GP-based hierarchical unmanned cluster task planning method

Abstract

The invention discloses a hierarchical unmanned cluster task planning method and device based on GP, equipment and a medium. The method comprises the following steps: constructing a hierarchical task individual tree with structured semantics; taking a strong type genetic programming algorithm as a population evolution baseline, inputting the hierarchical task individual tree into the strong type genetic programming algorithm to generate an initial population, and evaluating the performance of each initial individual to obtain a fitness evaluation result; judging that a termination condition is not met according to a fitness evaluation result, performing selection, crossover and mutation operation on each individual by utilizing a strong type genetic programming algorithm, and cleaning redundant nodes in each individual by utilizing a pruning algorithm based on domain knowledge in the crossover and mutation operation process; and converting the optimal hierarchical task individual tree into a task planning scheme of the unmanned cluster collaborative system. The technical problems of unclear task structure expression ability, imperfect feasibility guarantee mechanism, low search efficiency and no self-adaptability in the existing unmanned cluster collaborative task planning method are solved.

Description

Technical Field

[0001] The present application relates to the technical field of intelligent unmanned systems, and more specifically to a GP-based hierarchical unmanned cluster task planning method, device, equipment and medium. Background Art

[0002] With the widespread application of intelligent unmanned systems in disaster response, environmental monitoring, and urban security, multi-agent collaborative task planning has become a key research area for unmanned swarm systems. For example, in emergency firefighting scenarios, a collaborative system consisting of an unmanned ground vehicle (UGV) and multiple unmanned aerial vehicles (UAVs) must rapidly cover and address multiple fires, achieving efficient, low-cost, and coordinated firefighting scheduling.

[0003] In such tasks, the tasks have obvious hierarchical structural characteristics: UGV, as the upper-level command agent, is responsible for subtask allocation, path coordination and task parameter issuance, while UAV, as the lower-level executor, undertakes the specific target fire extinguishing task. In order to ensure scheduling efficiency and rational resource utilization, the system needs to realize automatic planning and intelligent scheduling of tasks based on conditions such as the task area, drone energy limit, and fire point distribution. However, the research on this type of hierarchical multi-agent collaborative task planning problem still has the following technical difficulties: (1) Unclear task modeling structure: Traditional methods (such as centralized planning, graph search or linear programming) often cannot explicitly express the hierarchical structure of tasks and the interactive relationship between upper and lower layers, lack structural semantic support, and make scheduling schemes difficult to explain and expand. (2) Weak task constraint processing capabilities: Traditional optimization algorithms lack effective structural level verification and repair mechanisms when faced with feasibility problems such as "drone task allocation overload", "duplicate coverage of target points" or "uncovered targets", and are prone to generating illegal or inefficient scheduling schemes. (3) The search space is large in dimension and highly redundant: Traditional optimization algorithms cannot actively avoid redundant areas in coding design, resulting in a large amount of unnecessary searches and slow convergence. This problem is particularly serious in multi-objective and multi-agent environments. Summary of the Invention

[0004] The present application aims to at least solve the technical problems existing in existing unmanned cluster collaborative task planning methods, such as unclear task structure expression capability, imperfect feasibility guarantee mechanism, inefficient search and non-adaptiveness. To this end, the first aspect of the present application proposes a hierarchical unmanned cluster task planning method based on GP, which includes the following execution steps: based on the domain knowledge of the hierarchical multi-agent system, five types of nodes of the unmanned cluster collaborative system jointly constructed by the fire unmanned vehicle and the multi-UAV group are determined, and a hierarchical task individual tree with structured semantics is constructed based on the five types of nodes; a strongly typed genetic programming algorithm is used as the population evolution baseline, and the hierarchical task individual tree is input into the strongly typed genetic programming algorithm to generate an initial population, wherein the initial individuals of each initial population are task individual trees of each level, and each level of task individual tree corresponds to the deployment plan of each UAV group and fire unmanned vehicle when performing tasks, and a predefined fitness function is used to evaluate the fitness of the individual trees. Estimate the performance of each initial individual and obtain the fitness evaluation result; determine whether the termination condition is met based on the fitness evaluation result. If the termination condition is met, output the corresponding optimal hierarchical task individual tree based on the optimal fitness evaluation result; otherwise, use the strongly typed genetic programming algorithm to perform selection, crossover, and mutation operations on each individual, and use the pruning algorithm based on domain knowledge to clean up the redundant nodes in each individual during the crossover and mutation operations to generate a descendant hierarchical task individual tree, use the fitness function to evaluate the performance of each group individual, obtain a new fitness evaluation result, iteratively execute the previous execution step, and convert the optimal hierarchical task individual tree into a task planning scheme for the unmanned swarm collaborative system.

[0005] Optionally, the five types of nodes include overall task nodes, lower-level subtask nodes, interactive subtask nodes and upper-level subtask nodes, wherein the overall task nodes, upper-level subtask nodes, interactive subtask nodes and lower-level subtask nodes have a parent-child inheritance relationship, and the upper-level subtask nodes include real subtask nodes and virtual subtask nodes; the redundant nodes in each individual are cleaned up using a pruning algorithm based on domain knowledge during the crossover and mutation operations, including: obtaining the initial individual after the crossover; based on the basic subtree mutation operator of the strongly typed genetic programming algorithm, selecting a subtree that is not a root node from each initial individual after the crossover as a randomly generated new subtree to obtain the mutated individual; calculating the current participating individual according to the lower-level subtask nodes of the mutated individual The number of lower-level intelligent agents for extinguishing fires, wherein the lower-level subtask nodes are used to determine the fire points corresponding to each lower-level intelligent agent, and the lower-level intelligent agent is a drone; judging whether the number of lower-level intelligent agents is greater than the number of fire points, if it is greater than the number of fire points, randomly generating an intermediate number within the range of the number of fire points, and judging whether the number of lower-level intelligent agents is greater than the intermediate number, if it is greater than the intermediate number, randomly pruning the lower-level subtask nodes until the number of lower-level intelligent agents is less than or equal to the number of fire points; wherein, in the process of pruning the lower-level subtask nodes, judging that all lower-level subtask nodes under the interactive subtask node are pruned, then the entire subtree with the real subtask node as the root is pruned, and using the virtual subtask node to replace the real subtask node to connect the overall task node.

