Public safety event emergency resource allocation method based on multi-agent collaborative optimization

Abstract

The invention discloses a public security event emergency resource allocation method based on multi-agent collaborative optimization. The method comprises the following steps: S1, forming a three-dimensional state tensor field; s2, initializing an event agent, a resource agent and a scheduling agent based on the three-dimensional state tensor field, constructing a resource scheduling solution space, configuring an evolution hybrid perception model for each agent, and outputting a scheduling potential map; s3, path search based on target driving is carried out in the resource scheduling solution space, honey source position selection is carried out based on the scheduling potential map through a bee-bird search algorithm, and a resource matching path is generated; s4, establishing a bidirectional coupling mechanism between the hummingbird search algorithm and the evolution hybrid perception model, and realizing dynamic collaborative optimization; and S5, outputting a final resource scheduling scheme and executing the final resource scheduling scheme in real time by each agent to form a scheduling decision closed loop of the emergency response of the public security event. According to the method, the multi-agent evolution hybrid perception model and the hummingbird search algorithm are fused, and public security event resource scheduling optimization is realized.

Description

Technical Field

[0001] The present invention relates to the technical fields of emergency management and artificial intelligence, and particularly to an emergency resource allocation method for public safety incidents based on multi-agent collaborative optimization. Background Art

[0002] Public safety incidents have significant characteristics such as strong suddenness, urgent response time, and complex resource coordination, involving the rapid linkage of multiple emergency systems such as policing, medical care, fire protection, and transportation. With the increase in urban complexity, public safety incidents show a trend of multi-point and multi-category concurrency in space and time. A single incident may quickly trigger regional resource conflicts, further resulting in a decline in response efficiency and the spread of secondary risks. Therefore, how to achieve rapid, accurate, and efficient emergency resource allocation in scenarios of multiple events, multiple resources, and multiple constraints has become a key problem that urgently needs to be solved in the current public safety technology field.

[0003] Currently, traditional emergency resource scheduling methods mostly take rule engines, static models, or single algorithms as the core, relying on scheduling schemes formulated based on preset response rules or expert experience. Such methods have certain practicability in the case of small-scale and single events, but it is difficult to adapt to complex scenarios such as dynamic evolution, event coupling, and resource competition. On the one hand, rule-based methods cannot cope with real-time changing environmental data and lack a dynamic feedback mechanism, resulting in rigid scheduling strategies and low response timeliness; on the other hand, static models do not consider the influence diffusion relationship between events and ignore the dynamic changes in the resource usage status, easily causing resource duplication or blind scheduling.

[0004] At the level of intelligent algorithms, some studies have attempted to introduce methods such as heuristic search, genetic algorithms, and ant colony algorithms to optimize and solve the resource scheduling problem. These algorithms usually simplify the event response process into a bipartite graph matching model of resources - tasks, which improves the resource utilization efficiency to a certain extent. However, such methods generally have the following deficiencies: First, the algorithm target is single, usually only optimizing the response time or scheduling distance, and unable to take into account multiple factors such as constraint conditions, resource status, and event levels at the same time; Second, the algorithm process runs independently and is not coupled with the event perception model and the state feedback system, unable to achieve real-time perception of external environmental changes and policy adjustment; Third, the generation of the scheduling optimization path lacks a global perspective and is often trapped in local optimality, especially showing problems such as poor strategy stability and strong system volatility under the condition of multi-event conflicts.

[0005] In recent years, with the development of the theory of multi-agent systems, the academic and industrial communities have gradually explored the application of multi-agent mechanisms in the scenario of emergency resource scheduling. This approach believes that various response entities in the emergency scenario (such as rescue vehicles, medical teams, event nodes, etc.) can be modeled as intelligent agent units with independent perception, decision-making, and collaboration capabilities, and the autonomous collaborative scheduling of resources is achieved through the state exchange and behavior game among the agents. Such methods have advantages such as being distributed, decentralized, and scalable, but multiple limitations have also emerged in actual implementation. First, the state information sharing mechanism among multi-agents often lacks the support of structured environment modeling, resulting in information fragmentation and chaotic interaction during the collaboration process. Second, the perception model designs of the agents are not unified, lacking a unified data structure input and evolution mechanism, making it difficult to form a joint force for collaborative scheduling behavior. Third, multi-agent systems often lack an optimization engine oriented to scheduling goals, resulting in the difficulty of aggregating agent behaviors into a system-level optimal solution.

[0006] In addition, in current research, the perception model and the scheduling optimization algorithm are often designed separately, that is, the model is responsible for identifying the external state, and the optimization algorithm runs independently to solve the resource scheduling path. Since no feedback path is established between the two, the scheduling optimization process cannot dynamically adjust the target area or resource direction according to the recognition ability of the model, and at the same time, the focus of model perception cannot self-evolve according to the historical results of algorithm search. This fragmented structure severely restricts the overall collaborative efficiency and response adaptability of the system.

[0007] As a new type of heuristic optimization method, the hummingbird search algorithm has advantages such as adjustable search path direction, strong local perturbation ability, and outstanding ability to jump out of local optima, and shows good global optimization ability in some complex search problems. However, at present, the hummingbird search algorithm has not been systematically embedded in the public safety resource scheduling scenario, and has not been deeply combined with the intelligent agent model with self-learning ability, making it difficult to exert its optimization potential in the high-dimensional dynamic solution space.

