Multi-agent collaborative decision-making and task allocation method for urban governance

Abstract

The invention discloses a multi-agent collaborative decision and task allocation method for urban governance, and belongs to the technical field of urban governance methods. Comprising the following steps: step S100, task modeling and publishing; step S200, performing agent cluster discovery and capability evaluation; step S300, task bidding and bidding based on an improved contract network protocol; step S400, making a bidder decision based on multi-objective optimization; step S500, performing distributed task execution and collaboration; and step S600, fusing and feeding back task results. According to the invention, through a layered architecture combining cloud global optimization and edge autonomous collaboration, the problems of delay and robustness of centralized decision making are effectively solved, and the real-time requirement of urban governance is met. Meanwhile, according to the method, global optimization of perception resource efficiency is achieved by fusing an intelligent task allocation model of multi-dimensional constraints, and the execution efficiency and the resource utilization rate of complex tasks are remarkably improved.

Description

Technical Field

[0001] This invention provides a multi-agent collaborative decision-making and task allocation method for urban governance, belonging to the technical field of urban governance methods. Background Technology

[0002] With the accelerating pace of urbanization, urban governance faces multiple complex challenges, including traffic congestion, emergencies, public safety, and environmental monitoring. To address these challenges, smart city systems have deployed a massive number of sensing terminals, including road surveillance cameras, unmanned aerial vehicles (UAVs), vehicle-mounted IoT devices, and environmental sensors, forming a vast urban sensing neural network. These terminals generate massive amounts of multimodal data in real time, providing a valuable information foundation for urban governance. However, the existing technological framework mainly relies on a centralized control center for data processing and decision-making. All sensing data is transmitted to the cloud or a central server for unified analysis before commands are issued. This model has significant bottlenecks: First, the remote transmission of massive amounts of data puts enormous pressure on communication network bandwidth, introducing high latency and making it difficult to meet the stringent timeliness requirements of emergency command and real-time traffic management scenarios. Second, centralized systems have heavy computational loads, easily creating performance bottlenecks and posing a single point of failure risk, resulting in insufficient system robustness. Third, sensing terminals from different sources often belong to different management systems, forming "data silos" and lacking effective collaboration mechanisms, leading to low resource utilization efficiency and difficulty in achieving optimized scheduling across regions and tasks from a global perspective.

[0003] In recent years, Multi-Agent System (MAS) technology has provided a new approach to solving distributed problems. This technology abstracts each sensing terminal or computing unit into an agent with autonomous decision-making capabilities, enabling complex tasks to be completed through communication and collaboration among agents. Although some research has attempted to introduce MAS into urban governance, existing methods mostly remain at the theoretical level or in simple application scenarios. Their main shortcomings are: 1. The task allocation model is overly simplified, usually based on fixed rules or simple auction algorithms, failing to fully consider the spatiotemporal dynamics of urban tasks, the differences in capabilities of heterogeneous agents (such as the complementarity of fixed camera perspectives and the maneuverability of drones), and the decay of time-sensitive value in task execution, resulting in insufficient global optimization of the allocation scheme; 2. The collaborative decision-making mechanism among agents is inefficient, lacking a hybrid decision-making architecture that integrates local sensing information and global task status. Agents often make decisions based on limited local information, easily getting trapped in local optima and unable to effectively cope with complex situations such as dynamic changes in task priorities and resource competition in the urban environment. Therefore, there is an urgent need for a multi-agent collaborative decision-making and task allocation method for large-scale, heterogeneous, and dynamic urban environments, in order to achieve efficient collaboration of sensing resources and global optimization of task execution. Summary of the Invention

[0004] The technical problem to be solved by this invention is how to overcome the shortcomings of high latency, poor robustness and low resource coordination efficiency of the centralized decision-making architecture in existing smart city systems, and realize global optimization and autonomous and efficient coordination of task allocation for large-scale heterogeneous sensing terminals in dynamic urban environments.