[0006] Optionally, the five types of nodes also include parameter nodes; the overall task node is connected to each upper-level sub-task node, and is used to determine the overall fire-fighting task of the unmanned cluster collaborative system during fire-fighting based on the upper-level sub-task nodes, and the execution order of the upper-level sub-task nodes is determined according to the connection order of the overall task nodes; the real sub-task nodes of the upper-level sub-task nodes include interactive sub-task nodes and parameter nodes, and the parameter nodes are used to store the interactive parameters inside and outside the unmanned cluster collaborative system, wherein the interactive parameters include path coordinate parameters and energy consumption parameters, and the interactive sub-task nodes represent the fire point corresponding to each drone, as well as the collaborative path between the drone group and the unmanned fire truck; the virtual sub-task nodes of the upper-level sub-task nodes are used to use the virtual sub-task nodes as placeholders when the number of real sub-tasks is less than the number of sub-nodes required according to the overall task node.

[0007] Optionally, the hierarchical task variable tree is a four-layer structure, and the four-layer structure is, from top to bottom, a root layer, an upper subtask layer, an interactive subtask layer, and a lower subtask layer; wherein, the overall task node belongs to the root layer; the upper subtask node belongs to the upper subtask layer; the interactive subtask node and the parameter node belong to the interactive subtask layer; and the lower subtask node belongs to the lower subtask layer.

[0008] Optionally, before obtaining the initial individuals after the crossover, the method further includes: randomly selecting multiple initial individuals from the initial population based on a tournament selection operator; randomly selecting intersection points of the same type from each initial individual based on a conventional subtree crossover operator of a strongly typed genetic programming algorithm, and exchanging subtrees with the intersection points as root nodes between the initial individuals to obtain the individuals after the crossover.

[0009] Optionally, before generating the initial population, the method further includes: predefining node sets and node connection rules, wherein the node sets and node connection rules are used to generate constraint conditions so that the hierarchical task variable tree structure is legal and complies with task constraints.

[0010] Optionally, the fitness function is expressed as:

[0011] in, , , , Indicates the fire point The percentage of firefighting tasks completed, Indicates the number of drones actually deployed, Indicates the number of drones to be deployed, represents the total mission time of each UAV, Indicates the maximum total working time of all drones, Indicates the degree of task completion, Indicates resource utilization efficiency, Indicates time consumption efficiency.

[0012] In order to achieve the above-mentioned objectives, the second aspect of the present application provides a hierarchical unmanned swarm task planning method based on GP, characterized by comprising: The individual number construction module is used to determine the five types of nodes in the unmanned swarm collaborative system jointly constructed by firefighting unmanned vehicles and multiple drones based on the domain knowledge of hierarchical multi-agent systems, and to construct a hierarchical task individual tree with structured semantics based on the five types of nodes; The fitness calculation module is used to use a strongly typed genetic programming algorithm as the population evolution baseline, input the hierarchical task individual tree into the strongly typed genetic programming algorithm to generate an initial population, wherein the initial individuals of each initial population are the task individual trees of each level, and each level of task individual tree corresponds to the deployment plan of each drone group and firefighting unmanned vehicle when performing a task. A predefined fitness function is used to evaluate the performance of each initial individual to obtain a fitness evaluation result; The task planning scheme generation module is used to determine whether the termination condition is met based on the fitness evaluation results. If the termination condition is met, the corresponding optimal hierarchical task individual tree is output according to the optimal fitness evaluation result; otherwise, the strongly typed genetic programming algorithm is used to select, crossover, and mutate each individual, and the pruning algorithm based on domain knowledge is used to clean up the redundant nodes in each individual during the crossover and mutation operations to generate a descendant hierarchical task individual tree. The fitness function is used to evaluate the performance of each group individual to obtain a new fitness evaluation result, and the previous execution step is iteratively executed, and the optimal hierarchical task individual tree is converted into a task planning scheme for the unmanned cluster collaborative system.

[0013] In order to achieve the above-mentioned purpose, the third aspect of the present application provides an electronic device, characterized in that the electronic device includes a processor and a memory, and the memory stores at least one instruction, at least one program, a code set or an instruction set, and the at least one instruction, the at least one program, the code set or the instruction set is loaded and executed by the processor to implement the hierarchical unmanned cluster task planning method based on GP described in any of the previous methods.

[0014] In order to achieve the above-mentioned purpose, the fourth aspect of the present application provides a computer-readable storage medium, characterized in that the storage medium stores at least one instruction, at least one program, code set or instruction set, and the at least one instruction, the at least one program, the code set or instruction set is loaded and executed by the processor to implement the GP-based hierarchical unmanned cluster task planning method described in any one method.

