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the 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.

[0020] Figure 1 This is a schematic diagram of the method steps of the present invention.

[0021] Figure 2 This is a diagram of the memory ring learning evolution system architecture of the present invention. Detailed Implementation

[0022] 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.

[0023] Please see Figure 1 and Figure 2 The present invention provides an evolutionary method based on memory ring learning, including: S1, locking the elite group in the current generation population and performing group calculations based on solution structure characteristics, and selecting representative individuals from each group to generate core solution clusters.

[0024] In a specific embodiment of the present invention, the specific steps of locking the elite group in the current generation population and performing group calculations based on the solution structure characteristics, and selecting representative individuals from each group to generate the core solution cluster include: when the evolutionary process reaches a preset convergence plateau period or when an external task phase switching signal is received, locking the individuals with the top one-third fitness in the current generation population to form an elite group.

[0025] It's important to note that, firstly, the system continuously monitors the evolutionary process, determining whether to trigger the generation of a core cluster in two ways. The first method involves determining if the evolutionary process has entered a convergence plateau. The system records the changes in the fitness of the best individuals in the population over recent generations. If the rate of improvement in the best fitness falls below a preset small threshold for a specified number of generations, it is determined to have entered a convergence plateau. The second method involves receiving an external task phase switching signal. This is a clear instruction issued by an external control system, indicating a change in the optimization objective or environmental constraints. Once either condition is met, the system immediately locks onto the top one-third of the individuals in the current generation based on fitness, forming an elite group.

[0026] It should also be noted that identifying the top third of the individuals in terms of fitness in the current generation to form an elite group is a reasonable and effective strategy in evolutionary algorithms. This ratio considers both the effectiveness of elite preservation and the balance between computational resources and efficiency, while also possessing a degree of flexibility and task specificity. Although the specific ratio may need to be adjusted based on actual applications, the top third, as an empirical setting, is reasonable in most cases.

[0027] The structural information of each individual solution in the elite group is transformed into a numerical vector, and a clustering algorithm is used to group individuals with similar performance and structure into the same group.

[0028] It should be noted that, next, the system analyzes each individual solution within the elite group. The core of this process is to group and calculate based on the structural characteristics of the individual solutions. The system first converts the structural information of each individual solution into a calculable numerical vector, and then uses a clustering algorithm, such as K-means clustering, as input to these vectors. At the same time, the performance index of the individual solutions is used as an auxiliary grouping criterion, and finally, individuals with similar performance and structure are grouped into the same group.

[0029] It should also be noted that the process by which the system first transforms the structural information of each individual solution into a computable numerical vector is as follows: extracting the key features of the solution structure, quantizing them into multi-dimensional values, normalizing the multi-dimensional values ​​to the [0,1] interval, and finally arranging them into a vector.

[0030] It should be further explained that when using the K-means clustering method, the system randomly initializes K cluster centers, calculates the Euclidean distance between each numerical vector and the center, assigns it to the cluster corresponding to the nearest center, recalculates the cluster center and iterates until the center is stable or the maximum number of iterations is reached. The Euclidean distance is achieved by taking the square root of the sum of the squared differences of each element.

[0031] The individuals with the highest fitness are selected from each group and combined to form a core solution cluster.

[0032] It should be noted that after grouping, the system will iterate through each group and select the most representative individual from each group. The selection criterion can be the individual with the highest fitness within that group. Finally, these most representative individuals selected from all groups will be grouped together to form a dataset, which is the core solution cluster of the output result.

[0033] It should be further explained that the fitness is obtained by evaluating individual solutions through a predefined fitness function. This fitness value plays a crucial role in the evolution process. The specific calculation method of the fitness function can be summarized as follows: the system quantifies the structural characteristics of individual solutions into numerical indicators based on the performance objectives of the optimization problem, such as flight path length, flight time, and energy consumption in UAV flight path planning, and key constraint parameters, such as maximum flight time, no-fly zone coordinates, and crosswind influence coefficient. These indicators are then mapped to real-valued fitness values ​​through weighted summation, where the weights are determined by the priority of the task performance objectives. Finally, a single fitness value is generated that comprehensively reflects the performance of individual solutions in satisfying constraints and optimization objectives.

[0034] The convergence plateau phase refers to a stagnant state in the evolutionary algorithm where the overall performance of the population or the performance of the best individual no longer shows significant improvement over several generations. It is defined based on statistical analysis of hundreds of independent evolutionary experiments. When the improvement rate of the best fitness value is less than one ten-thousandth over 20 consecutive generations, it is considered to have entered the convergence plateau phase. Therefore, the preset threshold value is one ten-thousandth. The external task phase switching signal is a Boolean or enumerated data signal actively issued by the external scheduling logic according to changes in task requirements, used to force the algorithm to enter the knowledge extraction phase. The elite group is the set of the top 10% of individuals in the current generation population, sorted by fitness from high to low. This 10% percentage is derived from parameter tuning experiments on multiple standard test problems, aiming to balance the quality of the solution with the computational overhead of subsequent calculations. An individual solution is a specific instance of the evolutionary algorithm in the solution space. Its data structure varies depending on the problem being solved; for example, for a path planning problem, it is a sequence of cities; for a parameter optimization problem, it is a vector of real numbers. Solution structure features refer to quantitative indicators that can describe the intrinsic structural characteristics of an individual solution. For example, in network topology optimization problems, solution structure features can be numerical representations of information such as the degree distribution and connection patterns of nodes.

