Intelligent agent online autonomous cooperative scheduling method based on Nash game and dynamic coupling

Abstract

The invention discloses an intelligent agent online autonomous cooperative scheduling method based on Nash game and dynamic coupling, and aims at solving the problems that in the prior art, centralized equipment and a global communication network excessively depend on, and the static coupling relation of an intelligent agent is rigid, the intelligent agent achieves a scheduling consensus through local communication and a local Nash equilibrium game. The dynamic coupling mechanism supports the intelligent agent to be online and offline according to needs, and synchronizes the state to the adjacent intelligent agent in real time. The method comprises the following specific steps: carrying out matching division based on expected earnings of the unmanned aerial vehicle, and selecting behavior points based on softmax probability; and carrying out local Nash equilibrium game according to the behavior income, and carrying out iterative adjustment until all agents reach an equilibrium state. Through the local Nash equilibrium game and the dynamic coupling scheduling, the robustness of the system is remarkably improved, and the communication overhead is reduced.

Description

Technical Field

[0001] This invention belongs to the technical field of multi-agent system scheduling, and further relates to a method for online autonomous collaborative scheduling of agents based on Nash games and dynamic coupling in the field of mobile crowd-sensing heterogeneous multi-agent system scheduling. By leveraging a real-time dynamic coupling relationship reconstruction mechanism and a local Nash equilibrium game, this invention achieves online autonomous collaborative scheduling of heterogeneous multi-agents, thereby improving the execution efficiency of perception data collection tasks. Background Art

[0002] With the rapid development of mobile crowdsensing technology, heterogeneous multi-agent collaborative scheduling systems have shown great potential for application in scenarios such as disaster relief and environmental monitoring. A multi-agent collaborative scheduling system is a network of heterogeneous agents, such as drones, unmanned vehicles, unmanned ships, robots, or mobile workers, that autonomously completes complex perception and search tasks in dynamic environments through mechanisms such as task allocation, path planning, and resource coordination. However, traditional multi-agent collaborative scheduling methods often suffer from low collaboration and inefficient perception and search.

[0003] Shandong University of Science and Technology proposed a reinforcement learning-based collaborative search method for drones and unmanned vessels in its patent application, "A reinforcement learning-based collaborative search method for drones and unmanned vessels with long endurance" (application number 202411473542.0, application publication number CN 119005305 A). This method effectively improves the collaboration and perception search efficiency between intelligent agents. The collaborative search method includes the following steps: first, multiple drones and unmanned vessels are deployed in the mission area, and the mission area is discretized into a grid. The probability of target existence in each cell is recorded using a probabilistic graphical model; when the drone's energy is insufficient, it sends a charging request to the unmanned vessel, which plans a path based on the drone's battery energy and the target existence probability; the unmanned vessel optimizes its own motion trajectory using a collaborative optimization scheduling algorithm, provides charging services for the drone, and searches for the target; if no sea surface target is found, the target existence probability of each cell is recalculated; if the drone's energy is insufficient to complete the mission, the drone charging request and the unmanned vessel optimize their own trajectory repeatedly until the target location is identified through the target probability. This search method has the following two shortcomings: first, this method relies on a multi-agent reinforcement learning method and needs to rely on centralized training equipment for global policy updates, and the agents need to share global network information in real time (such as the location of all drones, energy status, etc.). In actual extreme scenarios, centralized training equipment often cannot be deployed or operated normally, resulting in interruptions in policy updates; at the same time, frequent global communications between agents will generate high bandwidth occupancy and transmission delays. In large-scale scenarios, communication links are prone to congestion or even paralysis, resulting in decision-making delays or failures. If this limitation is ignored, the system will not be able to complete collaborative optimization due to information synchronization failure, causing task interruption or resource conflict, and even danger (such as collisions caused by overlapping planned paths of agents such as drones and unmanned ships). Second, this patented method relies on a static coupling relationship, that is, it assumes that the collaborative relationship between drones and unmanned ships is fixed (such as charging priority, search area allocation, etc.), and does not consider the impact of dynamic environments (such as sudden task insertion, temporary failure of agents, or access to new equipment) on collaborative logic. For example, if a drone goes offline due to a malfunction, the system cannot dynamically adjust the charging service targets of the remaining unmanned boats, resulting in idle charging resources in some areas and drones in other areas being shut down due to depleted batteries. Alternatively, when the actual mission volume is small, a large number of drones are idle online, while when the actual mission volume is large, there are insufficient drones to complete the various tasks, causing urgent tasks to be delayed. This rigid coordination mechanism severely weakens the system's ability to respond to dynamic environments, reducing search efficiency and robustness.