[0015] The present invention provides a hierarchical unmanned swarm task planning method, apparatus, device, and medium based on Genetic Modulation (GP). Compared with existing technologies, the method has the following advantages: Based on the domain knowledge of a hierarchical multi-agent system, five types of nodes in an unmanned swarm collaborative system jointly constructed by a firefighting unmanned vehicle and a swarm of multiple unmanned vehicles are identified, and a hierarchical task individual tree with structured semantics is constructed based on the five types of nodes. A strongly typed genetic programming algorithm is used as the population evolution baseline, and the hierarchical task individual tree is input into the strongly typed genetic programming algorithm to generate an initial population. The initial individuals of each initial population are task individual trees at each level, and each task individual tree corresponds to the deployment plan of each unmanned vehicle swarm and firefighting unmanned vehicle when executing the task. A predefined fitness function is used to evaluate the performance of each initial individual, and a fitness evaluation result is obtained. During the population initialization phase, the method abstracts five types of task nodes based on the hierarchical characteristics of HMAS tasks and combines domain knowledge to construct a node set structure with hierarchical semantics. By introducing a strongly typed genetic programming (STGP) mechanism, node connection relationships and syntax types are strictly limited during the individual construction process, ensuring that each individual tree has a clear division of tasks between upper and lower levels and reasonable scheduling logic. This design avoids the generation of illegal structures such as task allocation conflicts, redundant target assignments, and broken execution paths from the source, greatly reducing the occurrence of invalid combinations. By integrating structural constraints with domain knowledge, the feasibility of individuals in the initial population is greatly improved, effectively compressing the dimension of the search space, improving the structural quality of the population and the initial fitness of the scheduling scheme, and providing a more efficient and stable search starting point for the subsequent evolutionary process; based on the fitness evaluation results, it is determined whether the termination conditions are met. If the termination conditions are met, the corresponding optimal hierarchical task individual tree is output based on the optimal fitness evaluation results; otherwise, a strongly typed genetic programming algorithm is used to select, crossover, and mutate each individual, and a pruning algorithm based on domain knowledge is used during the crossover and mutation operations to clean up redundant nodes in each individual, generate a descendant hierarchical task individual tree, use the fitness function to evaluate the performance of each population individual, obtain a new fitness evaluation result, and iteratively execute the previous execution step, and convert the optimal hierarchical task individual tree into a task planning scheme for the unmanned cluster collaborative system. This application uses a pruning mechanism to remove redundant leaf nodes and redundant branches, dynamically shrinking the invalid structure space. This mechanism significantly improves the individual feasibility ratio, ensuring that the solutions generated during the evolution process meet the scheduling resource restrictions and task logic constraints; the pruning operation of this mechanism effectively compresses the search space dimension and improves the convergence efficiency of the optimization algorithm; this pruning mechanism effectively improves the structural adaptability and search performance of the algorithm on the basis of ensuring the legality of scheduling. This application can ensure that the task planning solution always maintains structural feasibility and scheduling rationality during the evolution process, and further solves the technical problems of unclear task structure expression ability, imperfect feasibility guarantee mechanism, inefficient search and non-adaptiveness in the existing unmanned cluster collaborative task planning methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] To more clearly illustrate the technical solution of this application, the following briefly introduces the drawings required for use in the embodiments or prior art descriptions. Obviously, the drawings described below are merely some embodiments of this application, and those skilled in the art can derive other drawings based on these drawings without inventive effort.

[0017] Figure 1 A flowchart of a GP-based hierarchical unmanned cluster task planning method provided in an embodiment of the present application; Figure 2 A schematic diagram of a firefighting task scenario in a hierarchical multi-agent system using a GP-based hierarchical unmanned cluster task planning method according to an embodiment of the present application; Figure 3 A schematic diagram of the structure of an individual tree of a GP-based hierarchical unmanned swarm task planning method provided in an embodiment of the present application; Figure 4 A schematic diagram of an infeasible solution generated during the individual tree intersection process of a GP-based hierarchical unmanned swarm task planning method provided in an embodiment of the present application; Figure 5 A hierarchical unmanned cluster task planning method based on GP is provided in the embodiment of the present application. and Example task flow diagram; Figure 6 This is a structural block diagram of a GP-based hierarchical unmanned cluster task planning device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0018] The following will be combined with the accompanying drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0019] This specification provides method operation steps as described in the embodiments or flowcharts, but more or fewer operation steps may be included based on routine or non-inventive work. When implemented in a real system or server product, the methods shown in the embodiments or figures may be executed sequentially or in parallel (for example, in a parallel processor or multi-threaded processing environment).

[0020] In response to the technical problems existing in the background technology of the existing technology, this application proposes a hierarchical unmanned cluster task planning method based on genetic programming (GP), which is used to solve the problems of difficulty in expressing task structure and low search efficiency when performing resource minimization and rapid coverage tasks for multiple fire points in the "firefighting unmanned vehicle (UGV)-multiple unmanned aerial vehicles (UAV)" collaborative system.

[0021] Based on a hierarchical multiagent system (HMAS), this application constructs a four-layer task variable tree with structured semantics, consisting of an overall task node, upper-level subtask nodes, interactive subtask nodes, lower-level subtask nodes, and parameter nodes. This structure leverages domain knowledge extracted from the task execution process to design node sets and model type connection constraints. This ensures expressive task encoding, clear structure, and legal connections, significantly improving the accuracy and systematicity of task scheduling modeling.

[0022] In terms of optimization algorithm design, this application uses strongly typed genetic programming (STGP) as the evolutionary baseline. Through selection, subtree crossover, and mutation, individual trees are evolved and optimized to gradually generate higher-quality scheduling solutions. To address issues such as node redundancy, target duplication, and task allocation overruns that may arise during crossover and mutation, a domain-knowledge-based pruning and repair mechanism is designed. This automatically prunes redundant subtrees and introduces virtual nodes to fill in their place, ensuring that the generated solution maintains structural feasibility and scheduling rationality throughout the evolutionary process.

[0023] Assuming that the unmanned swarm mission planning scenario is a firefighting unmanned vehicle / unmanned aerial vehicle (UGV / UAV) collaborative system, the unmanned swarm mission planning scenario is as follows: Figure 2 As shown, Figure 2 This is a firefighting mission scenario in HMAS. Figure 2 Includes 1 transport UGV (upper intelligent body) and up to The UGV and the loaded drone initially reside at the red dot on the lower left. The orange star indicates the fire point. The number of fire points is In this scenario, each agent moves at a constant speed, and the UGV is usually slower than the UAV. In this scenario, the goal of task scheduling is to have the minimum number of drones cover all fire points as quickly as possible. Since the number of drones called n is the decision variable in the task scheduling problem, its different choices will lead to dimensional changes in the task parameters and cross-layer interaction parameters of the upper and lower-level agents, which makes HMAS task scheduling a unique and challenging problem with a variable-sized decision space.

[0024] The technical solution of the present application is further described below with reference to the accompanying drawings and embodiments: refer to Figure 1 , Figure 1 This is a flowchart of the GP-based hierarchical unmanned swarm task planning method proposed in this application. The method may include the following execution process: S10. Based on the domain knowledge of hierarchical multi-agent systems, identify five types of nodes in the unmanned swarm collaborative system constructed by unmanned firefighting vehicles and multi-UAV swarms, and construct a hierarchical task individual tree with structured semantics based on the five types of nodes. Among them, the five types of nodes include overall task nodes, lower-level subtask nodes, interactive subtask nodes and upper-level subtask nodes, among which the overall task nodes, upper-level subtask nodes, interactive subtask nodes and lower-level subtask nodes have a parent-child inheritance relationship, and the upper-level subtask nodes include real subtask nodes and virtual subtask nodes.