[0035] For example, in an evolutionary task of UAV flight path planning, an individual solution is a sequence of 100 waypoint coordinates. When the evolutionary algorithm reaches the 300th generation, the system detects that the shortening rate of the optimal flight path length in the last 20 generations is less than one ten-thousandth, triggering a convergence plateau. The system then identifies the top 10 individuals with the shortest flight paths from the current population of 100 as the elite group. Next, the individual solutions in these 10 elite groups are grouped and calculated. Based on their solution structure characteristics, i.e., the spatial distribution pattern of the waypoint sequence, they are divided into 3 groups using the K-means clustering algorithm. The first group contains 4 individual solutions, all of which have the flight path characteristic of "first flying over the western region and then over the eastern region"; the second group contains 3 individual solutions, characterized by "crossing the entire region diagonally"; and the third group contains another 3 individual solutions, characterized by "adopting a spiral-like contraction towards the center" flight pattern. Finally, the individual solution with the shortest flight path length is selected from each group as the representative. These three selected individual solutions were eventually combined to form a core solution cluster containing three high-quality and diverse solutions.

[0036] S2. Extract the key constraint parameters and performance target weights corresponding to the task stage when the core decluster is generated, and serialize and perform identifier operations on them to encode the task constraint fingerprint.

[0037] In a specific embodiment of the present invention, the specific steps of the encoding task constraint fingerprint include: arranging and normalizing the extracted key constraint parameters and performance target weights according to a predetermined rule to form a standardized parameter vector.

[0038] Perform a hash function operation on the standardized parameter vector to generate a unique string of fixed length, and use the unique string as a task constraint fingerprint.

[0039] It's important to note that the process of encoding the task constraint fingerprint begins immediately after the generation of the core solution cluster. The input to this process is the task stage at which the core solution cluster is generated. First, the system extracts a set of key constraint parameters and performance target weights from the definition of the current task stage. These data form the core rules defining the current optimization problem. Then, the system serializes the extracted key constraint parameters and performance target weights according to a predetermined rule. This rule specifies the order, data format, and normalization method of all parameters and weights, ensuring that any identical task definition always generates a completely consistent sequence, thus forming a standardized parameter vector. Finally, the system performs an operation on this standardized parameter vector to generate a unique identifier. This operation is a deterministic hash function that maps an input parameter vector of arbitrary length to a fixed-length, unique string—this string is the task constraint fingerprint.

[0040] Key constraint parameters are a series of values ​​defining the boundary of the solution space. Their setting is based on the physical or logical limitations of the specific task, such as the maximum payload, battery range limit, or no-fly zone coordinates in a drone mission. Performance target weights are a set of values ​​used in multi-objective optimization problems to represent the relative importance of different optimization objectives. Their setting is determined by the task decision-maker based on strategic priorities, such as whether flight time or flight distance is more important. Predefined rules are a data processing flow fixed in the system, ensuring that the transformation from parameters to vectors is stable and repeatable. The standardized parameter vector is a one-dimensional array containing all key constraint parameters and performance target weights normalized to a uniform range, such as 0 to 1, facilitating subsequent unified computation. The operation used to generate a unique identifier specifically refers to a cryptographic hash function, such as the SHA-256 algorithm, which generates a unique "digital fingerprint" for a piece of data. The task constraint fingerprint is the final output of this step, a fixed-length hexadecimal string that serves as a unique identifier for the task stage; its data structure is a text string.

[0041] For example, following the previous example of UAV flight path planning, after generating the core decluster in the 300th generation, the system begins to encode task constraint fingerprints for this task phase. First, the system extracts the key constraint parameters for this phase, including a maximum flight time of 60 minutes, the requirement to avoid no-fly zone A with a radius of 2 kilometers centered at coordinates (10,20), the requirement to avoid no-fly zone B with a radius of 3 kilometers centered at coordinates (50,60), and a minimum flight altitude of 50 meters. Simultaneously, performance target weights are extracted: the importance of flight path length is 0.7, and the importance of flight time is 0.3 for this phase. Next, the system serializes and normalizes these data according to predetermined rules, forming a standardized parameter vector, the content of which may be [0.6,0.1,0.2,0.02,0.5,0.6,0.03,0.05,0.7,0.3]. Finally, the system performs a SHA-256 hash operation on this standardized parameter vector to generate a task constraint fingerprint consisting of 64 characters that can accurately summarize the characteristics of the current task, such as "A591A6D40BF420404A011733CFB7B190D62C65BF0BCDA32B57B277D9AD9F146E".

[0042] S3. Backtrack the evolutionary history of the task stages associated with the task constraint fingerprint, count the number of times the evolutionary operator generates offspring that are superior to the parent individual to calculate the performance contribution, and screen out the high-efficiency operators and their contribution values ​​to draw a performance operator profile.

