A hierarchical adaptive hyperheuristic robot path planning method

By employing a hierarchical adaptive hyperheuristic algorithm and a near-end policy optimization network, and dynamically scheduling a low-level heuristic operator pool, the problems of local optima and poor cross-scene adaptability in mobile robot path planning are solved, achieving efficient path generation in complex environments.

CN123113883APending Publication Date: 2026-11-17GUIZHOU UNIV +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611452380.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-18
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing mobile robot path planning methods are prone to getting stuck in local optima, having rigid search strategies, and exhibiting poor cross-scenario adaptability in complex, nonlinear, and high-dimensional search spaces. They are also difficult to generate high-quality paths in environments with dense obstacles and discontinuous feasible domains.

Method used

A hierarchical adaptive hyperheuristic algorithm is adopted to generate an initial candidate path population by acquiring a two-dimensional occupancy grid map. The hierarchical adaptive hyperheuristic algorithm is then used for iterative optimization. Combined with a near-end policy optimization network and an individual-level policy preference mechanism, a low-level heuristic operator pool is dynamically scheduled to generate collision-free paths.

Benefits of technology

It improves environmental adaptability and comprehensive solution performance in complex scenarios, enhances the stability and efficiency of path planning, and generates high-quality collision-free paths.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123113883A_ABST
    Figure CN123113883A_ABST
Patent Text Reader

Abstract

This invention discloses a hierarchical adaptive hyperheuristic robot path planning method, comprising: acquiring a two-dimensional occupancy grid map of the mobile robot's workspace, a starting point, and a target point; encoding the continuous row coordinates corresponding to each intermediate column between the starting point and the target point as candidate path individuals; constructing discrete grid paths by rounding the continuous row coordinates to generate an initial candidate path population; determining path feasibility based on boundary crossings and obstacle collisions of the initial candidate paths; performing local line-of-sight pruning on feasible paths; evaluating path quality based on the pruned path length and effective turning number; setting hierarchical penalties for infeasible paths; evaluating the fitness of each initial candidate path individual; and iteratively optimizing the candidate path individuals using a hierarchical adaptive hyperheuristic algorithm. This invention exhibits superior environmental adaptability and comprehensive solution performance in complex scenarios.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of machine technology, and more specifically to a hierarchical adaptive hyperheuristic robot path planning method. Background Technology

[0002] Mobile robots are gradually moving from structured environments to more complex and dynamic real-world application scenarios. As a core component of autonomous navigation, path planning aims to generate a feasible path with good overall performance in terms of path length, smoothness, number of turns, and computational cost, while meeting the requirements of obstacle avoidance, boundary constraints, and motion executability. However, in environments with dense obstacles and fragmented feasible regions, path planning problems often exhibit characteristics such as non-convexity, multi-peakedness, and strong constraints, making the optimization process more prone to getting trapped in local optima and exhibiting significant convergence fluctuations.

[0003] In existing technologies, mobile robot path planning methods can be broadly categorized into two types: traditional planning methods and heuristic optimization methods. Traditional methods, such as A*, Dijkstra's algorithm, and artificial potential field methods, typically rely on explicit graph structures and heuristic evaluation functions, exhibiting strong interpretability and widespread application in environments with relatively clear structures. However, as the environment scales up, obstacle distribution becomes more complex, or constraints increase, the search cost of these methods often rises significantly. In contrast, heuristic optimization methods typically model path planning as a constrained optimization problem, utilizing stochastic search or swarm evolution mechanisms to find approximate optimal solutions, thus demonstrating greater flexibility in complex, nonlinear, and high-dimensional search spaces. Metaheuristic algorithms such as Differential Evolution (DE) and Particle Swarm Optimization (PSO) have been widely used in path planning tasks, but these methods typically rely on fixed search operators or preset parameter rules, making it difficult to dynamically coordinate the relationship between exploration and development based on the search stage and scene difficulty. When exploration is insufficient, the algorithm is prone to getting trapped in local optima; while when exploration is excessive, it may lead to a decrease in convergence speed and unstable solution quality. This contradiction is particularly prominent for path planning problems with strong constraints and discontinuous feasible regions.

[0004] In recent years, reinforcement learning has provided a new approach for data-driven adaptive operator selection. By modeling the search process as a sequential decision problem, reinforcement learning can learn dynamic scheduling strategies based on environmental feedback, thus supporting search control in complex optimization problems. Among these, proximal policy optimization (PPO) is suitable for learning high-level scheduling strategies in evolutionary search due to its advantages such as stable training, controllable update magnitude, and suitability for continuous action modeling. However, performing discrete decisions directly on a large number of low-level operators still faces problems such as high action space dimensionality, low sampling efficiency, and unstable training; if only a population-level unified strategy is adopted, it is difficult to take into account the local search needs of different individuals. In summary, mobile robot path planning is prone to problems such as getting trapped in local optima, rigid search strategies, and poor cross-scene adaptability. Summary of the Invention

[0005] The purpose of this invention is to overcome the above-mentioned shortcomings and propose a hierarchical adaptive hyperheuristic robot path planning method that can exhibit better environmental adaptability and comprehensive solution performance in complex scenarios.

[0006] The present invention provides a hierarchical adaptive hyperheuristic robot path planning method, comprising the following steps:

[0007] A two-dimensional grid map of the mobile robot's workspace, the starting point, and the target point are obtained. The continuous row coordinates corresponding to each intermediate column between the starting point and the target point are encoded as candidate path individuals. Discrete grid paths are constructed by rounding the continuous row coordinates to generate an initial candidate path population. The feasibility of the path is determined based on the boundary crossing and obstacle collision of the initial candidate path. Local line-of-sight pruning is performed on feasible paths. The path quality is evaluated by the length of the pruned path and the number of effective turns. Layered penalties are set for infeasible paths. The fitness of each initial candidate path individual is evaluated.

[0008] A hierarchical adaptive hyperheuristic algorithm is used to iteratively optimize individual candidate paths. The solution process specifically includes:

[0009] Based on the initial candidate path population, the set of globally optimal individuals and elite individuals is determined according to the fitness. Multiple low-level heuristic operators are divided into three functional operator pools: development type, balanced type, and exploratory type. Historical preference probability vectors for each low-level heuristic operator are initialized for each candidate path individual.

[0010] Based on the current iteration progress, the relative improvement rate of the recent global best fitness, the rate of change of population diversity, the current fitness histogram, and the historical average fitness histogram, construct a search state whose dimensions do not change with map size or path decision dimensions.

[0011] The search state is input into an offline training network with fixed parameters for a near-end policy optimization strategy network to obtain a one-dimensional continuous scheduling signal representing local exploitation, balanced search, and global exploration tendencies. Based on the distance between the one-dimensional continuous scheduling signal and the ordered anchor points corresponding to the three functional operator pools, the macroscopic weights of the three functional operator pools are obtained through distance response and temperature normalization.

[0012] For each candidate path, the historical preference probability of each low-level heuristic operator is normalized within its respective functional operator pool, and then multiplied by the macro weight of the functional operator pool to obtain the joint selection probability of each low-level heuristic operator. The sum of the joint selection probabilities of each low-level heuristic operator in any functional operator pool is equal to the macro weight of the functional operator pool. Low-level heuristic operators are sampled according to the joint selection probability to generate offspring candidate paths.