[0025] It should be noted that during the population initialization phase, based on the hierarchical nature of HMAS tasks and incorporating domain knowledge, five types of task nodes are abstracted: overall tasks, upper-level subtasks, interactive subtasks, lower-level subtasks, and parameter nodes. This creates a node set structure with hierarchical semantics. By introducing a strongly typed genetic programming (STGP) mechanism, node connectivity and syntax types are strictly constrained during individual construction, ensuring that each individual tree has a clear upper- and lower-level task division and reasonable scheduling logic. This design fundamentally avoids the generation of illegal structures such as task allocation conflicts, redundant target assignments, and broken execution paths, significantly reducing the occurrence of invalid combinations. By integrating structural constraints with domain knowledge, the feasibility of individuals in the initial population is significantly improved, effectively compressing the dimensionality of the search space, improving the population's structural quality and the initial fitness of the scheduling scheme, and providing a more efficient and stable search starting point for subsequent evolution.

[0026] In one embodiment of the present application, the hierarchical task variable tree is a four-layer structure, and the four-layer structure is respectively a root layer, an upper subtask layer, an interactive subtask layer, and a lower subtask layer from top to bottom; Among them, the overall task node belongs to the root layer; the upper subtask node belongs to the upper subtask layer; the interactive subtask node and the parameter node belong to the interactive subtask layer; and the lower subtask node belongs to the lower subtask layer.

[0027] Specifically, Figure 3 is a schematic diagram of the structure of the individual tree. Figure 3 As shown in the figure, nodes are set according to the four-layer structure of the individual tree, and then the individual tree is constructed using the nodes. The nodes are divided into five categories: overall task nodes, upper-level subtask nodes, interactive subtask nodes, lower-level subtask nodes, and parameter nodes.

[0028] In one embodiment of the present application, the five types of nodes further include parameter nodes; the overall task node is connected to each upper-level subtask node, and is used to determine the overall fire extinguishing task of the unmanned cluster collaborative system during fire extinguishing according to the upper-level subtask nodes, and the execution order of the upper-level subtask nodes is determined according to the connection order of the overall task node; For example, the overall task node As the root node of the variable tree, connect The execution order of the upper-level subtask nodes depends on the order in which they are connected to the overall task node.

[0029] The real subtask node of the upper-level subtask node includes an interaction subtask node and a parameter node. The parameter node is used to store interaction parameters inside and outside the unmanned cluster collaborative system, wherein the interaction parameters include path coordinate parameters and energy consumption parameters. The interaction subtask node represents the fire point corresponding to each drone, as well as the collaborative path between the drone swarm and the unmanned fire truck. For example, the upper-level subtask node represents the upper-level subtask, and the overall task node is used as the parent node to form a tree structure. This type of node is divided into two categories: real subtask nodes and virtual subtask nodes: a real subtask node contains two child nodes, the first is the interaction subtask node, and the other is the parameter node that stores the corresponding interaction parameters; the virtual subtask node is only used as a placeholder when the number of required real subtasks is less than the number of child nodes required by the root node. The upper-level subtask node corresponds to the area allocation or drone task scheduling that the UGV is responsible for. All upper-level subtask nodes constitute a set .

[0030] The virtual subtask node of the upper subtask node is used as a placeholder when the number of real subtasks is less than the number of subnodes required by the overall task node.

[0031] For example, the interactive subtask node is both a child node of the upper subtask node and a parent node of the lower subtask node, and is used to describe the cross-level interaction between the upper and lower intelligent agents. The interactive subtask node reflects the collaboration and path interaction between UGV and UAV, and determines how many UAVs are deployed in the subtask. All interactive subtask nodes constitute the interactive subtask node set .

[0032] Table 1 Experimental node set

[0033] For example, the lower-level subtask node describes in detail the subtask that the lower-level agent ultimately performs. Since the subtasks of the lower-level agent are usually simple and have a single function, these nodes are the terminal nodes (i.e., leaf nodes) of the variable tree. The lower-level subtask node represents the fire point ID corresponding to each UAV. All lower-level subtask nodes constitute the lower-level subtask node set. .

[0034] For example, as a terminal node, the parameter node is connected to the upper subtask node to provide coordinate parameters for the interactive subtask. The parameter node contains specific task parameters such as path coordinates and energy consumption. All parameter nodes constitute the parameter node set .

[0035] From this we can get, Given When A variable tree covering upper-level agent path planning (upper-level subtask nodes and parameter nodes), cross-level interaction (interaction subtask nodes), and lower-level agent task allocation (lower-level subtask nodes), where Indicates the population size. Node set The definitions are shown in Table 1.

[0036] According to the task scenario settings, the hierarchical task variable tree can be set to follow a four-layer structure: from top to bottom, the root layer , upper subtask layer , interactive subtask layer and the lower subtask layer Based on the domain knowledge extracted from the execution of HMAS general tasks, a node set containing 5 types of nodes is defined: To build a variable tree, we need to ensure full coverage of the solution space. When using nodes to build a hierarchical task variable tree, we need to clearly define which types of nodes can serve as parent nodes, which types of nodes must contain parameters or child nodes, and the semantic meaning of the node connection order (such as the order in which tasks are executed) based on domain knowledge.

[0037] Before generating the initial population, the method may further include the following execution process: Node sets and node connection rules are predefined, where the node sets and node connection rules are used to generate constraint conditions so that the hierarchical task variable tree structure is legal and complies with task constraints.

[0038] That is to say, in the initialization phase, the processor needs to analyze the task scenario and participating agents, extract domain knowledge from task characteristics and constraints, and use the domain knowledge to guide the representation of decision variables to generate the initial population. This application adopts Strongly-Typed GP (STGP) as the evolutionary baseline, integrates HMAS domain knowledge, and proposes a tree-shaped individual representation method to generate the initial population. In the initial population, the node set and connection rules are defined according to the characteristics of the task to ensure that the generated individual tree structure is legal and meets the constraints. Each individual tree is an individual in the initial population, and each individual corresponds to a task scheduling plan. In the process of using nodes to construct an individual tree, the processor can clearly specify which type of node can be used as a parent node, which type of node must contain parameters or child nodes, and the semantic meaning of the node connection order (such as the order of task execution) based on domain knowledge.