[0005] To address the aforementioned problems, the present invention proposes the following technical solution: a multi-agent collaborative decision-making and task allocation method for urban governance, wherein the method operates on a hierarchical computing architecture, comprising a cloud-based global decision-making layer and an edge-side agent collaboration layer; the method includes the following steps:

[0006] Step S100: Task Modeling and Deployment: The cloud-based global decision-making layer receives urban governance task requests, performs formal modeling on each task T_i, and generates a task description tuple T_i =<LT_i,PT_i,DT_i,RT_i,VR(T_i)> Where LT_i represents the spatial location where the task occurs, PT_i represents the task priority, DT_i represents the task description information, RT_i represents the set of resource types required by the task, and VR(T_i) represents the value decay function of the task.

[0007] Step S200: Agent Cluster Discovery and Capability Assessment: For task T_i, based on its spatial location LT_i and the set of required resource types RT_i, agents whose spatial coverage and capability range intersect are discovered in the agent registry center, forming a candidate agent cluster A_candidate; for each agent A_j in the cluster, its comprehensive capability fit C_ij relative to task T_i is evaluated.

[0008] Step S300: Task bidding and tendering based on the improved contract network protocol: The cloud-based global decision layer, acting as the manager, publishes the bidding information for task T_i to the candidate agent cluster A_candidate; the candidate agent A_j calculates the bid value Bid_ij based on its own state, current load, and comprehensive capability suitability C_ij, and returns the bid information to the manager;

[0009] Step S400: Winner decision based on multi-objective optimization: The cloud-based global decision layer collects bidding information, constructs a multi-objective optimization model with the objectives of maximizing global task completion efficiency and minimizing total system energy consumption, solves the model using an improved multi-objective genetic algorithm, and selects the winning agent or agent alliance.

[0010] Step S500: Distributed Task Execution and Coordination: The winning agent moves to the designated location or adjusts its perception posture according to the task description DT_i to execute the task; during the execution process, the agents share information and coordinate behaviors through the edge communication network to cope with the dynamic environment;

[0011] Step S600: Task Result Fusion and Feedback: Each agent uploads its local execution results to a designated node for data fusion, generates the final task execution report, and feeds it back to the cloud-based global decision layer to update the system status and task library.

[0012] Preferably, in step S100, the value decay function VR(T_i) of the task is defined as a monotonically non-increasing function with respect to the task waiting time t, and its expression is:

[0013] VR(T_i)(t)=V_0*exp(-λ_i*t)

[0014] Where V_0 is the initial value of the task, λ_i>0 is the value decay coefficient unique to task T_i, and the higher the priority PT_i of the task, the larger its λ_i value.

[0015] Preferably, in step S200, the formula for calculating the comprehensive capability fit C_ij is:

[0016] C_ij=ω1*f1(SpatialCoverage_ij)+ω2*f2(AbilityMatch_ij)+ω3*

[0017] f3(CurrentLoad_j)+ω4*f4(EnergyCost_ij)

[0018] Where f1(SpatialCoverage_ij) is the spatial coverage function of agent A_j to task location LT_i. If A_j is a fixed camera, it is the intersection measure of the view coverage area and the task area. If A_j is a drone, it is the reciprocal of the time required to maneuver from its current position to the task area. f2(AbilityMatch_ij) is the capability matching function, which is calculated based on the degree of matching between the perception capabilities of agent A_j (such as video resolution and sensor type) and the resource type RT_i required by the task. f3(CurrentLoad_j) is the current load rate function of agent A_j. The higher the load, the smaller this value.

[0019] f4(EnergyCost_ij) is the energy cost function, which estimates the energy consumption required for agent A_j to perform task T_i; ω1,ω2,ω3,ω4 are weight coefficients, and ω1+ω2+ω3+ω4=1.

[0020] Preferably, in step S300, the formula for calculating the bid value Bid_ij calculated by the intelligent agent A_j is:

[0021] Bid_ij=α*C_ij-β*Cost_ij

[0022] Where C_ij is the overall capability fit of agent A_j for task T_i, Cost_ij is the estimated cost required by agent A_j to execute task T_i, including time cost, energy cost, and communication cost; α and β are weight coefficients greater than 0, used to adjust the proportion of capability and cost in the bidding.