[0013] The offspring candidate paths are subjected to boundary detection and statistics, discretization and fitness evaluation. When the offspring fitness is less than the parent fitness, the offspring replaces the parent, increasing the historical preference probability of the corresponding individual for the selected low-level heuristic operator and proportionally decreasing the historical preference probability of the other low-level heuristic operators. Otherwise, the parent and the corresponding historical preference probability vector are retained.

[0014] After all candidate path individuals have been updated, the elite individual set, the globally optimal individual, and the current and historical data required to construct the search state are updated. Population evolution is repeated until the maximum number of iterations is reached or the preset convergence condition is met. The path corresponding to the globally optimal individual is checked for boundary violations, connectivity, and collisions, and two local line-of-sight prunings are performed. The collision-free path node sequence from the starting point to the target point is output, and the path node sequence is sent to the navigation control module of the mobile robot to complete the robot path planning.

[0015] The above-mentioned hierarchical adaptive hyperheuristic robot path planning method, wherein the two-dimensional occupancy grid map is a square grid map composed of obstacle grids and passable grids, and the starting point and the target point are both located in the passable grids, and the starting point column and the target point column are different;

[0016] The continuous row coordinates corresponding to each intermediate column between the starting point and the target point are encoded as candidate path individuals. The encoding of the candidate path individuals is applicable to grid path planning scenarios that monotonically advance along the column direction. The intermediate columns between the starting point column and the target point column are fixed in the order from the starting point to the target point. Each candidate path individual uses the continuous row coordinates corresponding to each intermediate column as decision variables.

[0017] The coordinates of each consecutive row are rounded to the nearest integer to obtain the discrete raster row coordinates of the corresponding intermediate column. Following the order of the intermediate columns, the starting point, the discrete raster nodes in each intermediate column, and the target point are connected sequentially to form an initial path. This allows the initial candidate path individuals to be updated within the continuous space, and boundary crossing and collision detection are performed on the discrete raster. The candidate path individuals and the initial path are represented as follows:

[0018]

[0019]

[0020] in, For the first Individual candidate paths For this individual in the first The coordinates of consecutive rows in the middle column; This represents the number of intermediate columns between the starting column and the target column. For the first The fixed column coordinates of the middle column, for Discrete row coordinates obtained by rounding; and These are the starting point and the target point, respectively. For the first The initial path corresponding to each candidate path individual, with the superscript 0 indicating the path before pruning.

[0021] The aforementioned hierarchical adaptive hyperheuristic robot path planning method, wherein the feasibility of a path is determined based on the boundary crossing and obstacle collision situations of the initial candidate paths, local line-of-sight pruning is performed on feasible paths, the path quality is evaluated by the length of the pruned path and the number of effective turns, and hierarchical penalties are set for infeasible paths, and the fitness evaluation of each initial candidate path is performed, including:

[0022] The number of out-of-bounds decision variables among the initial candidate path individuals is counted, and the out-of-bounds decision variables are not truncated before fitness evaluation; when there is no out-of-bounds, the number of obstacle grids traversed by the initial path is counted; when there is no out-of-bounds and the initial path does not intersect with obstacle grids, the candidate path is determined to be a feasible path.

[0023] Fitness is determined using a hierarchical penalty method. When there are out-of-bounds decision variables, the fitness is the product of the square of the number of grid cells on each side of the map and the number of out-of-bounds decision variables. When there are no out-of-bounds decision variables but the initial path intersects with obstacle grid cells, the fitness is the product of the square of the number of grid cells on each side of the map and the number of intersections.

[0024] For a feasible path, two local line-of-sight pruning operations are performed. The sum of the Euclidean distances between adjacent feasible path nodes after pruning is taken as the path length, and the number of intermediate nodes retained after pruning is taken as the effective turning count. The sum of the path length and the effective turning count multiplied by a preset effective turning weight is taken as the fitness. The path length and fitness function are as follows:

[0025]

[0026]

[0027] in, For the first The length of the path after pruning. The total number of nodes reserved for this path; , The first The row and column coordinates of each node. , These are the row and column coordinates of the next node, respectively. For the first The number of out-of-bounds decision variables in each candidate path individual The number of obstacle grids traversed by its initial path; The equivalent path length weight for effective turning; For fitness evaluation function, Indicates the first The fitness of each candidate path individual; the lower the fitness, the better the candidate path. This is the base value for boundary crossing and collision penalties that varies with map scale.

[0028] The aforementioned hierarchical adaptive hyperheuristic robot path planning method, wherein performing two local line-of-sight pruning operations includes:

[0029] For three consecutive path nodes in the feasible path, detect the grids traversed by the alternative line segment that directly connects the previous node and the next node. When all the grids traversed are passable grids, delete the middle node among the three path nodes.

[0030] After the first scan is completed, the feasible paths are scanned again to delete redundant nodes that were newly exposed during the first scan.

[0031] The starting and target points are preserved in both scans, and collision detection is performed again for each alternative line segment.

[0032] The aforementioned hierarchical adaptive hyperheuristic robot path planning method, wherein the search state is constructed in the following manner:

[0033] The ratio of the current iteration count to the maximum iteration count is used as the current iteration progress;

[0034]

[0035] in, This represents the current iteration progress. Let the current iteration algebra be... This represents the maximum number of iterations.

[0036] The difference between the global optimal fitness before a preset number of generations and the current global optimal fitness is divided by the sum of the absolute value of the global optimal fitness before the preset number of generations and the numerical stability term to obtain the relative improvement rate of the recent global optimal fitness.

[0037]

[0038] in, The relative improvement rate of the global optimal fitness within the most recent preset algebra; and The first generation and first The global optimal fitness of the generation; The preset algebra used to calculate the recent improvement rate, Indicates looking back Generation; A positive numerically stable term. Indicates taking the absolute value;

[0039] The current population diversity is obtained by dividing the standard deviation of the current population fitness by the sum of the absolute value of its mean and the numerical stability term. The population diversity rate is obtained by dividing the difference between the population diversity before the preset generation and the current population diversity by the sum of the population diversity before the preset generation and the numerical stability term.

[0040]

[0041]

[0042] in, For the first Let std be the set of fitness values ​​of all candidate paths, and mean be the standard deviation function and mean function, respectively. and The first generation and first Population diversity of generations For the most recent The rate of change in population diversity over generations;

[0043] The fitness of each candidate path individual is normalized to the interval between zero and one. This interval is divided into a preset number of intervals. The proportion of individuals in each interval is counted to obtain the current fitness histogram. The historical average fitness histogram is obtained by maintaining a historical histogram queue of a preset length.

[0044] The fitness normalization and historical average fitness histogram are respectively represented as follows:

[0045]

[0046]

[0047] in, For the first The generation The fitness of each candidate path individual Its normalized fitness is indicated by the horizontal line above; For the first The current fitness histogram vector of the generation. The vector of the historical average fitness histogram; The length of the historical histogram queue; For the first The fitness histogram vector of the generation;

[0048] The search state is formed by concatenating the current iteration progress, relative improvement rate, rate of change of population diversity, current fitness histogram, and historical average fitness histogram, and is represented as follows:

[0049]

[0050] in, For the first The search state vector of the generation; , , , and These are the iteration progress, relative improvement rate, population diversity change rate, current fitness histogram vector, and historical average fitness histogram vector, respectively; each of the two histogram vectors contains the proportion of individuals in five intervals.