[0008] There is also a lack of a structural solution for bidirectionally coupling the hummingbird search algorithm with the multi-agent perception mechanism in the existing technology. On the one hand, the search path cannot act on the model in reverse, resulting in information fragmentation between perception and behavior; on the other hand, the perception result cannot accurately control the starting point selection and path direction generation of the search path, and the search behavior lacks clear guidance in space, making it difficult to improve the resource scheduling efficiency. Especially in the actual emergency response scenario involving multiple emergencies, diverse resource types, and complex constraints, the existing technology cannot provide a scheduling strategy generation framework that takes into account interpretability, real-time performance, and scheduling globalness.

[0009] Therefore, how to provide an emergency resource allocation method for public safety events based on multi-agent collaborative optimization is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0010] An object of the present invention is to propose an emergency resource allocation method for public security incidents based on multi-agent collaborative optimization. The present invention integrates a multi-agent evolutionary hybrid perception model and a hummingbird search algorithm, constructs a three-dimensional state tensor field of events, resources, and constraints, dynamically generates an efficient resource scheduling path, realizes intelligent response and optimal allocation in public security incidents, and has the advantages of fast response, reasonable path, strong system coordination ability, and high ability to adapt to complex environments.

[0011] The emergency resource allocation method for public security incidents based on multi-agent collaborative optimization according to an embodiment of the present invention includes the following steps:

[0012] S1. Construct an event set, a resource set, and a constraint set to form a three-dimensional state tensor field;

[0013] S2. Initialize an event agent, a resource agent, and a scheduling agent based on the three-dimensional state tensor field, construct a resource scheduling solution space, and configure an evolutionary hybrid perception model for each agent to output a scheduling potential map;

[0014] S3. Execute the hummingbird search algorithm to perform target-driven path search in the resource scheduling solution space. The hummingbird search algorithm selects the nectar source position based on the scheduling potential map and generates a resource matching path;

[0015] S4. Establish a bidirectional coupling mechanism between the hummingbird search algorithm and the evolutionary hybrid perception model, where the resource matching path of the hummingbird search algorithm is fed back to the evolutionary hybrid perception model of each agent to adjust the perception accuracy and the strategy deviation direction to achieve dynamic collaborative optimization;

[0016] S5. Output the final resource scheduling plan and execute it in real time by each agent to form a scheduling decision closed loop for emergency response to public security incidents.

[0017] Optionally, the three-dimensional state tensor field is used to describe the multi-dimensional dynamic correlation relationship between the event set, the resource set, and the constraint set, and the construction method is as follows:

[0018]

[0019] where T i,j,k represents the state score of resource r k responding to event e j under constraint condition c i , c k represents the kth constraint condition, r j represents the jth type of resource, e i represents the ith event, and ∈ represents a smoothing factor;

[0020] P(e i ,r j ) represents the task priority weight of event e i for resource r j , and D(r j ,e i ) represents the path cost from resource r j to event e i , including response time and traffic accessibility. The larger the value of P(e i ,r j ), the more difficult the response;

[0021] A(c k ) represents the constraint activation factor, which takes the value of 1 when the constraint condition c k is satisfied, and 0 otherwise;

[0022] L(r j ) represents the resource load factor, that is, the current task occupancy ratio of resource r j , ranging from [0,1]. The larger the value of L(r j ), the higher the load;

[0023] The three-dimensional state tensor field is updated by rolling over time, and is used to reflect the joint scheduling feasibility of event states, resource states, and environmental constraints in real time.

[0024] Optionally, the S2 specifically includes:

[0025] S21. Construct event agents according to each event in the event set. The attributes of the event agents include the spatial location, response time limit, risk level, and diffusion trend prediction parameters of the events, and are used to extract the associated resource subsets and constraint condition slices in the three-dimensional state tensor field;

[0026] S22. Construct resource agents according to each resource in the resource set. The attributes of the resource agents include resource type, current state, physical location, task load, and response ability, and are used to perceive the current potential event response tasks based on the three-dimensional state tensor field;

[0027] S23. The scheduling agent serves as a coordination unit, and is used to construct a mapping matrix between events, resources, and constraints, form a scheduling candidate set and a resource scheduling solution space structure, and is responsible for collecting perception information and organizing global feedback;

[0028] S24. Configure an evolutionary hybrid perception model for each agent. The evolutionary hybrid perception model includes a state perception embedding module, an evolutionary control module, and a scheduling tendency module;

[0029] The state perception embedding module is used to input tensor slice data and generate z through a graph attention encoder A, where z A represents the perceptual embedding representation of the A-th agent in the current state;

[0030] The evolution control module maintains the behavior trajectory, historical feedback results, and current state weight of each agent, and executes the state update function:

[0031]

[0032] where, ω A (t + 1) represents the behavior participation degree of the A-th agent at time t + 1, ω A (t) represents the behavior participation degree of the A-th agent at time t, tanh represents the hyperbolic tangent function, α, β, and γ represent adjustment parameters, and ΔR A (t) represents the resource response benefit improvement value after the A-th agent executes the policy, represents the policy gradient, and ∥·∥ represents the norm;

[0033] The scheduling tendency module outputs the scheduling response tendency value based on z A and ω A (t + 1):

[0034] η A = σ(ω A (t + 1)·W1·tanh(W2·z A + b));

[0035] where, η A represents the scheduling response tendency value of the A-th agent, η A ∈[0,1], σ represents the activation function, W1 and W2 represent the parameter matrices, and b represents the offset parameter;

[0036] S25. Fuse the scheduling response tendency values generated by all event agents and resource agents to construct a scheduling potential map:

[0037]

[0038] where, Ψ(x,y) represents the scheduling potential map, (x,y) represents the location coordinates of the target area, exp represents the exponential function, η ij represents the joint scheduling response score of the i-th event agent and the j-th resource agent, δ represents the spatial diffusion adjustment coefficient, (x i ,y i ) represents the spatial location coordinates of the event, (x j ,y j ) represents the spatial location coordinates of the resource, n represents the total number of event agents, and m represents the total number of resource agents;

[0039] S26. Use the scheduling potential map as the basic input map for nectar source selection and path search in the hummingbird search algorithm, and use it to dynamically guide the generation of resource scheduling paths and the adjustment of optimization directions.