[0039] After the initialization phase, the search phase begins. The processor evaluates each individual in each generation of the population and obtains its corresponding fitness function value. If the termination condition is met, the optimal individual is saved as the task scheduling solution. Otherwise, offspring are generated through mutation, crossover, and selection, and the optimal individual is further selected. The crossover, mutation, and selection operators essentially inherit the traditional STGP method. While these crossover and mutation operations can ensure the feasibility of node type connections, they cannot guarantee that the offspring will meet the global task constraints of the underlying agent. Figure 4 It is a schematic diagram of the infeasible solutions generated during the individual tree crossover process.

[0040] The above technical issues are further explained below. Figure 4 As shown, assume that there is a constraint requiring that the number of underlying subtasks represented by the green terminal node be less than 4. Both parent individuals (A) and (B) satisfy the constraint, but after exchanging the subtrees within the black dashed box, although the subtree root types are the same and all node positions are correct, the offspring (A') contains 4 green terminal nodes and violates the constraint. Mutation operations may also lead to infeasible solutions. In order to solve the problem of disordered individual tree structure caused by infeasible solutions generated during crossover and mutation, this application designs a pruning scheme based on domain knowledge to correct these infeasible individuals, so as to clean up redundant nodes, reconstruct the path order, and improve the rationality of the structure and the feasibility of the scheduling plan.

[0041] Figure 1This is a flowchart for solving the task planning model based on GP. As shown in Figure 1, the process of solving the task planning model based on GP is as follows: first, the processor can extract domain knowledge and build a task hierarchy model. Then, in the population initialization phase, the domain knowledge is used to guide the design of the node set to ensure that the generated individual trees are all feasible and have no redundant decision space, and the fitness function is used to evaluate the fitness of the individual trees to obtain the fitness evaluation result, which is the content of step S20 of this application.

[0042] Furthermore, the search phase may specifically include the following execution processes: S20. Using a strongly typed genetic programming algorithm as a population evolution baseline, inputting the hierarchical task individual tree into the strongly typed genetic programming algorithm to generate an initial population, wherein the initial individuals of each initial population are the hierarchical task individual trees, and each hierarchical task individual tree corresponds to the deployment plan of each drone swarm and firefighting unmanned vehicle when performing a task. A predefined fitness function is used to evaluate the performance of each initial individual to obtain a fitness evaluation result; Specifically, the first step in the search phase can be fitness evaluation: The processor can evaluate each individual in each generation of the population to obtain the corresponding fitness function value.

[0043] Among them, the fitness function is shown in formula (1): (1) Where, ; ; ; Indicates the fire point The percentage of firefighting tasks completed; Indicates the number of drones actually deployed; Indicates the number of drones to be deployed; Indicates the total mission time of each UAV; Indicates the maximum total operating time of all drones.

[0044] Among them, this function combines the three core indicators of task completion ( ), resource utilization efficiency ( ), time consumption efficiency ( ) into a single metric, which comprehensively evaluates the overall performance of scheduling solutions and avoids the bias caused by single-objective optimization. A higher fitness value indicates a better solution.

[0045] It is worth noting that task completion Appearing as a multiplication factor in the main term of the function, it directly drives the scheduling plan to optimize towards a higher target coverage; exponential decay term A joint penalty mechanism for resource usage and time consumption is introduced to explicitly suppress redundant scheduling and delayed execution while ensuring task effectiveness, thereby prompting the algorithm to generate a more compact and efficient task allocation structure. Furthermore, this function structure exhibits normalization properties, adapting to task parameter configurations of varying scales. It can be efficiently calculated without relying on complex simulations, facilitating integration into evolutionary search processes. Overall, this design effectively guides the algorithm to strike a balance between task completion, resource conservation, and execution efficiency, improving the optimization capabilities and practical usability of the overall scheduling system.

[0046] In one embodiment of the present application, before obtaining the initial individual after the crossover, the method may further include the following execution process: Randomly select multiple initial individuals from the initial population based on the tournament selection operator; For example, the processor uses the widely used tournament selection operator. Randomly select from the population Individuals with the best fitness in each batch are selected into the next generation population.

[0047] Based on the conventional subtree crossover operator of the strongly typed genetic programming algorithm, crossover points of the same type are randomly selected from each initial individual, and subtrees with the crossover points as root nodes are exchanged between the initial individuals to obtain individuals after each crossover.

[0048] Exemplarily, the processor follows the majority GP method, adopts a conventional subtree crossover operator, randomly selects a crossover point of the same type from the parent individuals, and exchanges the subtree with the point as the root.

[0049] S30. Determine whether the termination condition is met based on the fitness evaluation result. If the termination condition is met, output the corresponding optimal hierarchical task individual tree based on the optimal fitness evaluation result; otherwise, use the strongly typed genetic programming algorithm to select, cross, and mutate each individual, and use the pruning algorithm based on domain knowledge to clean up the redundant nodes in each individual during the crossover and mutation operations to generate a descendant hierarchical task individual tree, use the fitness function to evaluate the performance of each group individual, obtain a new fitness evaluation result, and iteratively execute the previous execution step, and convert the optimal hierarchical task individual tree into a task planning scheme for the unmanned cluster collaborative system. After executing the above pseudo code, the processor determines whether the termination condition is met. If the termination condition is met, In one embodiment of the present application, the process of using a pruning algorithm based on domain knowledge to clean up redundant nodes in each individual during the crossover and mutation operations may include the following execution process: Get the initial individual after crossover; Based on the basic subtree mutation operator of the strongly typed genetic programming algorithm, a subtree that is not a root node is selected from the initial individuals after each crossover as a randomly generated new subtree to obtain the mutated individuals; Calculate the number of lower-level agents currently participating in firefighting based on the lower-level subtask nodes of the mutated individuals, where the lower-level subtask nodes are used to determine the fire point corresponding to each lower-level agent, where the lower-level agents are drones; Determine whether the number of lower-level agents is greater than the number of fire points. If so, randomly generate an intermediate number within the range of the number of fire points and determine whether the number of lower-level agents is greater than the intermediate number. If so, randomly remove lower-level subtask nodes until the number of lower-level agents is less than or equal to the number of fire points. Among them, in the process of pruning lower-level subtask nodes, if it is determined that all lower-level subtask nodes under the interactive subtask node are pruned, the entire subtree with the real subtask node as the root will be pruned, and the virtual subtask node will be used to replace the real subtask node to connect to the overall task node.