[0023] Preferably, in step S400, the objective function of the multi-objective optimization model is as follows:

[0024] Maximize:F1(X)=Σ_i(VR(T_i)(t_i)*y_i)

[0025] Minimize:F2(X)=Σ_ij(E_ij*x_ij)

[0026] Where X is the set of decision variables, x_ij is a 0-1 variable, if agent A_j is assigned to task T_i, then x_ij = 1, otherwise 0; y_i is a 0-1 variable, if task T_i is successfully assigned (i.e. at least one agent is assigned), then y_i = 1, otherwise 0; E_ij is the estimated energy consumption of agent A_j executing task T_i; t_i is the estimated start time of task T_i; the improved multi-objective genetic algorithm adopts the fast non-dominated sorting genetic algorithm (NSGA-II) with elite retention strategy, and its chromosome encoding uses integer encoding to directly represent the assignment relationship from task to agent.

[0027] Preferably, in step S400, when using the multi-objective genetic algorithm to solve the problem, the task value decay function VR(T_i)(t) defined in claim 2 is incorporated into the optimization process as a time constraint. For allocation schemes where the expected start time t_i is much greater than the current time, the value of VR(T_i)(t_i) will be significantly reduced, thereby guiding the algorithm to select an allocation scheme that can execute high-priority tasks faster.

[0028] Preferably, in step S500, the information sharing and behavioral coordination among the agents specifically involves: agents performing the same or related tasks forming a temporary collaborative group, with a virtual leader agent defined within the group or a distributed consensus algorithm employed; agents periodically broadcasting their local observation states (such as their own position, observed target information, and remaining energy) to other members of the group, and each agent adjusting its behavior based on the received neighbor information using a local controller based on the potential field method or model predictive control (MPC) to avoid collisions, cover blind spots, or collaboratively track targets.

[0029] Preferably, the system for implementing the method described above includes:

[0030] The task management module, deployed in the cloud-based global decision-making layer, is used to receive, model, and publish urban governance tasks, and to execute the decisions of the winning bidder.

[0031] The agent management module is used to register, discover, and manage all agents, and maintain the status, capabilities, and location information of the agents.

[0032] The communication middleware module provides reliable, low-latency message passing services between the cloud and the edge, as well as edge intelligent agents.

[0033] The local decision-making and collaboration module is deployed on each agent or its associated edge computing node, enabling agents to autonomously bid, plan local paths, and collaborate with other agents.

[0034] The data fusion and feedback module is used to fuse local data submitted by multiple agents and generate a unified task report.

[0035] The beneficial effects of this invention are:

[0036] 1. By employing a layered collaborative architecture of "cloud-edge," complex global optimization calculations are placed in the cloud, while real-time, local collaborative decision-making is delegated to edge agents. This effectively reduces network bandwidth pressure, significantly lowers decision latency, and meets the real-time requirements of urban governance tasks (such as emergency response). Simultaneously, the distributed architecture avoids the risk of single points of failure, resulting in stronger system robustness.

[0037] 2. By introducing a cost-benefit model that integrates spatiotemporal attributes, task value decay, agent multidimensional capabilities, and energy consumption, and by using an improved contract network protocol and multi-objective optimization algorithm for task allocation, the allocation scheme is no longer a simple "nearby assignment" or "round-robin allocation," but a globally optimal solution that comprehensively considers task urgency, agent suitability, and total system energy consumption. This significantly improves the utilization efficiency of urban sensing resources and the overall completion efficiency of complex tasks. Attached Figure Description

[0038] Figure 1 This is a flowchart of a multi-agent collaborative decision-making and task allocation method for urban governance according to the present invention. Detailed Implementation

[0039] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. The embodiments of the present invention include, but are not limited to, the following examples.

[0040] The intelligent agents involved in this invention include, but are not limited to, fixed cameras, unmanned aerial vehicles (UAVs), and mobile patrol vehicles. The cloud-based global decision-making layer can be deployed on the cloud server of the city operations center (IOC), while the edge-side intelligent agent collaboration layer relies on each intelligent agent itself or the edge computing nodes it connects to.