[0004] In its patent application, "Energy Optimization Method for Multi-Agent UAV Search Tasks Based on Entropy Maximization Strategy" (Application No. 202411264528.X, Publication No. CN 119151062 A), Beijing Institute of Technology proposed an energy optimization method for multi-agent UAV search tasks based on entropy maximization. This search method includes: modeling multi-region search tasks for multi-agent UAVs, constructing an energy consumption model for the UAVs, prioritizing minimizing the energy consumption of the multi-agent system, and using a multi-agent entropy maximization reinforcement learning algorithm to make decisions on the actions of each UAV. Based on a centralized training and distributed execution method, each UAV's next action decision is adjusted in real time based on the environment and energy consumption observed by other UAVs. When a UAV reaches its target area, it conducts a full-path search of the target area. The drawbacks of this approach are that, similar to centralized training, it requires global environmental information and centralized training equipment. If the central server experiences a single point of failure due to hardware failure, cyberattack, or environmental disaster, all agents (drones, unmanned boats, etc.) that rely on it for decision-making will lose their scheduling instructions and fall into chaos. In this global communication state, all agents must report their status (such as location, battery level, and sensor data) to the central node in real time and wait for the central node to return a global optimization strategy. In large-scale scenarios, massive data transmission consumes high bandwidth and causes network congestion. Furthermore, the online status between drones remains fixed, making it difficult to adapt to scenarios with dynamically changing environments. For example, a high-priority target suddenly appears outside the original mission area, requiring drones to be urgently dispatched for detection. However, the system's fixed task allocation mechanism cannot dynamically release idle drone resources. Alternatively, a drone may be taken offline due to a mechanical failure, but the system has not dynamically brought other drones online to remedy the situation, resulting in a blind spot in that area. Summary of the Invention

[0005] The purpose of the present invention is to address the deficiencies in the above-mentioned prior art and propose an online autonomous collaborative scheduling method for intelligent agents based on Nash game and dynamic coupling, aiming to solve the problems in the prior art of single point failure and global collapse caused by excessive reliance on centralized equipment, the problem of communication link congestion and paralysis causing decision-making delays and failures caused by excessive reliance on the global communication network, and the problem of inability to respond to emergencies in a timely manner or waste of a large amount of online resources caused by excessive reliance on static coupling relationships between intelligent agents.

[0006] The technical idea for achieving the purpose of the present invention is that the method of the present invention uses distributed autonomous scheduling, so that the intelligent agent does not need to rely on centralized equipment and the global network for scheduling and communication. It only relies on its own communication capabilities to transmit information within the local communication range, and relies on the local Nash equilibrium game method to achieve scheduling consensus between nodes within the communication range, so as to efficiently achieve effective communication within each maximum complete subnet in the global network, and ensure the coordination of self-scheduling efficiency between each intelligent agent, effectively avoiding excessive dependence on centralized equipment and the global network. At the same time, the method of the present invention adopts a dynamic coupling method. The intelligent agent can go online and offline at any time according to the environmental conditions, and the state change information is synchronized to the neighboring nodes through local broadcast to avoid decision lag. At the same time, when an intelligent agent goes online and offline, the system does not need to reconstruct the scheduling strategy, effectively realizing dynamic coupling scheduling of multiple intelligent agents. In addition, this method significantly reduces the scheduling decision time of the intelligent agent through adaptive dimensionality reduction in the matching process, and ensures the efficiency of intelligent agent scheduling decision by probabilistic behavior selection. While ensuring the speed of the intelligent agent's self-scheduling decision-making, a large number of perception tasks are completed efficiently and collaboratively.