[0050] During the specific execution process, the processor can first determine whether the termination condition is met based on the fitness evaluation result. If the termination condition is met, the optimal individual tree is output and the scheduling plan is determined. If the termination condition is not met, the offspring individuals are selected, crossed, and mutated to generate a new individual tree.

[0051] During the generation of a new individual tree, the processor can further apply domain knowledge to prune infeasible individual trees and restructure the pruned individual trees to generate a new, constrained individual tree. The processor can then perform a fitness evaluation on the newly generated individual tree, obtaining a fitness evaluation result. Finally, based on the fitness evaluation result, the processor can again determine whether the termination condition is met. This cycle continues until the termination condition is met, resulting in the optimal individual tree and, in turn, a task scheduling solution.

[0052] For example, the processor can use the basic subtree mutation operator of the strongly typed genetic programming algorithm to replace the subtree with a randomly selected node as the root with a randomly generated new subtree. The root node type of the new subtree is consistent with the atomic tree to ensure the feasibility of the connection. Next, the processor first calculates the number of currently participating lower-level intelligent agents based on the lower-level subtask nodes. (Lines 1-2 in pseudo code). Infeasible individuals, randomly select and remove nodes until the (Lines 3-6 in the pseudocode). If all child nodes of an interactive subtask node are removed, the entire subtree rooted at that upper-level subtask node is pruned, and its connection to the root node is replaced by a virtual upper-level subtask node (lines 7-9 in the pseudocode). The introduction of virtual nodes allows truncating redundant subtrees when all agents are no longer needed, thereby compressing the decision space and improving search efficiency.

[0053] For example, when the processor is executing the pruning algorithm of the domain knowledge, the processor may set The number of layer nodes is , the corresponding constraint on the number of nodes can be expressed as shown in Expression (2): (2) in, and constitutes the feasible domain.

[0054] It should be noted that since only the interactive subtask nodes have uncertain output dimensions, this means that there is no constraint on the number of lower-level subtask nodes during the evolution process. Further considering the HMAS application scenario, the processor converts the global constraint of expression (2) into a limit on the number of participating lower-level agents, as shown in expression (3): (3) in, is a knowledge-based converter that converts Mapped to the number of participating lower-level agents .

[0055] Algorithm 1: Pseudocode of pruning operation

[0056] It is also important to note that since the node type and link constraints have ensured that each interactive subtask node has at least one lower-level subtask node, only Exceeding the upper limit The processor uses expression (3) to reasonably prune redundant lower-level subtask nodes, transforming infeasible offspring after crossover mutation into feasible individuals. This pruning operator not only ensures the feasibility of the solution but also introduces a certain degree of randomness by randomly selecting the nodes to be pruned, which is beneficial for maintaining population diversity. In actual operation, the pseudocode of the pruning operation can be found in Algorithm 1.

[0057] The beneficial effects of the above steps are as follows: the processor executes the above crossover and mutation steps, and during the mutation step, uses a pruning mechanism to remove redundant leaf nodes and redundant branches, dynamically shrinking the invalid structure space. The processor counts the number of lower-level subtask nodes in the individual tree and uses the knowledge converter mapping to calculate the actual number of participating lower-level agents to determine whether the resource constraint limit is exceeded. If there is task overload, the pruning operation will randomly trim excess lower-level subtask nodes to control the scheduling scale. At the same time, virtual task nodes will be automatically inserted to fill the gaps in the structure to maintain the structural closure and grammatical integrity of the individual tree. This mechanism offers multiple technical advantages: First, it significantly increases the proportion of individual feasibility, ensuring that solutions generated during the evolution process meet scheduling resource constraints and task logic constraints. Second, by actively removing redundant nodes and invalid task paths, pruning effectively compresses the search space dimension, improving the convergence efficiency of the optimization algorithm. Third, the introduction of virtual nodes enhances the scalability and flexibility of the scheduling structure, supporting dynamic changes in the number of tasks without disrupting the grammatical structure. Finally, the randomized pruning strategy ensures constraint satisfaction while preserving the structural diversity of the population, helping to improve overall search capabilities and avoid falling into local optima. Consequently, this pruning mechanism effectively improves the algorithm's structural adaptability and search performance while ensuring scheduling legitimacy.

[0058] Finally, if the algorithm's fitness has not been significantly improved or the maximum number of generations has been reached, the processor outputs the current optimal individual and converts the individual tree into a collaborative scheduling scheme for UGVs and UAVs; otherwise, the processor returns to the initial step of the pseudocode and continues to execute the next generation of evolution until the conditions mentioned above in this paragraph are met.

[0059] For example, the present application also provides simulation results: In this application, the mission planning scenario is set and In this scenario, set , indicating the fire point To simplify the simulation, The initial value is 0, each aircraft is assigned and successfully reaches the target Each step of the drone will make Increment by 1. Due to the energy of the drone in performing firefighting missions Therefore, by limiting the working time of each drone to Characterize the energy consumption of the UAV. The total working time of the UAV is limited to , the timed-out drone can neither move nor affect The scene size is , the time limit is set to In HMAS, UGV / UAV moves at a constant speed. Since UGV is usually slower than UAV, the UAV speed is set to , UGV cruising speed is Population size , mutation rate 0.2, crossover rate 0.8.