[0040] Optionally, the S3 specifically includes:

[0041] S31. Using the scheduling potential map as the input, randomly select spatial positions with potential values greater than the set threshold as the initial search nodes. The initial search nodes and reachable resources and events form the path search starting points, and all starting points form the initial position set of the hummingbird search algorithm;

[0042] S32. Based on the initial position set, construct a hummingbird search individual population. Each search individual corresponds to a set of resource scheduling path combinations. The resource scheduling path combinations include: starting point position, target event position, resource number, sequence of intermediate nodes in the resource movement path, estimated response time, and path passing status. All fields are stored in a structured manner;

[0043] S33. According to the spatial coordinate relationship of each node in the scheduling potential map, generate the initial search direction of the path combination. Each search individual uses the connection direction between the starting point and the target as the initial path direction, and gradually inserts reachable nodes along the initial path direction to generate the first-round candidate path plan;

[0044] S34. During the search process, perform two types of path expansion methods for each search individual:

[0045] Local path perturbation method: Insert an adjacent node that is reachable in the graph structure and has a normal passing status in any paragraph of the current path plan to expand the path selection space and optimize the timeliness of resource arrival at the event;

[0046] Remote path jump method: Forcefully insert a discontinuous node between the front and rear segments of the path, and calculate whether the connection path is within the allowable range of passing constraints. If allowed, retain this new path; otherwise, roll back to the previous path state;

[0047] S35. For search individuals whose path evaluation scores have not improved for three consecutive rounds, clear the current path record, and re-select a new path search starting point from the unselected high-potential nodes with potential values greater than the set threshold, and re-generate the starting point-target combination to replace the original search direction and path content;

[0048] S36. After completing all preset search rounds, select several groups of paths with complete paths, normal passing status, the shortest estimated response time, and the highest resource-event matching degree from all search individuals as the final resource matching paths for output.

[0049] Optionally, the S4 specifically includes:

[0050] S41. After each round of path search by the hummingbird search algorithm, based on the resource matching path, extract the path results corresponding to each search individual, including: the resource numbers used, the resource start positions, the scheduling target positions, the sequence of nodes passed by the path, the total path length, the evaluation value of the passage state, and the estimated arrival time, and encapsulate them into feedback data packets;

[0051] S42. Distribute the feedback data packets to the resource agents, scheduling agents, and corresponding event agents involved in the path. The feedback data packets received by each agent correspond one-to-one with the actual participation status in the path, ensuring that the path information feedback has an agent identity matching relationship;

[0052] S43. Establish a bidirectional coupling mechanism between the hummingbird search algorithm and the evolutionary hybrid perception model. Each agent that receives the feedback data packet performs two update tasks on the evolutionary hybrid perception model according to its own role in the path and the result status of the executed task:

[0053] Update the input fields of the state perception embedding module, including the resource passage state, the path scheduling success rate, and the adjacent event density distribution;

[0054] Update the offset parameter of the scheduling tendency module to adjust the priority direction of the resource pointing to the event in the scheduling response strategy;

[0055] S44. In the next round of scheduling tasks, the updated state perception embedding module will use the new path feedback as the historical input feature. When generating the scheduling response tendency value, the updated scheduling tendency module will automatically increase the scheduling response tendency value in the position area where the success rate in the historical scheduling is greater than 90%, and at the same time decrease the scheduling response tendency value in the position area where the failure rate is greater than 50%;

[0056] S45. After receiving the input fields of the updated state perception embedding module and the offset parameter of the scheduling tendency module, the hummingbird search algorithm re-initializes the path start point selection logic of the search individual, realizes the dependence of the search behavior on perception, and realizes the real-time coupling and dynamic collaborative optimization between the evolutionary hybrid perception model and the hummingbird search algorithm.

[0057] The beneficial effects of the present invention are:

[0058] First, the present invention adopts a multi-agent system composed of collaborative event agents, resource agents, and scheduling agents, effectively overcoming the problems of task concentration and rigid response paths in traditional scheduling systems. Each type of agent realizes independent perception and behavior decision-making based on its attribute fields and tensor slices, and achieves a closed-loop chain of state perception - behavior response - policy update through an evolutionary hybrid perception model, significantly enhancing the system's autonomous adaptability and task collaboration. Especially in complex scenarios with multiple concurrent events and strong resource heterogeneity, this model can support individual agents to autonomously adjust their strategies according to local environmental changes, improving the overall response efficiency.

[0059] Secondly, the present invention introduces the hummingbird search algorithm in the scheduling optimization layer, using the scheduling potential map as the guiding basis for the search behavior, making the search path have spatial selectivity and direction convergence. Compared with the traditional heuristic algorithm's unguided random search in high-dimensional space, the hummingbird search algorithm constructs the path starting point, path direction, and node jump strategy based on the scheduling tendency value, significantly improving both the search efficiency and the quality of the results. The path perturbation and path jump mechanisms in the hummingbird search further enhance the ability of the search to jump out of local optima, showing stronger convergence ability and search breadth in the generation of scheduling paths for high-density areas of emergencies.