[0043] In a specific embodiment of the present invention, the specific steps of calculating the efficiency contribution by statistically analyzing the number of times the evolutionary operator generates offspring with fitness values ​​superior to its parent individual, and selecting high-efficiency operators and their contribution values ​​to draw a profile of the efficiency operators include: traversing the usage records of each evolutionary operator within a specified task stage, and counting the number of times it successfully generates offspring with fitness values ​​superior to its parent individual.

[0044] The efficiency contribution is obtained by dividing the number of times a descendant with a fitness value better than its parent by the total number of times the evolution operator is used.

[0045] All evolution operators are sorted in descending order of performance contribution. The top-ranked operators and their corresponding performance contribution values ​​are selected to construct a performance operator profile.

[0046] It's important to note that the process of creating performance operator profiles begins with inputting a task stage associated with a task constraint fingerprint. The purpose of this process is to extract the most effective evolutionary strategies for that specific task environment from historical data. First, based on the input task stage identifier, the system traces back and retrieves the usage records of all evolutionary operators in the evolutionary history of that stage. This process relies on a continuously running log system that meticulously records how each new individual was generated from the first generation to the current generation, including information about its parent individuals, the specific evolutionary operator names used, and the fitness values ​​of both the old and new individuals. Next, the system performs statistical evaluation on the log data to calculate the performance contribution of each evolutionary operator. For each evolutionary operator recorded in the log, such as a specific mutation or crossover operation, the system iterates through all its usage records, counting the total number of times it was used, and simultaneously counting the number of times offspring superior to their parents were successfully generated during these uses. An offspring is considered superior to its parent if its fitness value is strictly superior to that of a single parent or the better of two parents. Then, the efficiency contribution of the operator is obtained by dividing the number of successful uses by the total number of uses. Finally, the system sorts all evolutionary operators from high to low according to their calculated efficiency contributions and selects the top-ranked operators, such as the top three. The names of these efficient operators and their corresponding efficiency contribution values ​​are combined into a key-value pair set, which constitutes the efficiency operator profile that records the combination of efficient operators.

[0047] Evolutionary operators are the core operations driving evolutionary algorithms in their search. Their function is to generate new candidate solutions by transforming existing solutions. Common evolutionary operators include different types of mutation or crossover operations, such as single-point crossover, uniform crossover, Gaussian mutation, or polynomial mutation. Efficiency contribution is a quantitative metric, with a floating-point data type. It precisely measures the efficiency or probability of a particular evolutionary operator having a positive impact on the population during a specific task phase in the past.

[0048] For example, continuing with the drone flight path planning example, the system inputs the task stage associated with the task constraint fingerprint "A591A6D40BF420404A011733CFB7B190D62C65BF0BCDA32B57B277D9AD9F146E" generated in the previous stage, which corresponds to generations 1 to 300 of the algorithm. The system retrospectively examines the usage records of all evolutionary operators throughout these 300 generations of evolution and finds that three evolutionary operators were used: "ordered crossover," "exchange mutation," and "flip mutation." Statistical evaluation yields the following data: the ordered crossover operation was used a total of 8000 times, with 2400 successfully generating offspring routes shorter than the parent's route; the exchange mutation operation was used 15000 times, with 1500 successful operations; and the flip mutation operation was used 15000 times, with 3000 successful operations. Based on this, the performance contribution of each operator is calculated: the performance contribution of ordered crossover is 2400 divided by 8000, which equals 0.3; the performance contribution of exchange mutation is 1500 divided by 15000, which equals 0.1; and the performance contribution of flip mutation is 3000 divided by 15000, which equals 0.2. Finally, the system selects the two operators with the highest performance contributions, namely ordered crossover and flip mutation, and constructs a performance operator profile with them and their corresponding contribution values. The data content is {"ordered crossover": 0.3, "flip mutation": 0.2}.

[0049] S4. Forcefully bind and encapsulate the core declustering, task constraint fingerprint, and performance operator profile to construct memory ring units, and store the memory ring units in the memory ring library.

[0050] It's important to note that the process of constructing and storing memory tree units is a data integration and persistence process, receiving all the core data generated in the first three steps as input. The system first receives three independent data objects: core declustering, task constraint fingerprints, and performance operator profiles. Then, the system performs an encapsulation operation, forcibly binding these three as an inseparable data whole. This binding process is achieved by creating a completely new, structured data container, defined as a memory tree unit. When creating a memory tree unit, the system assigns it three dedicated data fields and stores the input core declustering, task constraint fingerprints, and performance operator profiles completely into their respective fields. Simultaneously, to ensure subsequent chronological arrangement, the system also obtains the current system time, generates a precise timestamp, and stores it as the fourth field of the memory tree unit. Once this memory tree unit containing four parts of information is created, it is considered an atomic knowledge record. Finally, the system stores this newly generated memory tree unit in a pre-established memory tree repository. The design of the memory tree library ensures the temporal order of the data. The simplest implementation is a first-in-first-out queue or list structure. Each newly generated memory tree unit is appended to the end of the library, thus naturally forming a storage structure arranged in chronological order for subsequent queries and retrieval.