[0041] Example 1: Collaborative Traffic Management and Evidence Collection in Response to Sudden Traffic Incidents

[0042] Task Modeling and Deployment (S100): The cloud platform receives an alert for a "traffic accident at XX intersection" from the traffic detection system or a citizen. The platform immediately generates a task T1, whose tuples are: LT1 (GPS coordinates of the accident intersection), PT1 (height, corresponding to λ1 = 0.1), DT1 (capturing multi-angle high-definition video of the accident scene and monitoring traffic flow changes), RT1 ({high-definition video stream, traffic flow analysis}), VR(T1)(t) = 100*exp(-0.1t) (value unit representation).

[0043] Intelligent Agent Cluster Discovery and Capability Assessment (S200): Based on LT1, the platform discovers candidate intelligent agents within a 500-meter radius of the registry center: a fixed camera A1 (facing the intersection), a drone A2 (on standby) in a drone base station, and a smart police car A3 (patrolling nearby). Their capability suitability C_ij is assessed: A1 has high f1 (spatial coverage) but a fixed viewing angle; A2 has extremely high f1 (maneuverability) and can fly to the optimal angle; A3 has moderate f1 (coverage) but possesses on-site processing capabilities. The calculated C_1j values ​​are: A1: 0.85, A2: 0.90, A3: 0.70.

[0044] Task Bidding and Tendering (S300): The platform issues T1 tenders to A1, A2, and A3. A1 tender Bid_11 = 1.0 * 0.85 - 0.2 * 0.1 (low energy cost) = 0.83; A2 tender Bid_12 = 1.0 * 0.90 - 0.2 * 0.5 (higher flight energy consumption) = 0.80; A3 tender Bid_13 = 1.0 * 0.70 - 0.2 * 0.3 (mobile energy consumption) = 0.64.

[0045] Winner-of-the-Bid Decision (S400): The platform constructs a multi-objective optimization model. Considering the rapid decay of task value (high priority), a fast response is required. After balancing global efficiency (F1) and energy consumption (F2), the optimization algorithm decides to form an intelligent agent alliance: A1 immediately provides a real-time main perspective, while A2 is assigned to take off and go to the site to conduct multi-angle shooting and aerial monitoring. This solution has the highest F1 value (total task value), although F2 (total energy consumption) is not the lowest, it meets the core requirement of timeliness.

[0046] Distributed task execution and collaboration (S500): A1 continuously monitors. After A2 takes off, it establishes edge-side communication with A1. Based on the on-site images provided by A1, A2 autonomously plans its flight path to avoid congestion points and adjusts its camera angle to avoid overlapping with A1's images, achieving collaborative coverage. A2 transmits key images captured back to the edge nodes in real time.

[0047] Task Result Fusion and Feedback (S600): The edge node fuses the continuous flow of A1 and the key segments of A2, along with traffic flow analysis data, to generate a complete report containing the entire accident process, auxiliary information for liability determination, and traffic impact assessment. This report is then submitted to the cloud platform and the traffic police command center. The platform updates the task status and marks A2 and A3 as available.

[0048] Example 2: Regional Crowd Density Monitoring During Large-Scale Events

[0049] Task Modeling and Release (S100): A large-scale concert is held in the city. The cloud platform generates a periodic task T2: LT2 (1 square kilometer area around the stadium), PT2 (middle, λ2=0.03), DT2 (continuously monitor the heat map of the crowd and give early warning of abnormal gatherings), RT2 ({video crowd count, infrared thermal sensing}), VR(T2)(t)=80*exp(-0.03t).

[0050] Intelligent Agent Cluster Discovery and Capability Assessment (S200): The platform discovers fixed cameras A4, A5, and A6 within the area, as well as dispatchable drones A7 and A8. Their capabilities are assessed: A4 and A5 have crowd counting capabilities; A6 provides general surveillance; A7 and A8 are equipped with high-definition and infrared cameras.