[0060] In addition, the present invention constructs a bidirectional coupling mechanism between the hummingbird search algorithm and the evolutionary hybrid perception model, realizing information sharing and policy linkage between the model and the algorithm, and solving the problem in the prior art that the model and the optimization algorithm run separately and information cannot be synchronized in real time. In the present invention, after each round of path generation, the search algorithm takes the path feedback as the evolutionary input of the perception model to update the state evaluation and policy direction of the agent; conversely, the agent model can also adjust the search starting point selection logic and path generation rules according to its own evolutionary results, forming a policy co-evolution system centered on feedback. This deep coupling structure enables the system to have the ability of double-layer iterative optimization of paths and perception, significantly enhancing the sensitivity and stability of the scheduling strategy to dynamic environmental changes.

[0061] Finally, the present invention realizes a closed-loop of scheduling optimization of perception - search - feedback - re-perception. The system can automatically generate the optimal resource scheduling path in complex scenarios without relying on manual experience or predefined rules, realizing the intelligence, structuring, and adaptive evolution of resource scheduling, and having good practicality, scalability, and engineering deployment prospects. Description of the Drawings

[0062] The drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation to the present invention. In the drawings:

[0063] Figure 1This is the overall flowchart of the emergency resource allocation method for public security incidents based on multi-agent collaborative optimization proposed by the present invention. Detailed implementation manners

[0064] Now, the present invention will be further described in detail with reference to the accompanying drawings. These drawings are all simplified schematic diagrams, only illustrating the basic structure of the present invention in a schematic manner, so they only show the components related to the present invention.

[0065] Reference Figure 1 , the emergency resource allocation method for public security incidents based on multi-agent collaborative optimization includes the following steps:

[0066] S1. Construct an event set, a resource set, and a constraint set to form a three-dimensional state tensor field;

[0067] S2. Initialize an event agent, a resource agent, and a scheduling agent based on the three-dimensional state tensor field, construct a resource scheduling solution space, and configure an evolutionary hybrid perception model for each agent to output a scheduling potential map;

[0068] S3. Execute the hummingbird search algorithm to perform a target-driven path search in the resource scheduling solution space. The hummingbird search algorithm selects the nectar source position based on the scheduling potential map and generates a resource matching path;

[0069] S4. Establish a bidirectional coupling mechanism between the hummingbird search algorithm and the evolutionary hybrid perception model. Among them, the resource matching path of the hummingbird search algorithm is fed back to the evolutionary hybrid perception model of each agent to adjust the perception accuracy and the strategy deviation direction to achieve dynamic collaborative optimization;

[0070] S5. Output the final resource scheduling plan and execute it in real time by each agent to form a scheduling decision closed-loop for public security incident emergency response.

[0071] By proposing an emergency resource allocation method for public security incidents based on multi-agent collaborative optimization, the present invention realizes the unified modeling of event, resource, and constraint elements, the global optimization of the scheduling path, and the closed-loop response mechanism of the resource scheduling behavior. This method breaks the limitations of traditional rule-based and centralized emergency scheduling models, and for the first time integrates an evolutionary hybrid perception model and a hummingbird search algorithm to construct an information flow closed-loop structure among the model, the algorithm, and the agent. Its advantages are as follows: in scenarios with frequent emergencies, limited resources, and complex constraints, it can dynamically match the optimal resource path and adapt to external environmental changes through real-time evolution of the model, significantly improving the scheduling reaction speed and accuracy. This method has high flexibility and scalability, can adapt to various types of events, multi-source resources, and multi-level decision-making requirements, is applicable to complex scenarios such as urban emergency management, disaster response, and public security command systems, and has good engineering feasibility and practical application value.

[0072] In this embodiment, the three-dimensional state tensor field is used to describe the multi-dimensional dynamic association relationship among the event set, the resource set, and the constraint set, and the construction method is as follows:

[0073]

[0074] where T i,j,k represents the state score of resource r k responding to event e j under the constraint condition c i , c k represents the k-th constraint condition, r j represents the j-th type of resource, e i represents the i-th event, and ∈ represents the smoothing factor;

[0075] P(e i , r j ) represents the task priority weight of event e i on resource r j , D(r j , e i ) represents the path cost from resource r j to event e i , including response time and traffic accessibility. The larger the value of P(e i , r j ), the more difficult the response;

[0076] A(c k ) represents the constraint activation factor, which takes the value of 1 when the constraint condition c k is satisfied, and 0 otherwise;

[0077] L(r j ) represents the resource load factor, that is, the current task occupancy ratio of resource r j , and the range is [0, 1]. The larger the value of L(r j ), the higher the load;

[0078] The three-dimensional state tensor field is updated in a rolling manner over time to reflect the joint scheduling feasibility of the event state, the resource state, and the environmental constraints in real time.

[0079] Step S1 models the multi-dimensional dynamic relationships among the event set, resource set, and constraint set by introducing a three-dimensional state tensor field, constructing a structured perception data body with the ability of real-time rolling update. This tensor field provides accurate and quantifiable basic data support, enabling the agent to not only comprehensively obtain the interaction states among events-resources-constraints during the perception stage but also dynamically adjust strategies according to spatio-temporal changes. Compared with traditional two-dimensional table or graph structure representation methods, this tensor field can fuse multi-type, multi-scale, and multi-state data, enabling the scheduling system to more accurately identify potential adjustable resources, judge the urgency of event responses, and the scope of constraint impacts, greatly enhancing the system's sensitivity to sudden states and scheduling adaptability. At the same time, this structure provides a unified data interface for subsequent model perception, path search, and feedback update, serving as the basic support for the efficient operation of the entire system.