[0007] The steps for achieving the purpose of the present invention include the following:

[0008] Step 1: Calculate the expected benefits of each drone going to the task point within the local communication range and cooperating with workers to perform tasks. And the expected benefits of each drone going to a flight point within the local communication range and cooperating with the car to charge to replenish its own power Judging the same drone Is greater than or equal to If yes, the drone is selected to perform the task, and the “drone-worker-task point” matching scheduling is performed in the subsequent steps. Otherwise, the “drone-car-endurance point” matching scheduling is performed in the subsequent steps.

[0009] Step 2: Using the improved softmax function, the distance between each agent and the behavior point is converted into the probability p of selecting the behavior point. n , in order to make real-time online self-scheduling behavior decisions for dynamically coupled intelligent agents. Each intelligent agent goes online or offline strictly according to its own online and offline time. Each intelligent agent calculates the current behavior benefit based on the matching type. The benefits of drones and workers participating in the perception task are the number of successful perception task matches within their own communication range, and the benefits of drones and cars participating in battery charging are the sum of the power added to all drones within their own communication range.

[0010] Step 3: Based on the current behavior benefit of each agent, it plays a local Nash equilibrium game with other agents within its own communication range;

[0011] Step 4: Repeat steps 2 and 3 until each agent reaches a local Nash equilibrium state, that is, all agents reach a consensus, and at the current moment all agents cannot simply adjust their own behavior to increase their benefits, and execute the scheduling behavior of each agent at the current moment.

[0012] Furthermore, the expected return It is obtained from the following formula:

[0013]

[0014] Among them, uLeftT i Represents the i-th UAV uav i Remaining online time after executing the task, U_uptime i and U_downtime i Respectively represent UAV i Online time and offline time, ΔuPowSumT i UAV i The power consumed by executing the task, Fullpower i UAV i Fully charged.

[0015] Furthermore, the expected return It is obtained from the following formula:

[0016]

[0017] Among them, uLeftC i Represents the i-th UAV uav i Remaining online time after charging, ΔuPowSumC i Represents the i-th UAV uav i Charging replenishes the power.

[0018] Furthermore, the probability p of selecting a behavior point n as follows:

[0019]

[0020] Among them, e (·) Represents an exponential function with the natural constant e as the base, score n Indicates the agent's choice of behavior point uwvLoc n The score is the negative value of the distance, that is, score n =-uwvDis n .

[0021] Furthermore, the number of successful matches is:

[0022]

[0023] Reward_task ijk =∑matchSign ijx

[0024] Among them, uLoc i Indicates uav i Location, wLoc j Indicates worker j Location, tLoc x Indicates task x Position, matchSign ijx UAV i 、worker j and task x Match success or failure indicator, Reward_task ijk It represents the current behavior benefit of each drone going to the task point within the local communication range to cooperate with the workers to perform the task.

[0025] Furthermore, the total amount of electricity is:

[0026]

[0027] Reward_charge ijk =∑ΔPow iky

[0028] Among them, uLoc i Indicates uav i location, vLoc k Indicates vehicle k Location, chLoc y Indicates charge y Position, ΔPow iky Indicates vehicle vehocle k Charge at the endurance point y UAV i Increased power, Reward_charge ijk Represents the current behavior benefit of the intelligent agent, which is that the drone goes to a flight point within the local communication range and cooperates with the car to recharge its own power.