[0051] Task Bidding and Tendering (S300 / S400): The platform conducts bidding. When solving the problem, the optimization algorithm, considering the long task cycle and medium priority, prefers to select fixed cameras with lower energy consumption as the basic coverage. Therefore, the decision is as follows: fixed cameras A4 and A5 are assigned as the basic monitoring task, forming a static monitoring network. Simultaneously, drone A7 is designated as a mobile supplementary unit, tasked with verifying and performing detailed temperature measurements when A4 or A5 reports excessive density in a certain area.

[0052] Distributed task execution and coordination (S500): A4 and A5 report crowd density data at a predetermined frequency. When A5 detects a sudden increase in density at the southeast entrance, it immediately sends a coordination request to the mobile unit A7 through the edge coordination layer. Based on the received location and density data from A5, A7 autonomously plans the optimal path to fly to the southeast entrance and switches to infrared mode to assess crowd density and abnormal body temperature.

[0053] Task Result Fusion and Feedback (S600): A7 fuses the confirmed abnormal clustering data and infrared data with the data from A5 to generate a high-level early warning report, which is then sent to the platform. The platform can automatically trigger contingency plans based on this, such as notifying on-site security personnel to manage passenger flow. Upon completion of the task cycle, an overall crowd flow analysis report for the area is generated.

[0054] The above embodiments demonstrate that the method of the present invention can flexibly and efficiently address different types of urban governance tasks, realize dynamic optimization of sensing resources and autonomous collaboration among intelligent agents, and significantly improve the level of intelligence and emergency response capabilities of urban management.

[0055] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A multi-agent collaborative decision-making and task allocation method for urban governance, characterized in that, The method operates on a hierarchical computing architecture, which includes a cloud-based global decision-making layer and an edge-side intelligent agent collaboration layer; the method includes the following steps: Step S100: Task Modeling and Deployment: The cloud-based global decision-making layer receives urban governance task requests, performs formal modeling on each task T_i, and generates a task description tuple T_i =<LT_i,PT_i,DT_i,RT_i,VR(T_i)> Where LT_i represents the spatial location where the task occurs, PT_i represents the task priority, DT_i represents the task description information, RT_i represents the set of resource types required by the task, and VR(T_i) represents the value decay function of the task. Step S200: Agent Cluster Discovery and Capability Assessment: For task T_i, based on its spatial location LT_i and the set of required resource types RT_i, agents whose spatial coverage and capability range intersect are discovered in the agent registry center, forming a candidate agent cluster A_candidate; for each agent A_j in the cluster, its comprehensive capability fit C_ij relative to task T_i is evaluated. Step S300: Task bidding and tendering based on the improved contract network protocol: The cloud-based global decision layer, acting as the manager, publishes the bidding information for task T_i to the candidate agent cluster A_candidate; the candidate agent A_j calculates the bid value Bid_ij based on its own state, current load, and comprehensive capability suitability C_ij, and returns the bid information to the manager; Step S400: Winner decision based on multi-objective optimization: The cloud-based global decision layer collects bidding information, constructs a multi-objective optimization model with the objectives of maximizing global task completion efficiency and minimizing total system energy consumption, solves the model using an improved multi-objective genetic algorithm, and selects the winning agent or agent alliance. Step S500: Distributed Task Execution and Coordination: The winning agent moves to the designated location or adjusts its perception posture according to the task description DT_i to execute the task; during the execution process, the agents share information and coordinate behaviors through the edge communication network to cope with the dynamic environment; Step S600: Task Result Fusion and Feedback: Each agent uploads its local execution results to a designated node for data fusion, generates the final task execution report, and feeds it back to the cloud-based global decision layer to update the system status and task library.

2. The method according to claim 1, characterized in that, In step S100, the value decay function VR(T_i) of the task is defined as a monotonically non-increasing function with respect to the task waiting time t, and its expression is: VR(T_i)(t)=V_0*exp(-λ_i*t) Where V_0 is the initial value of the task, λ_i>0 is the value decay coefficient unique to task T_i, and the higher the priority PT_i of the task, the larger its λ_i value.