[0080] In this embodiment, S2 specifically includes:

[0081] S21. Construct event agents according to each event in the event set. The attributes of the event agents include the spatial location, response time limit, risk level, and diffusion trend prediction parameters of the events, and are used to extract the associated resource subsets and constraint condition slices in the three-dimensional state tensor field;

[0082] S22. Construct resource agents according to each resource in the resource set. The attributes of the resource agents include resource type, current state, physical location, task load, and response ability, and are used to perceive current potential event response tasks based on the three-dimensional state tensor field;

[0083] S23. The scheduling agent, as a coordination unit, is used to construct a mapping matrix among events-resources-constraints, form a scheduling candidate set and a resource scheduling solution space structure, and is responsible for collecting perception information and organizing global feedback;

[0084] S24. Configure an evolutionary hybrid perception model for each agent. The evolutionary hybrid perception model includes a state perception embedding module, an evolutionary control module, and a scheduling tendency module;

[0085] The state perception embedding module is used to input tensor slice data and generate z A , where z A represents the perception embedding representation of the A-th agent in the current state;

[0086] The evolutionary control module maintains the behavior trajectory, historical feedback results, and current state weights of each agent, and executes the state update function:

[0087]

[0088] where ωA (t + 1) represents the behavior participation degree of the A-th agent at time t + 1, ω A (t) represents the behavior participation degree of the A-th agent at time t, tanh represents the hyperbolic tangent function, α, β, and γ represent adjustment parameters, ΔR A (t) represents the value of the improvement in resource response benefit after the A-th agent executes the policy, represents the policy gradient, ∥·∥ represents the norm;

[0089] The said scheduling tendency module is based on z A and ω A (t + 1) to output the scheduling response tendency value:

[0090] η A = σ(ω A (t + 1)·W1·tanh(W2·z A + b));

[0091] where, η A represents the scheduling response tendency value of the A-th agent, η A ∈[0, 1], σ represents the activation function, W1 and W2 represent the parameter matrices, b represents the offset parameter;

[0092] S25. Integrate the scheduling response tendency values generated by all event agents and resource agents to construct a scheduling potential map:

[0093]

[0094] where, Ψ(x, y) represents the scheduling potential map, (x, y) represents the location coordinates of the target area, exp represents the exponential function, η ij represents the joint scheduling response score of the i-th event agent and the j-th resource agent, δ represents the spatial diffusion adjustment coefficient, (x i , y i ) represents the spatial location coordinates of the event, (x j , y j ) represents the spatial location coordinates of the resource, n represents the total number of event agents, m represents the total number of resource agents;

[0095] S26. Use the scheduling potential map as the basic input map for nectar source selection and path search in the hummingbird search algorithm to dynamically guide the generation of the resource scheduling path and the adjustment of the optimization direction.

[0096] Step S2 introduces a ternary cooperation mechanism of event agents, resource agents, and scheduling agents during the initialization of the agent system. By configuring the evolutionary hybrid perception model, it realizes the dynamic perception and behavior evolution of scheduling behaviors, enhancing the distributed perception ability and local decision-making quality of the entire scheduling system. By binding agents to tensor field slices, each agent can independently perceive its associated state data and self-adjust based on historical feedback, thus effectively avoiding the single-point bottleneck and information delay problems in global scheduling. The state perception embedding module, evolutionary control module, and scheduling tendency module in the evolutionary hybrid perception model work together, making the scheduling decision time-dependent, behaviorally memory-based, and environmentally adaptable. This structure strengthens the individual response ability of agents and the system-level synergy effect, providing high-quality input and scheduling direction guidance for subsequent path search and feedback optimization, significantly improving the scheduling accuracy and system stability.

[0097] In this embodiment, the specific content of S3 is as follows:

[0098] S31: Using the scheduling potential map as input, randomly select spatial positions with potential values greater than the set threshold as the initial search nodes. The initial search nodes and reachable resources and events form the path search starting points, and all starting points form the initial position set of the hummingbird search algorithm.

[0099] S32: Based on the initial position set, construct a hummingbird search individual population. Each search individual corresponds to a set of resource scheduling path combinations, and the resource scheduling path combinations include: starting point position, target event position, resource number, intermediate node sequence in the resource movement path, estimated response time, and path passing status. All fields are stored in a structured manner.

[0100] S33: According to the spatial coordinate relationship of each node in the scheduling potential map, generate the initial search direction of the path combination. Each search individual uses the connection direction between the starting point and the target as the initial path direction, and gradually inserts reachable nodes along the initial path direction to generate the first-round candidate path plan.

[0101] S34: During the search process, perform two types of path expansion methods on each search individual:

[0102] Local path perturbation method: Insert an adjacent node that is reachable in the graph structure and has a normal passing status in any paragraph of the current path plan to expand the path selection space and optimize the timeliness of resource arrival at the event.

[0103] Remote path jump method: Forcefully insert a discontinuous node between the front and back segments of the path, and calculate whether the connected path is within the allowable range of passing constraints. If allowed, retain this new path; otherwise, roll back to the previous path state.

[0104] S35. For search individuals whose path evaluation scores have not improved for three consecutive rounds, clear the current path record, and then re-select a new path search starting point from the unselected high-potential nodes with potential values greater than the set threshold. Regenerate the start-goal combination to replace the original search direction and path content.

[0105] S36. After completing all preset search rounds, screen out several groups of paths with complete paths, normal passage status, the shortest expected response time, and the highest resource-event matching degree from all search individuals as the final resource matching paths for output.

[0106] Step S3 clarifies the goal-driven strategy, path combination mechanism, perturbation expansion method, and local jump logic in the resource scheduling path search by constructing the whole-process behavior logic of the hummingbird search algorithm. It guides search individuals to carry out path combination in the high-response area through the scheduling potential map, solving the problems of large randomness and insufficient directionality in path initialization in traditional search algorithms. By introducing two mechanisms of local path perturbation and remote path jump, the search process takes into account both fine-tuning and large-scale exploration, thus enhancing the comprehensiveness and diversity of the solution space search. When the path gets stuck, the reset and re-selection strategy is adopted to effectively avoid the local optimal trap. The finally output path plan has characteristics such as short response time, high passage accessibility, and strong resource-event matching degree, which can be directly converted into agent scheduling instructions to improve the executability of resource dispatch and the task completion rate. Overall, step S3 improves the efficiency and quality of path search and is a key component to achieve high-quality scheduling output.