[0029] Furthermore, the local Nash equilibrium game refers to checking each agent UWV ijk Based on the selected behavior, adjust the self-scheduling behavior again to make it profitable: Reward_task ijk or Reward_chargeijk Is it improved? If so, its benefit indicator is: S ijk =0, otherwise, the income is marked as: S ijk =1; the UWV of the agent will not be upgraded ijk The benefit identifier of obtaining other agents within the communication range is: neighbors}; If the S of all agents within its communication range neighbors =1, then the agent is considered to be UWV ijk It has reached the local Nash equilibrium state, otherwise it is considered that it has not reached the local Nash equilibrium state.

[0030] Compared with the prior art, the present invention has the following advantages:

[0031] First, the distributed decision-making architecture designed by this invention allows each agent to autonomously schedule based solely on local communication information, without the need for central server support. This effectively addresses the existing over-reliance on centralized equipment and global communication networks. This eliminates the need for centralized equipment and global communication networks, allowing agents in this invention to communicate and coordinate within their local area, relying solely on their own communication capabilities. By leveraging local Nash equilibrium game methods to reach consensus among agents, they can efficiently collaborate and complete perception tasks.

[0032] Second, the present invention uses drones equipped with professional sensors, combined with worker-controlled obstacle avoidance and car charging support, to form a real-time collaborative scheduling mechanism for heterogeneous multi-agents. This overcomes the defect of existing technologies that are only applicable to heterogeneous agents with fixed scheduling coupling relationships. The present invention improves the professionalism of data collection to support autonomous collaboration among drones, workers, and cars in complex environments, and to promptly respond to complex and time-varying application scenarios under various emergencies.

[0033] Third, the agents of this invention can be brought online and offline in real time based on environmental conditions, overcoming the problem of static inter-agent coupling relationships in existing technologies, which leads to waste of online resources in emergencies. This allows each agent to go online or offline according to its own preset online and offline schedules, ensuring the normal operation of the online scheduling system at all times. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 is a flow chart of the method of the present invention;

[0035] Figure 2 It is the algorithm framework diagram of the present invention;

[0036] Figure 3 This is a flowchart of the matching dimensionality reduction based on the expected benefits of drones in the present invention;

[0037] Figure 4This is an example diagram of the autonomous scheduling of the intelligent agent based on softmax probability in the present invention;

[0038] Figure 5 Schematic diagram of the perception task matching process of the present invention. DETAILED DESCRIPTION

[0039] The embodiments of the present invention are described in further detail below with reference to the accompanying drawings.

[0040] Reference Figure 1 , the implementation steps of the embodiment of the present invention are further described in detail.

[0041] An embodiment of the present invention deploys the method of the present invention in a post-disaster environmental perception scenario, specifically involving three types of intelligent agents (drones, workers, and cars) and two types of behavior points (task points and endurance points). When a drone and a worker simultaneously match and arrive at a task point, they can begin collaborating to perform the task. When a drone and a car simultaneously match and arrive at an endurance point, they can begin collaborating to replenish the drone's battery. The present invention is characterized by relying on the intelligent agents' own communication capabilities to communicate information and match scheduling within a local area, reaching consensus among intelligent agents through a local Nash equilibrium game method, and achieving online scheduling based on dynamic coupling between intelligent agents.

[0042] Reference Figure 2 , the algorithm of the present invention is further described in detail.

[0043] Step 1: Calculate the expected benefits of each drone going to the task point within the local communication range and cooperating with workers to perform tasks. And the expected benefits of each drone going to a flight point within the local communication range and cooperating with the car to charge to replenish its own power Judging the same drone Is greater than or equal to If yes, the drone will choose to execute the task, and the “drone-worker-task point” matching scheduling will be carried out in the subsequent steps. Otherwise, the “drone-car-endurance point” matching scheduling will be carried out in the subsequent steps. Figure 3 As shown, drones are matched and divided according to power information and time status. The specific implementation is as follows:

[0044] Step 1.1, Estimation of expected benefits of UAV mission: First calculate the UAV i The nearest task within communication range distance And calculate its uLoc from the current position i Move to the mission point Reduced remaining battery life and the required flight time Then uav i Retrieve the closest task within its communication range worker j , and calculate the worker j Move to the mission point Time required Computing UAV i and workers j Complete the task Time required and UAV i Remaining online time Computing UAV i Execute the task The total power reduction is Define get drone uav i Execute the task Expected returns