[0051] The above-mentioned hierarchical adaptive hyperheuristic robot path planning method includes a proximal policy optimization network parameterized by a one-dimensional Gaussian policy. The original action is sampled from the one-dimensional Gaussian policy according to the search state. The original action is incremented by one and divided by two, and then clipped to the interval between 0 and 1 to obtain the one-dimensional continuous scheduling signal.

[0052] The one-dimensional continuous scheduling signals, from smallest to largest, correspond to local development, balanced search, and global exploration tendencies, respectively.

[0053] The one-dimensional continuous scheduling signal is represented as follows:

[0054]

[0055] in, The original actions sampled from a one-dimensional Gaussian policy, is the corresponding one-dimensional continuous scheduling signal; Clip is the interval clipping function, whose first parameter is the value to be clipped, and the second and third parameters are the lower limit 0 and the upper limit 1, respectively. When it is less than the lower limit, the lower limit is taken, and when it is greater than the upper limit, the upper limit is taken. In other cases, the value is kept.

[0056] The proximal policy optimization network is trained offline in the following way: the policy network interacts with the population evolution process, the difference between the current generation's global best fitness and the next generation's global best fitness is divided by the sum of the absolute value of the current generation's global best fitness and the numerical stability term, and used as an immediate reward. The advantage estimates within each sampled trajectory are standardized, and the proximal policy optimization objective with a pruning probability ratio is used to update the policy network parameters.

[0057] The instant reward is represented as follows:

[0058]

[0059] in, For the first The instant reward corresponding to the action. and The first generation and first The global optimal fitness of the generation; in the near-end policy optimization update, the probability ratio of the old and new policies is based on the original action. calculate;

[0060] After completing offline training, the network parameters of the policy are fixed, and forward inference is performed only once per generation on the map to be planned.

[0061] In the above-mentioned hierarchical adaptive hyperheuristic robot path planning method, the ordered anchor points corresponding to the developmental, balanced, and exploratory functional operator pools are 1 / 6, 3 / 6, and 5 / 6, respectively.

[0062]

[0063] in, , and These are the ordered anchor points for the development, balanced, and exploratory functional operator pools, respectively;

[0064] Calculate the absolute distance between the one-dimensional continuous scheduling signal and each ordered anchor point, normalize the absolute distance using the corresponding normalized distance factor, and multiply the value obtained by subtracting the normalized distance from the signal by the response intensity coefficient to obtain the distance response of the corresponding functional operator pool.

[0065]

[0066] in, For a one-dimensional continuous scheduling signal of generation t; For the first Ordered anchor points of a pool of functional operators; For the first A positive normalized distance factor for each functional operator pool; λ is a positive response intensity coefficient; λ is the distance response function. For the first Each functional operator pool in scheduling signals Distance response;

[0067] Divide each distance response by the temperature coefficient, take the exponent, and then divide by the sum of the exponent values ​​corresponding to the three functional operator pools to obtain the macroscopic weight of each functional operator pool. This ensures that the closer the one-dimensional continuous scheduling signal is to a certain ordered anchor point, the larger the distance response of the functional operator pool corresponding to that ordered anchor point, as expressed as:

[0068]

[0069] in, For the first The macroscopic weight function of the functional operator pool; The temperature coefficient is positive, and exp is the natural exponential function. For the first The distance response of the three functional operator pools under the same scheduling signal is equal to the sum of their macroscopic weights, which is 1.

[0070] The aforementioned hierarchical adaptive hyperheuristic robot path planning method, wherein the division of multiple low-level heuristic operators into three functional operator pools—exploratory, balancing, and exploratory—specifically involves:

[0071] The exploratory functional operator pool includes an elite search operator based on Lévy perturbation, a random difference mutation operator DE / rand / 1, and a random individual binomial crossover operator;

[0072] The balanced functional operator pool includes an elite search operator based on Gaussian perturbation, a current individual differential evolution operator DE / current / 1, an optimal individual differential evolution operator DE / best / 1, and an elite-guided differential evolution operator DE / current-to-pbest / 1.

[0073] The development-oriented functional operator pool includes an elite neighborhood search operator based on uniform perturbation, a global optimal guided differential evolution operator DE / current-to-best / 1 operator, and an elite individual binomial crossover operator.

[0074] The elite individual set consists of the top 10% of individuals in the population whose fitness is sorted from smallest to largest. The elite-guided differential evolution operator DE / current-to-pbest / 1 provides elite guidance through the centroids of the top 10% of individuals whose fitness is sorted from smallest to largest.

[0075] The above-mentioned hierarchical adaptive hyperheuristic robot path planning method, wherein the step of initializing the historical preference probability vector for each low-level heuristic operator for each candidate path individual includes: assigning the same probability to all low-level heuristic operators during the initialization of the historical preference probability vector of each candidate path individual;

[0076] When the The low-level heuristic operator belongs to the . When the function operator pool is in its first phase, the first... The candidate path individual selects the first... The joint selection probability of the low-level heuristic operators is:

[0077]

[0078] in, For the first Individual selection of candidate paths The joint selection probability of a low-level heuristic operator; For the first The macroscopic weight function of the functional operator pool; For the first The set of indexes of low-level heuristic operators contained in a functional operator pool; For the first Individual with candidate path for the first Historical preference probabilities of low-level heuristic operators;

[0079] Each functional operator pool satisfies the following probability conservation condition:

[0080]

[0081] Among them, the probability conservation formula For traversal The summation index of all operators within the pool; the probability conservation ensures that individual historical preferences only change the selection share of each low-level heuristic operator within the same functional operator pool, while keeping the macro weight of the functional operator pool unchanged.

[0082] The aforementioned hierarchical adaptive hyperheuristic robot path planning method, wherein the step of sampling low-level heuristic operators according to the joint selection probability to generate offspring candidate paths, wherein when the selected first... The low-level heuristic operator is the first... When the fitness of the offspring generated by an individual along a candidate path is less than that of its parent, the historical preference probability of that individual is updated as follows:

[0083]

[0084] And the historical preference probabilities of the remaining low-level heuristic operators are proportionally decayed:

[0085]

[0086] in, The previous iteration generation. This refers to the generation after this update; and The first The candidate path individual selects the first... Historical preference probabilities before and after operator update and These are the individual's views on the first... Historical preference probabilities before and after updating the remaining operators; The learning rate is greater than 0 and less than 1; the sum of all probabilities in the updated historical preference probability vector is 1.

[0087] When the fitness of the offspring is not less than that of the parent, the historical preference probability vector of the individual remains unchanged, and the historical preference probability is updated only based on whether the offspring is better than the parent.