[0051] For example, following the previous step, the system now receives three data inputs. The first input is the core solution cluster, which contains three UAV flight path schemes representing different flight strategies. The second input is the task constraint fingerprint, i.e., the string "A591A6D40BF420404A011733CFB7B190D62C65BF0BCDA32B57B277D9AD9F146E". The third input is the performance operator profile, i.e., the data {"ordered crossover": 0.3, "flip mutation": 0.2}. The system then creates a new memory tree unit. The internal structure of this unit contains four parts: the first part stores the core solution cluster composed of the three UAV flight path schemes; the second part stores the task constraint fingerprint in string form; the third part stores the performance operator profile recording the combination of efficient operators; and the fourth part records the current creation timestamp, such as "2023-10-27 10:30:00". Finally, this complete memory ring unit is stored as a whole in the memory ring library. If this is the first time this operation has been performed, then the memory ring library now contains its first record.

[0052] S5. Obtain the new key constraint parameters and performance target weights for the new task stage, and use the same operation rules as the encoding task constraint fingerprint to parse and generate the task fingerprint to be matched.

[0053] It should be noted that the process of parsing new tasks to generate matching task fingerprints is activated when the evolutionary process enters a completely new task stage. This activation signal can be the arrival of a preset iteration number or an instruction issued by the external environment monitoring system after detecting a change in the problem. Once a new stage is entered, the system first obtains the task definition for this new stage, which is a complete description containing all the rules and objectives of the new problem. The system precisely extracts the corresponding new key constraint parameters and performance target weights from the task definition. Next, the system strictly follows the exact same serialization and identifier generation rules used in step S2 to encode the task constraint fingerprint to process this batch of newly extracted data. This means that the system uses the same parameter arrangement order, data format, and normalization method to transform this newly extracted set of new key constraint parameters and performance target weights into a standardized parameter vector with a consistent structure. Subsequently, the system performs the same deterministic hash function operation, such as the SHA-256 algorithm, on this newly generated standardized parameter vector. The result of the operation is the generation of a fixed-length string that can uniquely represent the characteristics of this new task. This string is the fingerprint of the task to be matched, which will serve as the "key" for retrieving similar historical experiences in the memory bank later.

[0054] In this context, a new task phase refers to a substantial change in the optimization problem faced by the evolutionary algorithm, such as the addition or removal of constraints, a change in the optimization objective, or an adjustment of weights. The task definition is a structured data file or configuration information that fully describes the problem to be solved in the new phase, including all relevant constraints and objective information. New key constraint parameters are a series of updated values ​​specified in the new task definition to limit the range of feasible solutions. The task fingerprint to be matched is the final output of this step; its data structure and generation method are exactly the same as the task constraint fingerprint, both being a fixed-length text string.

[0055] For example, when the UAV flight path planning task has been running for 500 generations, the external system issues a command, and the task enters a new phase. The new task requires the UAV to fly in crosswind conditions and also needs to consider a newly established temporary no-fly zone. The system obtains the task definition for the new phase and extracts its corresponding new key constraint parameters. These parameters include the original maximum flight time of 60 minutes, no-fly zones A and B, and a newly added temporary no-fly zone C with a radius of 1 kilometer centered at coordinates (30,40), as well as a parameter value of 0.2 representing the impact of crosswinds. At the same time, because crosswinds increase energy consumption, the performance target weights have also changed, with the emphasis on flight path length adjusted to 0.6 and the emphasis on flight time adjusted to 0.4. The system uses the exact same serialization and identifier generation rules as in step S2 to process these data into a new, standardized parameter vector, such as [0.6,0.1,0.2,0.02,0.5,0.6,0.03,0.3,0.4,0.01,0.2,0.6,0.4]. Finally, the system performs a SHA-256 hash operation on this vector to generate a fingerprint representing the characteristics of the new task, such as “E3B0C44298FC1C149AFBF4C8996FB92427AE41E4649B934CA495991B7852B855”.

[0056] S6. Compare the fingerprint of the task to be matched with the task constraint fingerprint of each memory ring unit in the memory ring library, identify and locate the historical memory ring unit with the highest resonance degree, so as to match and extract the resonance ring data package.

[0057] In a specific embodiment of the present invention, the specific steps of identifying and locating the historical memory ring unit with the highest resonance degree in order to match and extract the resonance ring data package include: extracting a new standardized parameter vector to generate the fingerprint of the task to be matched.

[0058] Traverse the memory tree database and extract the historical standardized parameter vector corresponding to each historical memory tree unit.

[0059] Calculate the vector similarity between the new standardized parameter vector and each historical standardized parameter vector to quantify the degree of resonance.

[0060] The historical memory ring unit with the highest resonance value was selected.

[0061] All information contained in the historical memory ring unit with the highest resonance value is extracted to form a resonance ring data package.