[0045] Step 1.2, estimate the expected benefit of UAV charging endurance: First calculate the UAV i The nearest endurance point within the communication range distance And calculate its uLoc from the current position i Move to the endurance point Reduced remaining battery life and the required flight time Then uav i Retrieve the closest endurance point within its communication range vehicle k , and calculate vehicle k Move to the endurance point Time required Computing UAV i To the endurance point By car vehicle k Time required to fully charge and UAV i Remaining online time Computing UAV i Increase in battery life ΔuPowSumC i =Fullpower i -uPow i . Define the drone uav i go Expected benefits of charging range

[0046] Step 1.3, determine the same drone Is greater than or equal to If so, the drone is selected to perform the task, and the “drone-worker-task point” matching scheduling is performed in the subsequent steps. Otherwise, the “drone-car-endurance point” matching scheduling is performed in the subsequent steps.

[0047] Step 2: Using the improved softmax function, the distance between each agent and the behavior point is converted into the probability p of selecting the behavior point. n , in order to make real-time online self-scheduling behavior decisions for dynamically coupled intelligent agents. Each intelligent agent goes online or offline strictly according to its own online and offline time, such as Figure 4 As shown, the closer the behavior point is to the agent's own location, the higher the score and the greater the probability of selection; each agent calculates the current behavior benefit based on the matching type. The benefits of drones and workers participating in the perception task are the number of successful perception task matches within their communication range, and the benefits of drones and cars participating in battery charging are the total power added to all drones within their communication range. The specific implementation steps are as follows:

[0048] Step 2.1, each agent checks whether the current moment is within its own online and offline time interval, that is, whether it is online. Each online agent UWV ijk Calculate the distance between it and all selectable behavior points within the communication range {uwvDis0,...,uwvDis n ,...}, and record its negative value as score n =-ywvDis n Calculate the probability of selecting the behavior point based on the obtained score Among them, e (·) Represents an exponential function with the natural constant e as its base.

[0049] In step 2.2, each agent calculates the current behavior benefit according to the matching type, and obtains the benefits of the drone and worker participating in the perception task as the number of successful perception task matches within its own communication range.

[0050]

[0051] Reward_task ijk =∑matchSign ijx

[0052] Among them, uLoc i Represents the i-th UAV uav iLocation, wLoc j Represents the jth worker j Location, tLoc x Indicates task task x Position, matchSign ijx UAV i 、worker j and task x Match success or failure indicator, Reward_task ijk It represents the current behavior benefit of each drone going to the task point within the local communication range to cooperate with the workers to perform the task.

[0053] The benefits of drones and cars participating in the battery life charging are the sum of the power added by all drones within their communication range.

[0054]

[0055] Reward_charge ijk =∑ΔPow iky

[0056] Among them, vLoc k Indicates vehicle k Location, chLoc y Indicates the endurance point charge y Position, ΔPow iky Indicates vehicle k Charge at the endurance point y UAV i Increased power, Reward_charge ijk Represents the current behavior benefit of the intelligent agent, which is that the drone goes to a flight point within the local communication range and cooperates with the car to recharge its own power.

[0057] Step 3: Based on the current behavior benefit of each agent, it plays a local Nash equilibrium game with other agents within its own communication range. Figure 5 As shown in the figure, for example, the communication range of worker 3 overlaps with the communication ranges of worker 2, worker 4, and drone 2. When deciding whether to perform self-scheduling behavior, it must consider the local Nash equilibrium game among the three. The specific implementation is:

[0058] Check each agent UWV ijk Based on the selected behavior, adjust the self-scheduling behavior again to make it profitable: Reward_task ijk or Reward_charge ijkIs it improved? If so, the benefit mark is: S ijk =0, otherwise, the income is marked as: S ijk =1; the UWV of the agent will not be upgraded ijk The benefit identifier of obtaining other agents within the communication range is: neighbors}; If the S of all agents within the communication range neighbors =1, then the agent is considered to be UWV ijk The local Nash equilibrium state has been reached, otherwise, it is considered that the local Nash equilibrium state has not been reached.