[0107] In this embodiment, S4 specifically includes:

[0108] S41. After each round of path search of the hummingbird search algorithm, extract the path results corresponding to each search individual based on the resource matching path, including: the resource number used, the resource starting position, the scheduling target position, the sequence of nodes passed by the path, the total path length, the passage status evaluation value, and the expected arrival time, and encapsulate them into a feedback data packet.

[0109] S42. Distribute the feedback data packet to the resource agents, scheduling agents, and corresponding event agents involved in the path. The feedback data packet received by each agent corresponds one-to-one with its actual participation status in the path, ensuring that the path information feedback has an agent identity matching relationship.

[0110] S43. Establish a two-way coupling mechanism between the hummingbird search algorithm and the evolutionary hybrid perception model. Each agent that receives the feedback data packet performs two update tasks on the evolutionary hybrid perception model according to its own role in the path and the result status of the executed task:

[0111] Update the input fields of the state awareness embedding module, including resource passage status, path scheduling success rate, and adjacent event density distribution;

[0112] Update the offset parameter of the scheduling tendency module for adjusting the priority direction of resource - directed events in the scheduling response strategy;

[0113] S44. In the next - round scheduling task, the updated state awareness embedding module will use the new path feedback as the historical input feature. When generating the scheduling response tendency value, the updated scheduling tendency module will automatically increase the scheduling response tendency value in the position area where the success rate is greater than 90% in the historical scheduling, and at the same time decrease the scheduling response tendency value in the position area where the failure rate is greater than 50%;

[0114] S45. After receiving the input fields of the updated state awareness embedding module and the offset parameter of the scheduling tendency module, the hummingbird search algorithm re - initializes the path - starting point selection logic of the search individual, realizes the dependence of the search behavior on perception, and realizes the real - time coupling and dynamic co - optimization between the evolutionary hybrid perception model and the hummingbird search algorithm.

[0115] In step S4, by constructing a two - way coupling mechanism between the hummingbird search algorithm and the evolutionary hybrid perception model, the design pattern of the traditional model and optimization algorithm with the separation of perception and execution is broken, and the real - time linkage between the scheduling behavior and the model state is realized for the first time. After each round of search, the hummingbird search algorithm outputs a feedback data packet containing specific structured information such as path nodes, response time, and passage status, which is fed back to the intelligent agents corresponding to the participating paths; the intelligent agents update their own perception input fields and policy offset parameters accordingly, enabling the model to have an adaptive adjustment ability when generating scheduling tendencies in the next round. Conversely, the adjustment results of the model will affect the starting point and direction generation logic of the search path in real time, so that the search process is no longer a blind full - map traversal, but a dynamic behavior optimization process with historical perception constraints and feedback learning ability, effectively improving the scheduling accuracy, convergence speed, and scheduling strategy stability of the system, and significantly enhancing the robustness and self - recovery ability of the system in continuous emergency scenarios.

[0116] Embodiment 1:

[0117] To verify the feasibility of the present invention in implementation, the present invention is applied to a typical urban waterlogging emergency response case for simulation and verification. In this scenario, the central and southern regions of the city encounter continuous heavy rainfall in a short period of time, resulting in an overloaded drainage system, and waterlogging accumulates in some blocks, affecting key infrastructure such as multiple communities, subway entrances, tunnels, and substations. Multiple events occur simultaneously, and it is urgent to dispatch various emergency resources to carry out drainage, emergency repair, evacuation, and power restoration work.

[0118] In the short time after the event occurred, the system accessed the data of the urban perception platform and automatically identified 21 key dangerous situation points, including sudden problems of various types such as water ingress in the underground garage of residential buildings, waterlogging in transportation hubs, and collapse of perimeter walls around schools. Through the multi-agent collaborative optimization method proposed by the present invention, an event set, a resource set, and a scheduling constraint set are first constructed to form a three-dimensional state tensor field, and the risk level, resource demand type, and time limit requirements of each event point are uniformly encoded and modeled. The system automatically extracts the involved emergency resource units, including 63 groups of resources such as drainage vehicle groups, power repair vehicle groups, personnel emergency teams, drone inspection teams, and emergency power supply vehicles, and at the same time imports scheduling constraint conditions such as current traffic accessibility, personnel deployment restrictions, and task sequence priorities.

[0119] Next, according to the content of the tensor field, the system automatically initializes 21 event agents, 63 resource agents, and 1 global scheduling agent, and configures an evolutionary hybrid perception model for each agent. Each event agent generates a local perception result according to factors such as the risk diffusion trend in the area where it is located and the urgency of resource demand, and each resource agent predicts its available response ability according to its own type, current state, access path, etc. The perception results of all agents are fused to generate a real-time scheduling potential map, which provides a spatial scheduling guidance map for the hummingbird search algorithm.

[0120] Based on this potential map, the hummingbird search algorithm automatically generates search paths. Each path corresponds to a complete scheduling path for a group of resources to reach the target event location from the current location, and includes key fields such as estimated time consumption, resource matching degree, and constraint satisfaction. During the search process, through the path perturbation mechanism and the path jump mechanism, the search algorithm jumps out of multiple potential local optimal combinations, and finally generates 89 candidate resource-event matching path solutions, and automatically screens out 34 optimal paths as the final scheduling output.