[0062] It's important to note that the system first uses the fingerprint of the task to be matched as a query index to initiate a comparison operation on each memory ring cell stored in the memory ring database. This comparison process is not simply comparing whether the fingerprint strings are equal, but rather assessing the similarity between the new and old tasks. For each memory ring cell in the memory ring database, the system extracts the task constraint fingerprint stored within it, and further, extracts the standardized parameter vector of the original key constraint parameters and performance target weights that generated the historical fingerprint. Simultaneously, the system also holds a completely new standardized parameter vector for generating the fingerprint of the task to be matched. Then, the system calculates the resonance degree between the new task and each historical task based on the overlap of the key constraint parameters. This calculation is performed using a vector similarity algorithm, which takes the standardized parameter vectors of the two tasks as input and outputs a numerical value representing their similarity. The system iterates through all cells in the memory ring database, performing this similarity calculation one by one, and recording the score of each historical cell. After the traversal is complete, the system will find the highest score among all scores. The historical memory ring unit corresponding to this highest score is the historical memory ring unit that is identified and located with the highest degree of "resonance" with the characteristics of the current new task. Finally, the system will perform a complete deep copy of all the information contained in this selected memory ring unit with the highest degree of resonance, including its internally encapsulated core de-clustering, task constraint fingerprint, and performance operator profile. This data will be encapsulated into a new temporary data container to form a resonance ring data package as the final output of this step.

[0063] It should also be noted that the formula for calculating the vector similarity between the new standardized parameter vector and each historical standardized parameter vector is: ,in, Represents the degree of resonance between the new task and the historical task. It is a dimensionless floating-point number between -1 and 1. The closer the value is to 1, the higher the similarity. The standardized parameter vector representing the new task is a numerical vector formed by serializing new key constraint parameters and performance target weights. Represents the standardized parameter vector corresponding to a specific historical memory ring unit in the memory ring library; the symbol " " " represents the dot product operation of vectors, that is, the sum of corresponding elements multiplied together; the symbol ‖V‖ represents the Euclidean norm of a vector, that is, the square root of the sum of the squares of all elements in the vector.

[0064] For example, the system inputs the fingerprint of the task to be matched generated in the previous stage, "E3B0C44298FC1C149AFBF4C8996FB92427AE41E4649B934CA495991B7852B855", and the corresponding normalized parameter vector. At this moment, the memory ring library stores only one memory ring unit generated in the previous step, with a task constraint fingerprint of "A591A6D40BF420404A011733CFB7B190D62C65BF0BCDA32B57B277D9AD9F146E", and a corresponding historical normalized parameter vector of... The system will and As input, based on the overlap of key constraint parameters, the cosine similarity formula is used to calculate the resonance degree between them, which is 0.92. Since there is only one unit in the library, this unit is naturally identified and located as the historical memory ring unit with the highest "resonance" degree with the current new task features. Subsequently, the system extracts all the information contained in this memory ring unit, namely the core solution cluster composed of 3 route schemes, the task constraint fingerprint of the string "A591A6D4...AD9F146E", and the performance operator profile with the content {"ordered crossover": 0.3, "flip mutation": 0.2}, to form a resonance ring data package.

[0065] S7. Extract the core solution cluster encapsulated within the resonant tree ring data packet and use it to replace the individual with the lowest fitness ranking in the current population, so as to apply the core solution cluster to reshape and form the foundation of the reshaped population.

[0066] In a specific embodiment of the present invention, the specific steps of applying the core de-clustering reshaping to form the reshaped population basis include: sorting all individuals in the population of the current new task stage in descending order of fitness value.

[0067] Determine the number of individuals that need to be replaced. The number of individuals that need to be replaced is equal to the number of individual solutions contained in the core solution cluster.

[0068] Using individual solutions from the core solution cluster, the corresponding number of individuals with the lowest ranking in the current population are replaced one by one, thereby reshaping the population foundation.

[0069] It's important to note that the process of reshaping the population foundation using the core solution cluster is executed immediately upon receiving the resonant tree-ring data packet. First, the system inputs the resonant tree-ring data packet and extracts its encapsulated core solution cluster. This core solution cluster contains multiple high-quality and diverse individual solutions originating from similar historical tasks. Simultaneously, the system evaluates the population in the current new task phase, sorting each individual from highest to lowest fitness value. This sorting accurately identifies the lowest-ranked individuals in the current population. The number of individuals to be replaced is equal to the number of individual solutions contained in the core solution cluster. Next, the system performs the replacement operation, using the diverse high-quality solutions from the core solution cluster to replace the lowest-ranked individuals in the current population one-to-one. For example, the first solution in the core solution cluster replaces the lowest-fitting individual, the second replaces the second-to-last, and so on, until all solutions in the core solution cluster are injected into the population. After this replacement process, the overall quality lower bound of the population is instantly improved because the previously worst-performing parts are replaced by historically validated excellent solutions. The new population resulting from this operation constitutes an optimized and reshaped population foundation, which will serve as the starting point for the next round of evolutionary iterations.