[0060] The process of obtaining the optimal individual using the GP-based hierarchical unmanned swarm task planning method is as follows: Figure 5 (a) is the UGV's driving path and the UAV's status at t=1000s, and (b) is the UGV's navigation to the first deployment point at t=2400s. , deploying two UAVs with targets 5 and 2 respectively. (c) is when t=3700, the UGV cruises to the second deployment point , and deploy the remaining five UAVs with targets 3, 1, 1, 4, and 0. (d) At t=4035, the UGV is still at the second deployment point , the remaining five UAVs with targets 3, 1, 1, 4, and 0 have already arrived at the fire point. In the above process, the individual tree corresponding to the optimal individual is: Root(Deploy([1.42,1.03], Deploy2(5,2)),Deploy([0.77,1.66], Deploy5(3,1,1,4,0)), Virtual, Virtual, Virtual, Virtual,Virtual, Virtual), and its fitness is 296.13. This indicates that the UGV will first navigate to the first deployment point , deploying two UAVs with targets 5 and 2 respectively; then continue cruising to the second deployment point , and launch the remaining 5 UAVs with targets 3, 1, 1, 4 and 0 respectively. Figure 5 It can be seen that under this scheduling scheme, all UAVs can reach the corresponding targets within the energy limit, and finally the HMAS successfully covers all fire points.

[0061] refer to Figure 6 Based on the above embodiments, the present application further provides a hierarchical unmanned swarm task planning device based on GP to solve the same technical problem as any of the above method embodiments. The hierarchical unmanned swarm task planning device 100 based on GP may include an individual number construction module 1001, a fitness calculation module 1002, and a task planning scheme generation module 1003. The individual number construction module 1001 is used to determine the five types of nodes of the unmanned swarm collaborative system jointly constructed by the fire-fighting unmanned vehicle and the multi-UAV group based on the domain knowledge of the hierarchical multi-agent system, and construct a hierarchical task individual tree with structured semantics based on the five types of nodes; The fitness calculation module 1002 is used to use a strongly typed genetic programming algorithm as a population evolution baseline, input the hierarchical task individual tree into the strongly typed genetic programming algorithm to generate an initial population, wherein the initial individuals of each initial population are the hierarchical task individual trees, and each hierarchical task individual tree corresponds to the deployment plan of each drone group and firefighting unmanned vehicle when performing a task. A predefined fitness function is used to evaluate the performance of each initial individual to obtain a fitness evaluation result; The task planning scheme generation module 1003 is used to determine whether the termination condition is met based on the fitness evaluation result. If the termination condition is met, the corresponding optimal hierarchical task individual tree is output according to the optimal fitness evaluation result; otherwise, a strongly typed genetic programming algorithm is used to select, crossover, and mutate each individual, and a pruning algorithm based on domain knowledge is used during the crossover and mutation operations to clean up redundant nodes in each individual, generate a descendant hierarchical task individual tree, use the fitness function to evaluate the performance of each group individual, obtain a new fitness evaluation result, and iteratively execute the previous execution step, and convert the optimal hierarchical task individual tree into a task planning scheme for the unmanned cluster collaborative system.

[0062] It is worth mentioning that all modules involved in this embodiment are logical modules. In actual applications, a logical unit can be a physical unit, a part of a physical unit, or a combination of multiple physical units. In addition, to highlight the innovation of this application, this embodiment does not introduce units that are not closely related to solving the technical problems proposed by this application. However, this does not mean that other units do not exist in this embodiment.

[0063] In another embodiment provided in the present application, a device is also provided, comprising a processor and a memory, wherein the memory stores at least one instruction, at least one program, a code set or an instruction set, and the at least one instruction, the at least one program, the code set or the instruction set is loaded and executed by the processor to implement the GP-based hierarchical unmanned cluster task planning method described in the embodiment of the present application.

[0064] In another embodiment provided in the present application, a computer-readable storage medium is also provided, in which at least one instruction, at least one program, a code set or an instruction set is stored. The at least one instruction, the at least one program, the code set or the instruction set is loaded and executed by a processor to implement the GP-based hierarchical unmanned cluster task planning method described in the embodiment of the present application.

[0065] In the above embodiments, all or part of the embodiments can be implemented using software, hardware, firmware, or any combination thereof. When implemented using software, all or part of the embodiments can be implemented in the form of a computer program product. The computer program product includes multiple computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via a wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) method. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or data center that integrates multiple available media. The available medium can be a magnetic medium (e.g., a floppy disk, hard disk, tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid-state drive (SSD)).

[0066] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of additional identical elements in the process, method, article, or device comprising the element.

[0067] Each embodiment in this specification is described in a related manner. Similar portions between the various embodiments can be referenced to each other. Each embodiment focuses on the differences between the other embodiments. In particular, the system embodiments are described briefly because they are generally similar to the method embodiments. For related portions, reference can be made to the description of the method embodiments.

[0068] The above description is only a preferred embodiment of the present application and is not intended to limit the scope of protection of the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application are included in the scope of protection of the present application.

Claims

1. A hierarchical unmanned swarm task planning method based on GP, characterized in that: The following steps are included: Based on the domain knowledge of hierarchical multi-agent systems, the five types of nodes in the unmanned swarm collaborative system jointly constructed by firefighting unmanned vehicles and multi-UAV swarms are identified, and a hierarchical task individual tree with structured semantics is constructed based on the five types of nodes. A strongly typed genetic programming algorithm is used as the population evolution baseline. The hierarchical task individual tree is input into the strongly typed genetic programming algorithm to generate the initial population. The initial individuals of each initial population are the hierarchical task individual trees, and each hierarchical task individual tree corresponds to the deployment plan of each drone group and firefighting unmanned vehicle when performing the task. A predefined fitness function is used to evaluate the performance of each initial individual to obtain the fitness evaluation result. Whether the termination condition is met is determined based on the fitness evaluation result. If the termination condition is met, the corresponding optimal hierarchical task individual tree is output based on the optimal fitness evaluation result; otherwise, a strongly typed genetic programming algorithm is used to perform selection, crossover, and mutation operations on each individual, and a pruning algorithm based on domain knowledge is used during the crossover and mutation operations to clean up redundant nodes in each individual, generate a descendant hierarchical task individual tree, and use the fitness function to evaluate the performance of each group individual to obtain a new fitness evaluation result, and iteratively execute the previous execution step, and convert the optimal hierarchical task individual tree into a task planning scheme for the unmanned swarm collaborative system.