[0088] Compared with existing technologies, this invention has significant advantages: under the constraints of complex grid maps, this invention adaptively adjusts operator usage preferences based on changes in the search state, aiming to improve optimization performance, search stability, and environmental adaptability. The main advantages are as follows:

[0089] (1) An adaptive operator pool scheduling strategy based on PPO is proposed. This strategy adaptively allocates operator pool weights according to population statistical characteristics, thereby balancing exploration and exploitation. Unlike existing reinforcement learning hyperheuristic (RL-HH) methods that directly select low-level operators as discrete actions, this paper divides ten low-level operators into exploratory, balancing, and exploitative operator pools; the high-level PPO outputs one-dimensional continuous actions, which are converted into dynamic operator pool weights through soft mapping. Thus, the direct selection of the ever-expanding discrete operator set is reconstructed into one-dimensional continuous policy control, reducing the complexity of high-level scheduling.

[0090] (2) Design an individual-level strategy preference (ISP) mechanism to coordinate with the high-level operator pool scheduling strategy. To address the limitation that a unified population-level scheduling cannot characterize individual search differences, each individual maintains an independent low-level operator preference distribution and updates it online based on historical improvement feedback. Then, the individual preferences are fused with the weights of the high-level operator pool to generate the final operator selection probability, thereby achieving differentiated search behavior under a unified macro search rhythm.

[0091] The following specific embodiments further illustrate the beneficial effects of the present invention. Attached Figure Description

[0092] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0093] The following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, features, and effects of a hierarchical adaptive hyperheuristic robot path planning method proposed according to the present invention.

[0094] See Figure 1 The present invention discloses a hierarchical adaptive hyperheuristic robot path planning method, which includes the following steps:

[0095] Step 1: Obtain a two-dimensional grid map of the mobile robot's workspace, the starting point and the target point, establish a continuous-discrete coding, feasibility determination and path quality evaluation model for candidate paths, and generate an initial path population;

[0096] Step 1.1: Represent the robot's workspace as Occupied grid map ,in, Indicates the first line, number Listed as a barrier grid, This indicates that the corresponding grid cell is passable; the starting point is denoted as... The target point is denoted as Both the starting point and the target point are located within a passable grid.

[0097] Step 1.2: Fix the intermediate columns between the starting point column and the target point column in the order from the starting point to the target point. ,in, ;No. Each candidate path individual encodes only the coordinates of consecutive rows corresponding to each intermediate column; its model is as follows:

[0098]

[0099] in, Indicates the first The candidate paths are in the middle column. Continuous row coordinates on; perform on them Obtain the corresponding discrete raster row coordinates, and then construct the original path:

[0100]

[0101] The encoding enables evolutionary search to update path individuals in a continuous space, while also allowing for direct boundary crossing and collision detection on discrete grids; the encoding is suitable for grid path planning scenarios that can monotonically advance along column directions.

[0102] Step 1.3: Count the number of out-of-bounds decision variables for each candidate path individual. The number of intersections between the original path and the obstacle grid If and only if When this happens, the candidate path is determined to be a feasible path;

[0103] Step 1.4: Perform two local line-of-sight pruning operations on the feasible original path; for three consecutive path nodes... , and If by Directly connected If all the grid cells traversed by the line segment are passable grid cells, then delete the intermediate node. After the first scan is completed, a second scan is performed to delete redundant nodes newly exposed due to the first scan; the start point and target point are retained in both scans, and collision detection is performed again for each alternative line segment;

[0104] Step 1.5: Let the path after pruning be... ,in, , The path length is:

[0105]

[0106] Retained after pruning The number of intermediate nodes is used as the effective number of turns, and a hierarchical penalty method is used to construct the fitness function:

[0107]

[0108] in, The base for boundary crossing and collision penalties varies with map scale. The equivalent path length weight for effective turning; in one specific implementation, This evaluation method prioritizes eliminating infeasible solutions and simultaneously shortens the path and reduces turning within the feasible region.

[0109] Step 2: Initialize the path population, elite individual set, and individual-level strategy preferences, and extract the search state with fixed dimensions from the path population;

[0110] Step 2.1: Randomly generate a list containing... The initial population of candidate path individuals is determined by calculating the fitness of each individual according to step 1, thus identifying the current globally optimal individual and the set of elite individuals with the top 10% fitness ranking; for each path individual... Establish separately by A vector consisting of the preference probabilities of several low-level heuristic operators:

[0111]

[0112] in, Indicates the first Individuals on the path to the first Historical preferences of low-level heuristic operators; initialization timing This ensures that each operator has the same initial selection opportunity;

[0113] Step 2.2: In the first... In the next iteration, the state vector of the high-level policy is constructed:

[0114]

[0115] in, Indicates the current iteration progress. This represents the maximum number of iterations. Indicates recent The relative improvement rate of the global optimal fitness is modeled as follows:

[0116]

[0117] in, and They represent the first generation and first The global optimal fitness of the generation, To prevent the denominator from approaching zero, a stable term is needed; when A smaller value indicates limited recent improvement and can be used to identify search stagnation;

[0118] Step 2.3: Characterize population diversity using the coefficient of variation of population fitness, whereby the diversity and its nearest neighbors... The rates of change for each generation are as follows:

[0119]

[0120] in, For the first The fitness set of all individuals; This indicates a recent decline in population diversity. This indicates that diversity has not significantly declined or has recovered;

[0121] Step 2.4: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.] Normalize the fitness of each individual to The interval is divided into equal parts. The fitness histogram is generated by dividing the data into intervals and calculating the proportion of individuals within each interval. Maintenance length is Calculate the historical average histogram from the historical histogram queue:

[0122]

[0123] State vector The dimension is In one specific implementation, Therefore, the state dimension is fixed at 13. The state is not directly concatenated with map pixels or path coordinates. Map size, obstacle topology, collision constraints, and path quality are incorporated into the state through fitness distribution and its changes. Therefore, the policy input dimension does not change with map size or path decision dimension.

[0124] Step 3: Input the fixed-dimensional state into the offline-trained near-end policy optimization policy network, and the policy network outputs a one-dimensional continuous scheduling signal;

[0125] Step 3.1: The Actor network parameterizes the one-dimensional Gaussian policy of the proximal policy optimization policy network. According to the status Sampling original action And convert the original action into a range One-dimensional continuous scheduling signal within:

[0126]

[0127] in, It only indicates the macroscopic tendency of the current population on the ordered search axis of development-equilibrium-exploration, and does not directly indicate specific low-level heuristic operators; smaller This corresponds to a strong tendency towards localized development, while intermediate values ​​correspond to a tendency towards balanced search, with larger values... It corresponds to a strong tendency towards global exploration;

[0128] Step 3.2: During the offline training phase, the policy network interacts with the population evolution process, using the relative improvement rate of the global optimal fitness between two adjacent generations as the immediate reward:

[0129]

[0130] The reward uses relative improvement rather than absolute improvement to reduce the impact of the target scale of different training functions and the magnitude of improvement in different search stages on policy updates; the advantage estimation within each sampled trajectory is standardized separately, and a proximal policy with a pruning probability ratio is used to optimize the target update policy parameters to avoid excessive shift between the old and new policies in a single update; after offline training is completed, the policy network parameters are fixed, and forward inference is only performed on the map to be planned, without retraining for each map.