[0070] For example, following the previous step, the system inputs the resonant tree-ring data package and successfully parses its encapsulated core solution cluster, which contains three high-quality UAV flight path planning schemes. Simultaneously, the system analyzes the population of the current new task phase, which contains 100 individuals. The system calculates the fitness of each individual in the new task environment and sorts these 100 individuals from highest to lowest fitness. Through sorting, the system identifies the group of individuals with the lowest fitness ranking in the current population, namely, the three individuals ranked 100th, 99th, and 98th. Subsequently, the system performs a replacement operation: replacing the 100th ranked individual with the first flight path scheme from the core solution cluster, replacing the 99th ranked individual with the second flight path scheme, and replacing the 98th ranked individual with the third flight path scheme. After the replacement, the original three worst solutions in the population disappear, replaced by three solutions that have proven to be excellent in similar historical tasks. This optimized new population forms a reshaped population foundation, providing higher-quality initial conditions for subsequent evolution.

[0071] S8. Synchronously parse the performance operator profile encapsulated within the resonant tree ring data package, and weight the selection probability of the high-efficiency operators recorded in the performance operator profile to generate a dynamic operator weight table using the performance operator profile.

[0072] In a specific embodiment of the present invention, the specific steps of weighting and increasing the selection probabilities of high-efficiency operators recorded in the performance operator profile to generate a dynamic operator weight table using the performance operator profile include: creating a temporary operator selection probability table, wherein each operator in the temporary operator selection probability table has an initial selection probability.

[0073] For each efficient operator recorded in the efficiency operator profile, its probability value in the temporary operator selection probability table is multiplied by a set weighting factor.

[0074] The entire temporary operator selection probability table is normalized to form a dynamic operator weight table.

[0075] It's important to note that the process of generating a dynamic operator weight table using the applied performance operator profile is performed concurrently with the previous step of applying the core de-clustering operation. First, the system also inputs the resonant tree-ring data package and parses out its internally encapsulated performance operator profile. This performance operator profile records which evolutionary operators have shown higher efficiency in solving similar historical tasks. Next, the system creates a temporary operator selection probability table. In the initial settings of the evolutionary algorithm, all available evolutionary operators are typically assigned equal selection probabilities. However, in this step, the system modifies this probability table. Specifically, the system iterates through all the efficient operators recorded in the performance operator profile and significantly increases their selection probabilities in the probability table. This increase can be achieved by multiplying their base probability by a weight factor greater than 1, or by directly adding a fixed probability increment, and then normalizing the entire probability table to ensure that the sum of the selection probabilities of all operators remains 1. The direct result of this operation is that operators that have historically proven effective in solving similar problems will have a significantly increased chance of being selected in subsequent evolutionary processes. Finally, the system solidifies this modified probability table, generating a dynamic operator weight table that will be effective for a specified number of future generations and has a selection bias towards historically efficient operators. This dynamic operator weight table will serve as a temporary evolutionary strategy configuration, guiding the selection process of evolutionary operators within the specified number of generations.

[0076] It should also be noted that the core of this step is to perform a weighted adjustment of the operator selection probability, and its update process can be described by a specific mathematical formula. The formula is as follows: , ,in, Represents the adjusted evolution operator The new choice probability is a dimensionless floating-point number; It is an operator The initial selection probabilities are usually equal in the initial state. For example, if there are N operators, then... =1 / N; It is obtained from the performance operator profile and is related to the operator. The performance contribution of the association, if a certain operator is not in the portrait, then its =0; It is a normalization factor used to ensure that the sum of the probabilities of all new choices is 1; Representation and operator The associated performance contribution, Operator The original choice probability, Indicates the operator number. .

[0077] For example, continuing with the drone flight path planning example, the system, while inputting the resonant tree-ring data package, also parses its internally encapsulated performance operator profile, which contains {"Ordered Crossover": 0.3, "Flip Mutation": 0.2}. Assume there are three available evolutionary operators in the system: "Ordered Crossover", "Exchange Mutation", and "Flip Mutation", with their initial operator selection probability table being {"Ordered Crossover": 0.333, "Exchange Mutation": 0.333, "Flip Mutation": 0.333}. The system now creates a temporary operator selection probability table and weights it according to the performance operator profile. For "Ordered Crossover", its weight W is 0.3, and the new unnormalized probability is 0.333 × (1 + 0.3) = 0.4329. For "Flip Mutation", its weight W is 0.2, and the new unnormalized probability is 0.333 × (1 + 0.2) = 0.3996. "Exchange Mutation" is not in the profile, its weight W is 0, and its probability remains 0.333. The normalization factor Z = 0.4329 + 0.3996 + 0.333 = 1.1655. After normalization, the new selection probabilities are: "Ordered Crossover" ≈ 0.4329 / 1.1655 ≈ 0.371, "Flip Mutation" ≈ 0.3996 / 1.1655 ≈ 0.343, and "Exchange Mutation" ≈ 0.333 / 1.1655 ≈ 0.286. The system ultimately generates a dynamic operator weight table that will be effective for the next 50 generations, with the following content: {"Ordered Crossover": 0.371, "Exchange Mutation": 0.286, "Flip Mutation": 0.343}.

[0078] S9. Using the reshaped population base as the starting population, and selecting evolutionary operators for subsequent iterations based on the weighted probabilities defined in the dynamic operator weight table, to execute dual-channel collaborative guided evolution.