2. A hierarchical unmanned swarm task planning method based on GP as claimed in claim 1, characterized in that: The five types of nodes include overall task nodes, lower-level subtask nodes, interactive subtask nodes, and upper-level subtask nodes, wherein the overall task nodes, upper-level subtask nodes, interactive subtask nodes, and lower-level subtask nodes have a parent-child inheritance relationship, and the upper-level subtask nodes include real subtask nodes and virtual subtask nodes; The method of using a pruning algorithm based on domain knowledge to clean up redundant nodes in each individual during the crossover and mutation operations includes: Get the initial individual after crossover; Based on the basic subtree mutation operator of the strongly typed genetic programming algorithm, a subtree that is not a root node is selected from the initial individuals after each crossover as a randomly generated new subtree to obtain the mutated individuals; Calculate the number of lower-level agents currently participating in firefighting based on the lower-level subtask nodes of the mutated individuals, where the lower-level subtask nodes are used to determine the fire point corresponding to each lower-level agent, where the lower-level agents are drones; Determine whether the number of lower-level agents is greater than the number of fire points. If so, randomly generate an intermediate number within the range of the number of fire points and determine whether the number of lower-level agents is greater than the intermediate number. If so, randomly remove lower-level subtask nodes until the number of lower-level agents is less than or equal to the number of fire points. Among them, in the process of pruning lower-level subtask nodes, if it is determined that all lower-level subtask nodes under the interactive subtask node are pruned, the entire subtree with the real subtask node as the root will be pruned, and the virtual subtask node will be used to replace the real subtask node to connect to the overall task node.

3. A GP-based hierarchical unmanned swarm task planning method as claimed in claim 2, characterized in that: The five types of nodes also include parameter nodes; The overall task node is connected to each upper-level subtask node, and is used to determine the overall fire extinguishing task of the unmanned cluster collaborative system when extinguishing a fire according to the upper-level subtask nodes, and the execution order of the upper-level subtask nodes is determined according to the connection order of the overall task node; The real subtask node of the upper-level subtask node includes an interaction subtask node and a parameter node. The parameter node is used to store interaction parameters inside and outside the unmanned cluster collaborative system, wherein the interaction parameters include path coordinate parameters and energy consumption parameters. The interaction subtask node represents the fire point corresponding to each drone, as well as the collaborative path between the drone swarm and the unmanned fire truck. The virtual subtask node of the upper subtask node is used as a placeholder when the number of real subtasks is less than the number of subnodes required by the overall task node.

4. A GP-based hierarchical unmanned swarm task planning method as claimed in claim 3, characterized in that: The hierarchical task variable tree is a four-layer structure, which is a root layer, an upper subtask layer, an interactive subtask layer, and a lower subtask layer from top to bottom. Among them, the overall task node belongs to the root layer; The upper subtask nodes belong to the upper subtask layer; Interaction subtask nodes and parameter nodes belong to the interaction subtask layer; The lower subtask node belongs to the lower subtask layer.

5. The GP-based hierarchical unmanned swarm task planning method according to claim 2, characterized in that: Before obtaining the initial individual after the crossover, the method further includes: Randomly select multiple initial individuals from the initial population based on the tournament selection operator; Based on the conventional subtree crossover operator of the strongly typed genetic programming algorithm, crossover points of the same type are randomly selected from each initial individual, and subtrees with the crossover points as root nodes are exchanged between the initial individuals to obtain individuals after each crossover.

6. The GP-based hierarchical unmanned swarm task planning method according to claim 1, characterized in that: Before generating the initial population, the method further includes: Node sets and node connection rules are predefined, where the node sets and node connection rules are used to generate constraint conditions so that the hierarchical task variable tree structure is legal and complies with task constraints.

7. The GP-based hierarchical unmanned swarm task planning method according to claim 1, characterized in that: The expression of the fitness function is: in, , , , Indicates the fire point The percentage of firefighting tasks completed, Indicates the number of drones actually deployed, Indicates the number of drones to be deployed, represents the total mission time of each UAV, Indicates the maximum total working time of all drones, Indicates the degree of task completion, Indicates resource utilization efficiency, Indicates time consumption efficiency.

8. A hierarchical unmanned swarm task planning method based on GP, characterized in that: include: The individual number construction module is used to determine the five types of nodes in the unmanned swarm collaborative system jointly constructed by firefighting unmanned vehicles and multiple drones based on the domain knowledge of hierarchical multi-agent systems, and to construct a hierarchical task individual tree with structured semantics based on the five types of nodes; The fitness calculation module is used to use a strongly typed genetic programming algorithm as the population evolution baseline, input the hierarchical task individual tree into the strongly typed genetic programming algorithm to generate an initial population, wherein the initial individuals of each initial population are the task individual trees of each level, and each level of task individual tree corresponds to the deployment plan of each drone group and firefighting unmanned vehicle when performing a task. A predefined fitness function is used to evaluate the performance of each initial individual to obtain a fitness evaluation result; The task planning scheme generation module is used to determine whether the termination condition is met based on the fitness evaluation results. If the termination condition is met, the corresponding optimal hierarchical task individual tree is output according to the optimal fitness evaluation result; otherwise, the strongly typed genetic programming algorithm is used to select, crossover, and mutate each individual, and the pruning algorithm based on domain knowledge is used to clean up the redundant nodes in each individual during the crossover and mutation operations to generate a descendant hierarchical task individual tree. The fitness function is used to evaluate the performance of each group individual to obtain a new fitness evaluation result, and the previous execution step is iteratively executed, and the optimal hierarchical task individual tree is converted into a task planning scheme for the unmanned cluster collaborative system.

9. An electronic device, characterized in that: The electronic device includes a processor and a memory, wherein the memory stores at least one instruction, at least one program, a code set or an instruction set, and the at least one instruction, the at least one program, the code set or the instruction set are loaded and executed by the processor to implement the GP-based hierarchical unmanned cluster task planning method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that The storage medium stores at least one instruction, at least one program, a code set or an instruction set, and the at least one instruction, the at least one program, the code set or the instruction set are loaded and executed by the processor to implement the GP-based hierarchical unmanned cluster task planning method as described in any one of claims 1-7.

Images