[0131] Step 3.3: The one-dimensional continuous scheduling is the high-level decision-making interface of this method. When adding, replacing, or deleting specific low-level operators in the existing functional operator pool, the high-level policy still only outputs a continuous scalar, without needing to expand the discrete action set or policy output layer as the number of specific operators increases. Thus, the high-level policy only learns the macroscopic search rhythm, and the adaptation of specific operators is handled by the individual-level mechanism described later.

[0132] Step 4: Based on the distance between the one-dimensional continuous scheduling signal and the three ordered anchor points, the one-dimensional continuous scheduling signal is soft-mapped into macroscopic weights for the development, balance, and exploration operator pools;

[0133] Step 4.1: Following the functional order of local development, balanced search, and global exploration, within the interval... Three equally spaced anchor points are set on the top, namely:

[0134]

[0135] in, , and These represent the anchor points for the developmental, balanced, and exploratory operator pools, respectively; as... As the macroscopic search tendency increases, it changes smoothly in the order of development, balancing, and exploration;

[0136] Step 4.2: Calculate the one-dimensional continuous scheduling signal pair. Distance response of an operator pool:

[0137]

[0138] in, For normalized distance factor, The response intensity coefficient is used; the closer the operator pool is to the corresponding anchor point, the greater the original response it receives.

[0139] Step 4.3: Convert the three original responses into macroscopic weights using temperature normalization:

[0140]

[0141] in, Indicates the first The proportion of search resources obtained by each functional operator pool in the current iteration. Temperature coefficient; relatively small This makes the weights more concentrated, larger To make the mixing of the three operator pools smoother, in one specific implementation, we take... The distance response and temperature normalization ensure that similar consecutive actions generate similar macroscopic weights, avoiding behavioral abrupt changes caused by switching search modes according to hard thresholds; the three ordered anchors also restrict the policy output to a smooth weight family with clear development-exploration semantics, avoiding disordered combinations when the policy directly outputs three independent weights.

[0142] Step 5: Establish a set of low-level heuristic operators with complementary search functions, and divide them into exploratory, balanced and developmental operator pools based on the search step size, direction guidance and randomness of each operator;

[0143] Step 5.1: The exploratory operator pool is used to expand the spatial coverage in the early stage or stagnant phase of the search, including the elite search operator based on Lévy perturbation, the DE / rand / 1 random difference mutation operator, and the random individual binomial crossover operator; among them, Lévy perturbation is used to generate a small number of large step jumps, DE / rand / 1 provides unbiased search without relying on the current optimal solution, and random individual binomial crossover introduces external individual genes to slow down population homogenization;

[0144] Step 5.2: The balanced operator pool is used to gradually enhance convergence while maintaining a certain search breadth. It includes an elite search operator based on Gaussian perturbation, the DE / current / 1 operator, the DE / best / 1 operator, and the DE / current-to-pbest / 1 operator. Among them, Gaussian perturbation forms an isotropic search in the elite neighborhood, DE / current / 1 retains the memory of the current individual position, DE / best / 1 introduces the globally optimal direction, and DE / current-to-pbest / 1 provides a more robust elite guidance through the statistical information of the top 10% of individuals in terms of fitness.

[0145] Step 5.3: The development operator pool is used to implement fine-grained search in the neighborhood of high-quality solutions, including an elite neighborhood search operator based on uniform perturbation, a DE / current-to-best / 1 operator, and an elite individual binomial crossover operator; wherein, uniform perturbation narrows the search range within the elite neighborhood, DE / current-to-best / 1 enhances the driving force toward the current global optimal solution, and elite individual binomial crossover directly inherits the favorable components in the elite individuals;

[0146] Step 5.4: The number and specific form of the ten low-level heuristic operators are used to illustrate a specific implementation method and do not constitute a limitation on the number of operators; as long as the newly added operators can be included in the existing development, balanced or exploratory function pools, the one-dimensional output structure of the high-level strategy and the macro-scheduling structure of the three operator pools do not need to be changed.

[0147] Step 6: For each path individual, the macro weights of the operator pool are fused with the individual's historical preferences through pool normalization to obtain the joint selection probability of specific low-level heuristic operators, and the preferences of the corresponding individual are updated online based on the improvement results of the offspring.

[0148] Step 6.1: When the first The low-level heuristic operator belongs to the . When the function operator pool is in its first phase, the first... The candidate path individual selects the first... The joint selection probability of the low-level heuristic operators is:

[0149]

[0150] in, Used to allocate macroscopic search resources among the three types of functional operator pools. For use in accordance with the The historical returns of individuals along each path are in the... Within each operator pool, the selection share of specific operators is reallocated; for each entity, according to all... Perform random sampling and call the selected operator to generate candidate child paths;

[0151] Step 6.2: From Step 6.1, we can obtain:

[0152]

[0153] The equation constitutes the hierarchical probability conservation condition of this method, namely, the sum of the joint selection probabilities of all specific operators within any operator pool is always strictly equal to the macroscopic weight assigned to that operator pool by the higher-level strategy. Therefore, individual-level preferences can only change which operator uses resources within the same operator pool, and cannot increase or decrease the total resources of that operator pool; thus, the population can maintain a unified macroscopic development-exploration rhythm, while allowing different individuals to adopt different low-level operators based on their fitness level, local search position, and stagnation history.

[0154] Step 6.3: If the first The low-level heuristic operator is the first... The offspring generated by each path individual satisfy Then the child generation replaces the parent generation, and the individual's preference for the operator is strengthened:

[0155]

[0156] in, To determine the preferred learning rate, in one specific implementation, we take... Simultaneously, the preferences of the other operators are reduced proportionally:

[0157]

[0158] The proportional decay ensures that the updated individual preference vector still satisfies a probability sum of 1; if the offspring is not better than the parent, the individual's preference vector remains unchanged. The preference is updated only based on whether it has improved, without directly using the absolute improvement amount, thereby reducing the impact of different map target scales on individual-level learning.

[0159] Step 6.4: The aforementioned one-dimensional ordered macro-scheduling and individual preferences within the pool are not two independent additional strategies, but rather form a two-level decision chain constrained by probability conservation through steps 6.1 and 6.2: the higher level answers "how much search resource should be used for development, balancing, and exploration in this generation," and the lower level answers "which specific operator within the pool should use this resource for each individual." This structure decouples the macro-search rhythm from the micro-operator adaptation, which is the key difference between this method and directly selecting specific operators, directly outputting multiple unordered weights, or sharing a single operator probability across the entire population.

[0160] Step 7: Perform path population evolution, state history update, and path output;

[0161] Step 7.1: In each iteration, construct the state according to Step 2, obtain the one-dimensional continuous scheduling signal according to Step 3, obtain the macroscopic weights of the three types of operator pools according to Step 4, and perform the joint probability sampling, operator update, out-of-bounds detection and statistics, discretization, fitness evaluation, parent-child selection and individual preference update of Step 6 for each individual in the population.