[0079] In a specific embodiment of the present invention, the specific steps of performing dual-channel collaborative guided evolution include: by reshaping the population base as the starting population, injecting high-quality prior knowledge directly into the solution space level to provide a higher starting point for evolution.

[0080] By selecting evolutionary operators based on a dynamic operator weight table, the search direction is guided at the policy space level to favor historically proven efficient operations.

[0081] The guidance at the solution space level and the guidance at the policy space level work together to accelerate the convergence of the evolutionary process in the new task phase.

[0082] It should be noted that high-quality prior knowledge refers to: by identifying the top one-third of the current generation's population based on fitness, and selecting representative individuals based on solution structure characteristics to generate core solution clusters, these core solution clusters contain high-quality solutions that have been validated in similar historical tasks, exhibiting excellent performance and diverse structures; these core solution clusters are directly injected into the starting population of the new task stage, forming a reshaped population foundation, thereby providing a starting point with excellent genes and problem adaptability for the evolutionary process at the solution space level, significantly improving the convergence speed and solution performance of the evolutionary algorithm in the new environment.

[0083] It should also be noted that the process of executing the dual-channel collaborative guided evolution is the application and harvest phase of the entire methodology. This process is initiated by receiving the reshaped population base generated in step S7 and the dynamic operator weight table generated in step S8. First, the system uses the reshaped population base as the starting population for the new round of evolution. This means that at the beginning of the first iteration of the new task phase, the algorithm no longer operates on a randomly generated population or a population naturally continuing from the previous phase, but rather on this population that has been optimized by high-quality historical solutions and has a higher starting point. Then, the iterative cycle of the evolutionary algorithm officially begins. In subsequent iterations, whenever an evolutionary operator needs to be selected from the operator library to perform crossover or mutation operations on individuals, the system makes the selection based on the weighted probabilities defined in the dynamic operator weight table. This selection process is usually implemented through a roulette wheel selection mechanism, where the sector area occupied by each operator on the virtual roulette wheel is proportional to its probability in the weight table; operators with higher probabilities are more likely to be selected. This biased selection mechanism remains in effect for the specified number of generations. In this way, the search behavior of the evolutionary algorithm is guided by two channels: one channel directly improves the population's gene pool through the injection of high-quality solutions, and the other channel optimizes the algorithm's search strategy through operator preference selection. These two channels work together to guide the evolutionary direction with historically proven efficient strategies, thereby enabling the algorithm to find high-quality solutions more quickly in new task phases and achieving accelerated convergence.

[0084] The core idea of ​​this method is dual-channel collaborative guided evolution, a composite evolutionary guidance strategy. Dual-channel refers to simultaneously influencing the population at two levels: the first channel is the solution space channel, which guides the evolution at the solution level by reshaping the population foundation through input; the second channel is the policy space channel, which guides the evolutionary operator's policy selection through inputting a dynamic operator weight table. Collaborative guidance means that these two channels work simultaneously, cooperating to guide the evolutionary process in a more promising direction. Accelerated convergence is the ultimate effect of this guidance strategy, meaning that compared to standard evolutionary algorithms without historical guidance, this method can achieve the desired average fitness or optimal fitness of the population in fewer iterations.

[0085] For example, following the previous steps, the evolutionary algorithm officially begins dual-channel cooperative guided evolution upon entering generation 501. First, the system inputs a reshaped population base, which is a new population formed by replacing the three worst-performing schemes from the original 100 UAV flight path schemes with historically high-quality schemes. Simultaneously, the system also inputs a dynamic operator weight table, containing {"Ordered Crossover": 0.371, "Exchange Mutation": 0.286, "Flip Mutation": 0.343}, with a validity period of 50 generations. In each iteration from generation 501 to 550, when crossover is required, the algorithm chooses "Ordered Crossover" with a probability of 37.1%, significantly higher than other operators. When mutation is required, the probability of choosing "Flip Mutation" is also relatively higher. Due to the higher quality of the initial population and the algorithm's preference for using "Ordered Crossover" and "Flip Mutation" operators proven effective in similar tasks, the overall performance of the population improves very rapidly. Finally, in the 535th generation, the algorithm found a good flight path that met the requirements of the new task (with crosswinds and a new no-fly zone). In the control experiment without this mechanism, it might have taken up to the 620th generation to achieve a similar performance level. This clearly demonstrates the effect of accelerating convergence in the new task stage.

[0086] The above content is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the concept of the invention or exceed the scope defined by the present invention, and all such modifications and additions should fall within the protection scope of the present invention.