3. The method according to claim 1, characterized in that, In step S200, the formula for calculating the comprehensive capability fit C_ij is: C_ij=ω1*f1(SpatialCoverage_ij)+ω2*f2(AbilityMatch_ij)+ω3*f3(CurrentLoad_j)+ω4*f4(EnergyCost_ij) Where f1(SpatialCoverage_ij) is the spatial coverage function of agent A_j to task location LT_i. If A_j is a fixed camera, it is the intersection measure of the view coverage area and the task area. If A_j is a drone, it is the reciprocal of the time required to maneuver from its current position to the task area. f2(AbilityMatch_ij) is the capability matching function, which is calculated based on the degree of matching between the perception capabilities of agent A_j (such as video resolution and sensor type) and the resource type RT_i required by the task. f3(CurrentLoad_j) is the current load rate function of agent A_j. The higher the load, the smaller this value. f4(EnergyCost_ij) is the energy cost function, which estimates the energy consumption required for agent A_j to perform task T_i; ω1,ω2,ω3,ω4 are weight coefficients, and ω1+ω2+ω3+ω4=1.

4. The method according to claim 1, characterized in that, In step S300, the formula for calculating the bid value Bid_ij calculated by the intelligent agent A_j is as follows: Bid_ij=α*C_ij-β*Cost_ij Where C_ij is the overall capability fit of agent A_j for task T_i, Cost_ij is the estimated cost required by agent A_j to execute task T_i, including time cost, energy cost, and communication cost; α and β are weight coefficients greater than 0, used to adjust the proportion of capability and cost in the bidding.

5. The method according to claim 1, characterized in that, In step S400, the objective function of the multi-objective optimization model is as follows: Maximize:F1(X)=Σ_i(VR(T_i)(t_i)*y_i) Minimize:F2(X)=Σ_ij(E_ij*x_ij) Where X is the set of decision variables, x_ij is a 0-1 variable, if agent A_j is assigned to task T_i, then x_ij = 1, otherwise 0; y_i is a 0-1 variable, if task T_i is successfully assigned (i.e. at least one agent is assigned), then y_i = 1, otherwise 0; E_ij is the estimated energy consumption of agent A_j executing task T_i; t_i is the estimated start time of task T_i; the improved multi-objective genetic algorithm adopts the fast non-dominated sorting genetic algorithm (NSGA-II) with elite retention strategy, and its chromosome encoding uses integer encoding to directly represent the assignment relationship from task to agent.

6. The method according to claim 5, characterized in that, In step S400, when using the multi-objective genetic algorithm to solve the problem, the task value decay function VR(T_i)(t) defined in claim 2 is incorporated into the optimization process as a time constraint. For allocation schemes where the expected start time t_i is much greater than the current time, the value of VR(T_i)(t_i) will be significantly reduced, thereby guiding the algorithm to select an allocation scheme that can execute high-priority tasks faster.

7. The method according to claim 1, characterized in that, In step S500, the information sharing and behavioral coordination among the agents are specifically as follows: agents performing the same or related tasks form a temporary collaborative group, within which a virtual leader agent is defined or a distributed consensus algorithm is adopted; agents periodically broadcast their local observation states (such as their own position, observed target information, and remaining energy) to other members of the group, and each agent adjusts its own behavior based on the received neighbor information using a local controller based on the potential field method or model predictive control (MPC) to avoid collisions, cover blind spots, or collaboratively track targets.

8. A multi-agent collaborative decision-making and task allocation system for urban governance, characterized in that, The system for implementing the method of claim 1 includes: The task management module, deployed in the cloud-based global decision-making layer, is used to receive, model, and publish urban governance tasks, and to execute the decisions of the winning bidder. The agent management module is used to register, discover, and manage all agents, and maintain the status, capabilities, and location information of the agents. The communication middleware module provides reliable, low-latency message passing services between the cloud and the edge, as well as edge intelligent agents. The local decision-making and collaboration module is deployed on each agent or its associated edge computing node, enabling agents to autonomously bid, plan local paths, and collaborate with other agents. The data fusion and feedback module is used to fuse local data submitted by multiple agents and generate a unified task report.

Images