[0162] Step 7.2: After all individuals have been updated, the elite individual set and the globally optimal individual are redefined, and the globally optimal fitness, population diversity, and current and historical fitness histograms are updated simultaneously to construct the state for the next iteration;

[0163] Step 7.3: Repeat steps 7.1 to 7.2 until the maximum number of iterations is reached or the preset convergence condition is met; perform final boundary check, connectivity check, collision check and two local line-of-sight pruning on the global optimal individual, output the collision-free path node sequence from the starting point to the target point, and send the path node sequence to the navigation control module of the mobile robot;

[0164] Step 7.4: During the deployment phase, the high-level strategy performs forward inference only once per generation on a fixed-size network, and individual-level preference updates only process networks of length [missing information]. The probability vector; when When the value is constant, the hierarchical scheduling mechanism does not change the asymptotic complexity order determined by population evolution and fitness evaluation, thus maintaining the same main computational level as the basic population evolution process while increasing adaptive scheduling capabilities.

[0165] Based on the above steps, the greatest advantage of this invention is not simply employing a near-end strategy optimization algorithm or simply stacking ten low-level operators, but rather establishing a complete technical chain of "fixed-dimensional state—one-dimensional ordered macro-scheduling—inter-pool soft allocation—individual adaptation within the pool." The high-level output dimension is fixed at 1, similar actions result in similar search resource allocations, and low-level adaptation is constrained by the probability conservation condition within the pool. Therefore, the high-level decision space will not expand due to an increase in the number of specific operators, nor will the macro-search rhythm of the entire population be disrupted due to individual differentiated learning.

[0166] Effect Analysis

[0167] To verify the correctness and effectiveness of this invention, the path planning experiment used a population size of 300 and a maximum number of iterations of 100. Each algorithm was run independently 30 times on a fixed map and 50 times on a random map. All algorithms used the same initial population, path encoding, feasibility handling, fitness function, and search budget. For the fixed map, average path length, best path length, worst path length, and average number of effective turns were used as evaluation metrics. For the random map, the success rate of finding a connected and collision-free path was used.

[0168] First, in a 30×30 grid environment with fragmented obstacle distribution, this method is combined with random operator scheduling. Using only individual-level preferences Only using high-level near-end strategies to optimize scheduling Historical adaptive selection and the use of discrete high-level decision-making The ablation comparison was performed, and the results are shown in Table 1.

[0169] Table 1 Comparison of ablation results in fixed grid environments

[0170]

[0171] As shown in Table 1, this method achieves the lowest average path length and worst-case path length. Compared to the other five scheduling variants, the average path length is reduced by 1.65% to 14.53%, and the worst-case path length is reduced by 2.69% to 25.85%. Specifically, compared to... The average and worst values ​​decreased by 1.65% and 2.69% respectively, indicating that individual-level preferences can further improve path quality and repeated run stability without changing the macro-pool weights; compared to The average value decreased by 9.79%, indicating that one-dimensional continuous scheduling is superior to discrete search mode selection. and The average number of turns is lower, but its path length and worst-case performance are significantly worse than this method.

[0172] Secondly, in In irregular obstacle environments, our method was compared with representative swarm intelligence algorithms such as DE, GWO, PSO, and CL-PSO. The results are shown in Table 2.

[0173] Table 2 Comparison Results of Representative Swarm Intelligence Algorithms

[0174]

[0175] As shown in Table 2, this method achieves optimal results across all four metrics. Compared to the best comparison values ​​for each metric, the average path length, optimal path length, worst path length, and average number of turns of this method are reduced by 8.12%, 1.24%, 12.95%, and 14.89%, respectively. The difference between the worst and best values ​​is 9.25, which is also smaller than that of the comparison algorithms, indicating that this method combines path quality, smoothness, and operational stability in environments with significant irregularities.

[0176] Again, in In a grid environment, this method is compared with reinforcement learning-assisted optimization algorithms such as DRLDE, RLDMDE, TSRL-CMM, and RLDE-PFR. The results are shown in Table 3.

[0177] Table 3 Comparison Results of Reinforcement Learning-Assisted Optimization Algorithms

[0178]

[0179] As shown in Table 3, this method achieves the lowest average path length, optimal path length, and average number of turns, which are reduced by 0.67%, 0.92%, and 2.40% respectively compared to the corresponding best comparison values. The worst path length of RLDMDE is lower than that of this method, indicating that there is still room for improvement in the extreme performance of this method in this environment; however, its average path quality and path simplicity are generally the best.

[0180] Finally, in four independently generated In a fragmented random map, feasible path search stress tests were conducted on all fourteen path planning algorithms. The results are shown in Table 4.

[0181] Table 4 Summary of the success rates of all path planning algorithms on random maps

[0182]

[0183] Table 4 shows that our proposed method achieved the highest success rate across all four random maps, with an average success rate of 77.0%. This represents an improvement of 11.5 percentage points compared to the optimal comparison algorithm DRLDE, and 21.5 percentage points compared to HHA-PPO, which removes individual-level preferences. The results indicate that fusing high-level continuous pool scheduling with individual-level in-pool preferences can improve search reliability in fragmented feasible regions, and this effect is independent of specific obstacle layouts.

[0184] To further validate the scheduling structure itself, comparisons were made on the CEC2017 test function (which was not used in training) between HHA-SR, HHA-ISP, HHA-PPO, HHA-SHA, HHA-DDQN, HHA-PPO-DW (which directly outputs the weights of three pools), HHA-PPO-OD (which directly selects ten low-level operators), and the complete method PPO-HAHH. The average Friedman rankings of the above algorithms in 10-dimensional and 30-dimensional settings were as follows: HHA-SR 5.19 and 4.78, HHA-ISP 4.64 and 4.43, HHA-PPO 3.48 and 3.31, HHA-SHA 3.64 and 3.48, HHA-DDQN 3.97 and 4.17, HHA-PPO-DW 5.34 and 6.07, HHA-PPO-OD 7.93 for all, and our method 1.81 and 1.83. The complete approach ranks first in both dimensions, indicating that its advantage comes from the synergy of one-dimensional ordered continuous scheduling, anchor soft mapping, and individual preferences within the pool, rather than solely from the near-end policy optimization algorithm or the number of low-level operators.

[0185] In summary, this invention maintains superior average path quality across three comparison categories: ablation variants, traditional swarm intelligence algorithms, and reinforcement learning-assisted optimization algorithms. Furthermore, it achieves the highest feasible path search success rate in a random map stress test involving all fourteen algorithms. The results demonstrate that the synergistic effect of one-dimensional continuous pool-level scheduling and individual preferences within the pool can improve path optimization quality and feasible path search reliability in complex grid environments while maintaining a unified macroscopic search rhythm for the population.

[0186] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any simple modifications, equivalent changes, and alterations made to the above embodiments without departing from the technical essence of the present invention shall still fall within the scope of the present invention.