[0121] During the execution of the plan, all resource agents start according to the scheduling path instructions issued by the system, and upload information such as location information, task progress, and response status to the system in real time. The system compares and analyzes the resource execution path with the original matching path, and feeds it back to the evolutionary hybrid perception model of the agent for updating the state perception embedding and the policy response offset parameters. In the subsequent response to newly added events, the system automatically increases the response priority value of the high-success-rate area and reduces the scheduling score of the area with frequent constraint conflicts according to the feedback of the previous round, so as to realize the automatic optimization and iteration of the overall strategy in the second-round scheduling.

[0122] Through the scheduling system of the method of the present invention, within the first scheduling cycle (30 minutes), resource arrival instructions for 18 out of 21 event points were issued. The average time taken to generate the response path was 4.2 seconds, and the average number of convergence rounds for a single hummingbird search was 12 rounds. Compared with the original rule - engine - based scheduling system in this city, the average time taken by the original system to generate the scheduling path was 17.5 seconds, and there were problems such as repeated issuance of resource instructions and task area conflicts. Through manual review and comparison, after using the system of the present invention, the scheduling hit rate increased by 41.3%, the resource utilization rate increased by 38.7%, the average event response was completed 15 minutes earlier, and the failure rate of scheduling tasks decreased by 62% in subsequent feedback.

[0123] In the event simulation of this embodiment, the average resource path length generated by the scheduling system of the present invention was 6.8 kilometers, and the average resource arrival event time was 18.3 minutes; while the average path length of the traditional system was 8.5 kilometers, and the arrival time was 23.4 minutes. In the same scenario, the system of the present invention generated a total of 89 paths, among which 34 were the optimal execution paths selected by the system, and 32 tasks were successfully responded to, with a task success rate reaching 94.1%. The traditional system scheduled a total of 27 tasks, among which 19 were successfully responded to, with a success rate of 70.4%. In addition, the response time of each agent in the system of the present invention was less than 0.5 seconds, the hummingbird search algorithm could converge after an average of 12 iterations, and the average refresh period of the scheduling potential map was 15 seconds. The resource repeated scheduling rate decreased to 4.7% in this system, while this indicator was 11.2% in the traditional scheduling method.

[0124] It can be seen from this that the method for allocating emergency resources for public security events based on multi - agent collaborative optimization proposed by the present invention can quickly establish matching relationships, optimize path allocation in a complex, dynamic, multi - target scheduling environment, and has the iterative optimization ability of scheduling - perception - feedback linkage, significantly improving the emergency response efficiency, resource allocation rationality and task success rate, and having extremely high engineering adaptability and popularization value.

[0125] The above - mentioned are only the preferred specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, making equivalent substitutions or changes should be covered within the protection scope of the present invention.

Claims

1. A method for allocating emergency resources for public security incidents based on multi-agent collaborative optimization, characterized in that It includes the following steps: S1. Construct an event set, a resource set, and a constraint set to form a three-dimensional state tensor field; S2. Initialize an event agent, a resource agent, and a scheduling agent based on the three-dimensional state tensor field, construct a resource scheduling solution space, and configure an evolutionary hybrid perception model for each agent to output a scheduling potential map; S3. Execute the hummingbird search algorithm to perform target-driven path search within the resource scheduling solution space. The hummingbird search algorithm selects nectar source positions based on the scheduling potential map and generates a resource matching path; S4. Establish a bidirectional coupling mechanism between the hummingbird search algorithm and the evolutionary hybrid perception model. The resource matching path of the hummingbird search algorithm is fed back to the evolutionary hybrid perception model of each agent to adjust the perception accuracy and the strategy deviation direction, realizing dynamic collaborative optimization; S5. Output the final resource scheduling plan and have each agent execute it in real time to form a scheduling decision closed-loop for emergency response to public security events.

2. The emergency resource allocation method for public security incidents based on multi-agent collaborative optimization according to claim 1, wherein The three-dimensional state tensor field is used to describe the multi-dimensional dynamic association relationship between the event set, the resource set, and the constraint set. The construction method is as follows: Among them, T i,j,k represents the state score of resource r k under the constraint condition c j in response to event e i , where c k represents the k-th constraint condition, r j represents the j-th type of resource, and e i represents the i-th event, and ∈ represents the smoothing factor; P(e i ,r j ) represents the task priority weight of event e i for resource r j , and D(r j ,e i ) represents the path cost from resource r j to event e i , including response time and traffic accessibility. The larger the value of P(e i ,r j ), the more difficult the response is; A(c k ) represents a constraint activation factor, which takes a value of 1 when the constraint condition c k is satisfied, and 0 otherwise; L(r j ) represents the resource load factor, that is, the proportion of the current task of resource r j occupied, with a range of [0, 1]. The larger the value of L(r j ), the higher the load; The three-dimensional state tensor field is updated by rolling with time, and is used to reflect the joint scheduling feasibility of the event state, the resource state, and the environmental constraints in real time.