Claims

1. An evolutionary method based on memory ring learning, characterized in that, Includes the following steps: S1. Locate the elite group in the current generation population and perform group calculations based on the solution structure characteristics. Select representative individuals from each group to generate the core solution cluster. S2. Extract the key constraint parameters and performance target weights of the task stage corresponding to the generation of the core uncluster cluster, and serialize and perform identifier operations on them to encode the task constraint fingerprint. S3. Backtrack the evolutionary history of the task stages associated with the task constraint fingerprint, count the number of times the evolution operator generates offspring that are superior to the parent individual to calculate the performance contribution, and screen out the high-efficiency operators and their contribution values ​​to draw a performance operator profile. S4. Forcefully bind and encapsulate the core declustering, task constraint fingerprint, and performance operator profile to construct memory ring units and store the memory ring units in the memory ring library. S5. Obtain the new key constraint parameters and performance target weights for the new task stage, and use the same operation rules as the encoding task constraint fingerprint to parse and generate the task fingerprint to be matched. S6. Compare the fingerprint of the task to be matched with the task constraint fingerprint of each memory ring unit in the memory ring library, identify and locate the historical memory ring unit with the highest resonance degree, so as to match and extract the resonance ring data package. S7. Extract the core solution cluster encapsulated within the resonant tree ring data packet and use it to replace the individual with the lowest fitness ranking in the current population, so as to apply the core solution cluster to reshape and form the basis of the reshaped population. S8. Synchronously parse the performance operator profile encapsulated within the resonant tree ring data packet, and weight the selection probability of the high-efficiency operators recorded in the performance operator profile to generate a dynamic operator weight table using the performance operator profile. S9. Using the reshaped population base as the starting population, and selecting evolutionary operators for subsequent iterations based on the weighted probabilities defined in the dynamic operator weight table, to execute dual-channel collaborative guided evolution.

2. The evolutionary method based on memory ring learning according to claim 1, characterized in that: The specific steps of locking the elite group in the current generation population and grouping them according to the solution structure characteristics, and selecting representative individuals from each group to generate the core solution cluster include: When the evolutionary process reaches a predetermined convergence plateau or receives an external task phase switching signal, the top one-third of the individuals in the current generation population with the highest fitness are locked to form an elite group. The structural information of each individual solution in the elite group is transformed into a numerical vector, and a clustering algorithm is used to group individuals with similar performance and structure into the same group. The individuals with the highest fitness are selected from each group and combined to form a core solution cluster.

3. The evolutionary method based on memory ring learning according to claim 1, characterized in that: The specific steps for constraining fingerprints in the encoding task include: The extracted key constraint parameters and performance target weights are arranged and normalized according to predetermined rules to form a standardized parameter vector. Perform a hash function operation on the standardized parameter vector to generate a unique string of fixed length, and use the unique string as a task constraint fingerprint.

4. The evolutionary method based on memory ring learning according to claim 1, characterized in that: The specific steps for calculating the efficiency contribution by calculating the number of times the statistical evolutionary operator generates offspring superior to the parent individual, and for screening out high-efficiency operators and their contribution values ​​to create efficiency operator profiles, include: Iterate through the usage records of each evolution operator in the specified task phase and count the number of times it successfully generates offspring with fitness values ​​better than its parent individual. The efficiency contribution is obtained by dividing the number of times a descendant with a fitness value better than its parent by the total number of times the evolutionary operator is used. All evolution operators are sorted in descending order of performance contribution. The top-ranked operators and their corresponding performance contribution values ​​are selected to construct a performance operator profile.

5. The evolutionary method based on memory ring learning according to claim 1, characterized in that: The specific steps for identifying and locating the historical memory ring units with the highest resonance level, in order to match and extract the resonance ring data package, include: Extract a novel, standardized parameter vector to generate the fingerprint of the task to be matched; Traverse the memory tree database and extract the historical standardized parameter vector corresponding to each historical memory tree unit; Calculate the vector similarity between the new standardized parameter vector and each historical standardized parameter vector to quantify the degree of resonance; Select the historical memory ring unit with the highest resonance value; All information contained in the historical memory ring unit with the highest resonance value is extracted to form a resonance ring data package.

6. The evolutionary method based on memory ring learning according to claim 1, characterized in that: The specific steps for forming the reshaped population basis through the application of core cluster de-clustering and reshaping include: Sort all individuals in the population in the current new task phase in descending order of fitness value; Determine the number of individuals that need to be replaced. The number of individuals that need to be replaced is equal to the number of individual solutions contained in the core solution cluster. Using individual solutions from the core solution cluster, the corresponding number of individuals with the lowest ranking in the current population are replaced one by one, thereby reshaping the population foundation.

7. The evolutionary method based on memory ring learning according to claim 1, characterized in that: The specific steps for weighting and boosting the selection probabilities of high-efficiency operators recorded in the performance operator profile to generate a dynamic operator weight table using the performance operator profile include: Create a temporary operator selection probability table, in which each operator has an initial selection probability; For each high-efficiency operator recorded in the efficiency operator profile, its probability value in the temporary operator selection probability table is multiplied by a set weighting factor. The entire temporary operator selection probability table is normalized to form a dynamic operator weight table.

8. The evolutionary method based on memory ring learning according to claim 1, characterized in that: The specific steps for performing dual-channel collaborative guided evolution include: By reshaping the population base as the starting population, high-quality prior knowledge is directly injected into the solution space level, providing a higher starting point for evolution. By selecting evolutionary operators based on a dynamic operator weight table, the search direction is guided at the policy space level to favor historically proven efficient operations. The guidance at the solution space level and the guidance at the policy space level work together to accelerate the convergence of the evolutionary process in the new task phase.

Images