Claims

1. A hierarchical adaptive hyperheuristic robot path planning method, characterized in that, Includes the following steps: A two-dimensional grid map of the mobile robot's workspace, the starting point, and the target point are obtained. The continuous row coordinates corresponding to each intermediate column between the starting point and the target point are encoded as candidate path individuals. Discrete grid paths are constructed by rounding the continuous row coordinates to generate an initial candidate path population. The feasibility of the path is determined based on the boundary crossing and obstacle collision of the initial candidate path. Local line-of-sight pruning is performed on feasible paths. The path quality is evaluated by the length of the pruned path and the number of effective turns. Layered penalties are set for infeasible paths. The fitness of each initial candidate path individual is evaluated. A hierarchical adaptive hyperheuristic algorithm is used to iteratively optimize individual candidate paths. The solution process specifically includes: Based on the initial candidate path population, the set of globally optimal individuals and elite individuals is determined according to the fitness. Multiple low-level heuristic operators are divided into three functional operator pools: development type, balanced type, and exploratory type. Historical preference probability vectors for each low-level heuristic operator are initialized for each candidate path individual. Based on the current iteration progress, the improvement rate of the recent global best fitness, the change rate of population diversity, the current fitness histogram, and the historical average fitness histogram, construct a search state whose dimensions do not change with map size or path decision dimensions. The search state is input into an offline training network with fixed parameters for a near-end policy optimization strategy network to obtain a one-dimensional continuous scheduling signal representing local exploitation, balanced search, and global exploration tendencies. Based on the distance between the one-dimensional continuous scheduling signal and the ordered anchor points corresponding to the three functional operator pools, the macroscopic weights of the three functional operator pools are obtained through distance response and temperature normalization. For each candidate path, the historical preference probability of each low-level heuristic operator is normalized within its respective functional operator pool, and then multiplied by the macro weight of the functional operator pool to obtain the joint selection probability of each low-level heuristic operator. The sum of the joint selection probabilities of each low-level heuristic operator in any functional operator pool is equal to the macro weight of the functional operator pool. Low-level heuristic operators are sampled according to the joint selection probability to generate offspring candidate paths. The offspring candidate paths are subjected to boundary detection and statistics, discretization and fitness evaluation. When the offspring fitness is less than the parent fitness, the offspring replaces the parent, increasing the historical preference probability of the corresponding individual for the selected low-level heuristic operator and proportionally decreasing the historical preference probability of the other low-level heuristic operators. Otherwise, the parent and the corresponding historical preference probability vector are retained. After all candidate path individuals have been updated, the elite individual set, the globally optimal individual, and the current and historical data required to construct the search state are updated. Population evolution is repeated until the maximum number of iterations is reached or the preset convergence condition is met. The path corresponding to the globally optimal individual is checked for boundary violations, connectivity, and collisions, and two local line-of-sight prunings are performed. The collision-free path node sequence from the starting point to the target point is output, and the path node sequence is sent to the navigation control module of the mobile robot to complete the robot path planning.

2. The hierarchical adaptive hyperheuristic robot path planning method according to claim 1, characterized in that, The two-dimensional occupancy grid map is a square grid map composed of obstacle grids and passable grids. The starting point and the target point are both located within the passable grids, and the starting point column and the target point column are different. The continuous row coordinates corresponding to each intermediate column between the starting point and the target point are encoded as candidate path individuals. The encoding of the candidate path individuals is applicable to grid path planning scenarios that monotonically advance along the column direction. The intermediate columns between the starting point column and the target point column are fixed in the order from the starting point to the target point. Each candidate path individual uses the continuous row coordinates corresponding to each intermediate column as decision variables. The coordinates of each consecutive row are rounded to the nearest integer to obtain the discrete raster row coordinates of the corresponding intermediate column. Following the order of the intermediate columns, the starting point, the discrete raster nodes in each intermediate column, and the target point are connected sequentially to form an initial path. This allows the initial candidate path individuals to be updated within the continuous space, and boundary crossing and collision detection are performed on the discrete raster. The candidate path individuals and the initial path are represented as follows: in, For the first Individual candidate paths For this individual in the first The coordinates of consecutive rows in the middle column; This represents the number of intermediate columns between the starting column and the target column. For the first The fixed column coordinates of the middle column, for Discrete row coordinates obtained by rounding; and These are the starting point and the target point, respectively. For the first The initial path corresponding to each candidate path individual, with the superscript 0 indicating the path before pruning.

3. The hierarchical adaptive hyperheuristic robot path planning method according to claim 1, characterized in that, The process involves determining path feasibility based on boundary crossings and obstacle collisions of the initial candidate paths, performing local line-of-sight pruning on feasible paths, evaluating path quality based on the length of the pruned path and the number of effective turns, and imposing tiered penalties on infeasible paths. The fitness evaluation of each initial candidate path includes: The number of out-of-bounds decision variables among the initial candidate path individuals is counted, and the out-of-bounds decision variables are not truncated before fitness evaluation; when there is no out-of-bounds, the number of obstacle grids traversed by the initial candidate path is counted; when there is no out-of-bounds and the initial candidate path does not intersect with obstacle grids, the initial candidate path is determined to be a feasible path. A hierarchical penalty method is used to determine fitness. When there are out-of-bounds decision variables, the product of the square of the number of grid cells on each side of the map and the number of out-of-bounds decision variables is used as fitness. When there are no out-of-bounds decision variables but the initial candidate path intersects with the obstacle grid cells, the product of the square of the number of grid cells on each side of the map and the number of intersections is used as fitness. For a feasible path, two local line-of-sight pruning operations are performed. The sum of the Euclidean distances between adjacent feasible path nodes after pruning is taken as the path length, and the number of intermediate nodes retained after pruning is taken as the effective turning count. The sum of the path length and the effective turning count multiplied by a preset effective turning weight is taken as the fitness. The path length and fitness function are as follows: in, For the first The length of the path after pruning. The total number of nodes reserved for this path; , The first The row and column coordinates of each node. , These are the row and column coordinates of the next node, respectively. For the first The number of out-of-bounds decision variables in each candidate path individual For the first The number of obstacle grids traversed by each candidate path; The equivalent path length weight for effective turning; For fitness evaluation function, Indicates the first The fitness of each candidate path individual; the lower the fitness, the better the candidate path. This is the base value for boundary crossing and collision penalties that varies with map scale.

4. The hierarchical adaptive hyperheuristic robot path planning method according to claim 3, characterized in that, The two local line-of-sight pruning operations include: For three consecutive path nodes in the feasible path, detect the grids traversed by the alternative line segment that directly connects the previous node and the next node. When all the grids traversed are passable grids, delete the middle node among the three path nodes. After the first scan is completed, the feasible paths are scanned again to delete redundant nodes that were newly exposed during the first scan. The starting and target points are preserved in both scans, and collision detection is performed again for each alternative line segment.