3. The emergency resource allocation method for public security incidents based on multi-agent collaborative optimization according to claim 1, wherein The specific content of S2 includes: S21. Construct an event agent according to each event in the event set. The attributes of the event agent include the spatial position, response time limit, risk level, and diffusion trend prediction parameters of the event, and are used to extract the associated resource subset and constraint condition slice in the three-dimensional state tensor field; S22. Construct a resource agent according to each resource in the resource set. The attributes of the resource agent include the resource type, current state, physical position, task load, and response ability, and are used to perceive the current potential event response tasks based on the three-dimensional state tensor field; S23. The scheduling agent serves as a coordination unit, is used to construct a mapping matrix between events, resources, and constraints, form a scheduling candidate set and a resource scheduling solution space structure, and is responsible for collecting perception information and organizing global feedback; S24. Configure an evolutionary hybrid perception model for each agent. The evolutionary hybrid perception model includes a state perception embedding module, an evolutionary control module, and a scheduling tendency module; The state perception embedding module is used to input tensor slice data and generate z through a graph attention encoder A , where z A represents the perception embedding representation of the A-th agent in the current state; The evolutionary control module maintains the behavior trajectory, historical feedback results, and current state weight of each agent, and executes the state update function: Among them, ω A (t + 1) represents the behavior participation degree of the A-th agent at time t + 1, ω A (t) represents the behavior participation degree of the A-th agent at time t, tanh represents the hyperbolic tangent function, α, β, and γ represent adjustment parameters, ΔR A (t) represents the increased value of the resource response benefit after the A-th agent executes the policy, represents the policy gradient, ||·|| represents the norm; The scheduling tendency module is based on z A and ω A (t + 1) to output a scheduling response tendency value: η A = σ(ω A (t + 1)·W1·tanh(W2·z A + b)); Among them, η A represents the scheduling response tendency value of the A-th agent, η A ∈[0,1], σ represents the activation function, W1 and W2 represent the parameter matrices, and b represents the offset parameter; S25. Fuse the scheduling response tendency values generated by all event agents and resource agents to construct a scheduling potential map: Among them, Ψ(x, y) represents the scheduling potential map, (x, y) represents the location coordinates of the target area, exp represents the exponential function, and η ij represents the joint scheduling response score of the i-th event agent and the j-th resource agent, δ represents the spatial diffusion adjustment coefficient, and (x i , y i ) represents the spatial location coordinates of the event, and (x j , y j ) represents the spatial location coordinates of the resource, n represents the total number of event agents, and m represents the total number of resource agents; S26. Use the scheduling potential map as the basic input map for nectar source selection and path search in the hummingbird search algorithm, and is used to dynamically guide the generation of the resource scheduling path and the adjustment of the optimization direction.

4. The emergency resource allocation method for public security incidents based on multi-agent collaborative optimization according to claim 1, wherein The specific content of S3 includes: S31. Use the scheduling potential map as the input, randomly select spatial positions with potential values greater than the set threshold as the initial search nodes. The initial search nodes and the reachable resources and events form the path search starting points, and all starting points form the initial position set of the hummingbird search algorithm; S32. Construct a population of hummingbird search individuals based on the initial position set. Each search individual corresponds to a set of resource scheduling path combinations. The resource scheduling path combinations include: starting position, target event position, resource number, intermediate node sequence in the resource movement path, expected response time, and path traffic status. All fields are stored in a structured manner. S33. Generate an initial search direction for the path combination based on the spatial coordinate relationship of each node in the scheduling potential graph. Each search individual uses the direction of the line between the starting point and the target as the initial path direction, and gradually inserts reachable nodes along the initial path direction to generate the first round of candidate path solutions. S34. During the search process, two types of path expansion methods are performed on each search entity: Local path perturbation: Insert a graph-reachable, normally accessible adjacent node into any segment of the current path plan to expand the path selection space and optimize the timeliness of resource arrival events. Remote path jump method: A non-continuous node is forcibly inserted between the previous and next segments of the path, and the connection path is calculated to see if it is within the allowed range of the access constraints. If so, the new path is retained; otherwise, the path is rolled back to the previous round of path status. S35. For search individuals whose path evaluation scores have not improved for three consecutive rounds, clear the current path record, select a new path search starting point from a high-potential node that has not been selected and whose potential value is greater than the set threshold, and regenerate a starting point-target combination to replace the original search direction and path content; S36. After completing all preset search rounds, several groups of paths with complete paths, normal traffic status, shortest expected response time, and highest resource-event matching degree are selected from all search individuals as the final resource matching path output.

5. The emergency resource allocation method for public security incidents based on multi-agent collaborative optimization according to claim 1, wherein The S4 specifically includes: S41. After each round of path search by the Hummingbird search algorithm is completed, the path result corresponding to each search individual is extracted based on the resource matching path, including: the resource number used, the resource starting position, the scheduling target position, the node sequence passed by the path, the total path length, the traffic status evaluation value and the estimated arrival time, and encapsulated into a feedback data packet; S42: Distribute the feedback data packet to the resource agents, scheduling agents, and corresponding event agents involved in the path. The feedback data packet received by each agent corresponds to its actual participation status in the path, ensuring that the path information feedback has an agent identity matching relationship. S43. A bidirectional coupling mechanism is established between the Hummingbird search algorithm and the evolutionary hybrid perception model. Each agent that receives the feedback data packet performs two update tasks on the evolutionary hybrid perception model based on its own role in the path and the result status of the task executed: Updating the input fields of the state-aware embedding module, including resource traffic status, path scheduling success rate, and adjacent event density distribution; Updating the offset parameter of the scheduling tendency module to adjust the priority direction of resource-oriented events in the scheduling response strategy; S44. In the next round of scheduling tasks, the updated state perception embedding module will use the new path feedback as the historical input feature. When generating the scheduling response tendency value, the updated scheduling tendency module will automatically increase the scheduling response tendency value in the location area where the success rate in historical scheduling is greater than 90%, and at the same time decrease the scheduling response tendency value in the location area where the failure rate is greater than 50%. S45. After receiving the input field of the updated state perception embedding module and the offset parameter of the scheduling tendency module, the hummingbird search algorithm re-initializes the path starting point selection logic of the search individuals, realizes the dependence of the search behavior on perception, and realizes the real-time coupling and dynamic cooperative optimization between the evolutionary hybrid perception model and the hummingbird search algorithm.

Images