5. The hierarchical adaptive hyperheuristic robot path planning method according to claim 1, characterized in that, The search state is constructed in the following manner: The ratio of the current iteration count to the maximum iteration count is used as the current iteration progress; in, This represents the current iteration progress. Let the current iteration algebra be... This represents the maximum number of iterations. The difference between the global optimal fitness before a preset number of generations and the current global optimal fitness is divided by the sum of the absolute value of the global optimal fitness before the preset number of generations and the numerical stability term to obtain the improvement rate of the recent global optimal fitness. in, The improvement rate of the globally optimal fitness within the most recent preset algebra; and The first generation and first The global optimal fitness of the generation; The preset algebra used to calculate the recent improvement rate, Indicates looking back Generation; A positive numerically stable term. Indicates taking the absolute value; The current population diversity is obtained by dividing the standard deviation of the current population fitness by the sum of the absolute value of its mean and the numerical stability term. The population diversity rate is obtained by dividing the difference between the population diversity before the preset generation and the current population diversity by the sum of the population diversity before the preset generation and the numerical stability term. in, For the first Let std be the set of fitness values ​​of all candidate paths, and mean be the standard deviation function and mean function, respectively. and The first generation and first Population diversity of generations For the most recent The rate of change in population diversity over generations; Normalize the fitness of each candidate path individual to the interval between 0 and 1, divide the interval into a preset number of intervals, count the proportion of individuals in each interval to obtain the current fitness histogram, and obtain the historical average fitness histogram by maintaining a historical histogram queue of preset length. The fitness normalization and historical average fitness histogram are respectively represented as follows: in, For the first The generation The fitness of each candidate path individual Its normalized fitness is indicated by the horizontal line above; For the first The current fitness histogram vector of the generation. The vector of the historical average fitness histogram; The length of the historical histogram queue; For the first The fitness histogram vector of the generation; The search state is formed by concatenating the current iteration progress, improvement rate, rate of change of population diversity, current fitness histogram, and historical average fitness histogram, and is represented as follows: in, For the first The search state vector of the generation; , , , and The first The iteration progress, improvement rate, population diversity change rate, current fitness histogram vector, and historical average fitness histogram vector; each of the two histogram vectors contains the proportion of individuals in five intervals.

6. The hierarchical adaptive hyperheuristic robot path planning method according to claim 1, characterized in that, The action network parameterized one-dimensional Gaussian policy of the near-end policy optimization policy network is used to sample the original action from the one-dimensional Gaussian policy according to the search state, add one to the original action, divide by two, and prune to the interval of 0 to 1 to obtain the one-dimensional continuous scheduling signal. The one-dimensional continuous scheduling signals, from smallest to largest, correspond to local development, balanced search, and global exploration tendencies, respectively. The one-dimensional continuous scheduling signal is represented as follows: in, The original actions sampled from a one-dimensional Gaussian policy, is the corresponding one-dimensional continuous scheduling signal; Clip is the interval clipping function, whose first parameter is the value to be clipped, and the second and third parameters are the lower limit 0 and the upper limit 1, respectively. When it is less than the lower limit, the lower limit is taken, and when it is greater than the upper limit, the upper limit is taken. In other cases, the value to be clipped is kept. The proximal policy optimization network is trained offline in the following way: the policy network interacts with the population evolution process, the difference between the current generation's global best fitness and the next generation's global best fitness is divided by the sum of the absolute value of the current generation's global best fitness and the numerically stable term, and the result is used as an immediate reward. The proximal policy optimization objective with a pruning probability ratio is used to update the policy network parameters. The instant reward is represented as follows: in, For the first The instant reward corresponding to the action. and The first generation and first The global optimal fitness of the generation; in the near-end policy optimization update, the probability ratio of the old and new policies is based on the original action. calculate.

7. The hierarchical adaptive hyperheuristic robot path planning method according to claim 1, characterized in that, The ordered anchor points corresponding to the development, balanced, and exploratory functional operator pools are 1 / 6, 3 / 6, and 5 / 6, respectively; in, , and These are the ordered anchor points for the development, balanced, and exploratory functional operator pools, respectively; Calculate the absolute distance between the one-dimensional continuous scheduling signal and each ordered anchor point, normalize the absolute distance using the corresponding normalized distance factor, and obtain the distance response of the corresponding functional operator pool, expressed as: in, For a one-dimensional continuous scheduling signal of generation t; For the first Ordered anchor points of a pool of functional operators; For the first A positive normalized distance factor for each functional operator pool; λ is a positive response intensity coefficient; λ is the distance response function. For the first Each functional operator pool in scheduling signals Distance response; Calculate the macroscopic weights of each functional operator pool so that the closer the one-dimensional continuous scheduling signal is to a certain ordered anchor point, the greater the distance response of the functional operator pool corresponding to that ordered anchor point. The calculation formula is as follows: in, For the first The macroscopic weight function of the functional operator pool; The temperature coefficient is positive, and exp is the natural exponential function. For the first The distance response of the three functional operator pools under the same scheduling signal is equal to the sum of their macroscopic weights, which is 1.

8. The hierarchical adaptive hyperheuristic robot path planning method according to claim 1, characterized in that, The division of multiple low-level heuristic operators into three functional operator pools—developmental, balanced, and exploratory—is specifically as follows: The exploratory functional operator pool includes an elite search operator based on Lévy perturbation, a random difference mutation operator DE / rand / 1, and a random individual binomial crossover operator; The balanced functional operator pool includes an elite search operator based on Gaussian perturbation, a current individual differential evolution operator DE / current / 1, an optimal individual differential evolution operator DE / best / 1, and an elite-guided differential evolution operator DE / current-to-pbest / 1. The development-oriented functional operator pool includes an elite neighborhood search operator based on uniform perturbation, a global optimal guided differential evolution operator DE / current-to-best / 1 operator, and an elite individual binomial crossover operator. The elite individual set consists of the top 10% of individuals in the population whose fitness is sorted from smallest to largest. The elite-guided differential evolution operator DE / current-to-pbest / 1 provides elite guidance through the centroids of the top 10% of individuals whose fitness is sorted from smallest to largest.

9. The hierarchical adaptive hyperheuristic robot path planning method according to claim 1, characterized in that, The step of initializing the historical preference probability vector for each low-level heuristic operator for each candidate path individual includes: assigning the same probability to all low-level heuristic operators during the initialization of the historical preference probability vector for each candidate path individual; When the The low-level heuristic operator belongs to the . When the pool of functional operators is in its first phase, the first... The candidate path individual selects the first... The joint selection probability of the low-level heuristic operators is: in, For the first Individual selection of candidate paths The joint selection probability of a low-level heuristic operator; For the first The macroscopic weight function of the functional operator pool; For the first The set of indexes of low-level heuristic operators contained in a functional operator pool; For the first Individual with candidate path for the first Historical preference probabilities of low-level heuristic operators; Each functional operator pool satisfies the following probability conservation condition: Among them, the probability conservation formula For traversal The summation index of all operators within the pool; the probability conservation ensures that individual historical preferences only change the selection share of each low-level heuristic operator within the same functional operator pool, while keeping the macro weight of the functional operator pool unchanged.

10. A hierarchical adaptive hyperheuristic robot path planning method according to claim 1, characterized in that, The process involves sampling low-level heuristic operators according to the joint selection probability to generate offspring candidate paths, wherein when the selected first... The low-level heuristic operator is the first... When the fitness of the offspring generated by an individual along a candidate path is less than that of its parent, the historical preference probability of that individual is updated as follows: And the historical preference probabilities of the remaining low-level heuristic operators are proportionally decayed: in, The previous iteration generation. This refers to the generation after this update; and The first The candidate path individual selects the first... Historical preference probabilities before and after operator update and These are the individual's views on the first... Historical preference probabilities before and after updating the remaining operators; The learning rate is greater than 0 and less than 1; the sum of all probabilities in the updated historical preference probability vector is 1. When the fitness of the offspring is not less than that of the parent, the historical preference probability vector of the individual remains unchanged, and the historical preference probability is updated based on whether the offspring is better than the parent.