[0059] Step 4: Repeat steps 2 and 3 until each agent reaches a local Nash equilibrium state, that is, all agents reach a consensus, and at the current moment all agents cannot simply adjust their own behavior to increase their benefits, and execute the scheduling behavior of each agent at the current moment.

[0060] The effect of the present invention can be further demonstrated through the following simulation.

[0061] 1. Simulation experiment conditions.

[0062] The software platform for the simulation experiment of the present invention is: Windows 11 operating system and Python 3.9.0.

[0063] 2. Simulation content and result analysis.

[0064] The simulation experiment of the present invention is as follows:

[0065] 2.1, the simulation experiment evaluates the performance of the proposed algorithm based on the perception task completion rate.

[0066] The simulation experiment of the present invention evaluates the performance of the proposed algorithm based on real data sets and simulated data sets. Among them, the real data comes from actual scenarios, including bicycle order data (BicycleOrder), taxi trajectory data (TaxiTrajectories) and Didi order data (DidiOrders). First, the present invention discretizes the coverage area of ​​the real data set. Then, the data is screened, classified and labeled to obtain data that meets the experimental specifications. The random simulated data set is generated through parameterization to flexibly adjust the area scale, number of tasks, agent scale, etc. to verify the performance of the algorithm in different scenarios. The detailed data set description is shown in Table 1. Table 1 Experimental data set list

[0067]

[0068]

[0069] In simulation experiment 1, the present invention's method (HoAs-PALN) was used to conduct six comparative experiments with five existing technologies: the greedy method (GREETY), K-winner-take-all (K-WTA), multi-agent deep learning (MADL), multi-agent reinforcement learning (MARL), and multi-agent random self-scheduling (HoAs-RALN). The completion rate of the perception task for each technology in each experiment was obtained. The comparative experiment settings are shown in Table 2.

[0070] Table 2 Comparative experimental group settings

[0071]

[0072]

[0073]

[0074] The experimental results based on these 6 groups of comparative tests are shown in Table 3.

[0075] Table 3 Comparison table of 6 groups of comparative test results

[0076]

[0077]

[0078]

[0079]

[0080] It can be seen from the comparison table of the results of the six groups of comparative tests that the HoAs-PALN method proposed in the present invention shows significant advantages in all groups of experiments.

[0081] 2.2, simulation experiments evaluate the effect of the softmax probability behavior selection adopted in this invention.

[0082] Based on the real datasets (TaxiTrajectories, BicycleOrder, and DidiOrders) and the simulated dataset (Random_1) mentioned in 2.1, the effect of the softmax probabilistic behavior selection adopted in the present invention is evaluated by comparing the HoAs-RALN method based on random behavior selection and the HoAs-PALN method based on softmax probabilistic behavior selection.

[0083] Table 4 Comparison of softmax probability behavior selection effects

[0084]

[0085] As shown in Table 4, the HoAs-PALN proposed in this paper, with its probabilistic behavior selection mechanism guided by a softmax function, significantly outperforms the purely random strategy of the HoAs-RALN in terms of task completion rate, agent decision-making efficiency, and agent movement distance. The agent is more likely to select behavior points closer to the target, reducing ineffective movement distance and dedicating more time to completing the task, thereby improving the overall task completion rate.

[0086] 2.3. Simulation experiments evaluate the effect of the matching dimensionality reduction based on the expected benefits of drones adopted in this invention.

[0087] Based on the real datasets (TaxiTrajectories, BicycleOrder, and DidiOrders) and the simulated dataset (Random_1) mentioned in 2.1, the effect of the matching dimensionality reduction based on the expected return of the drone adopted in this invention is evaluated by comparing the average time of a single decision of the agent before and after dimensionality reduction, as shown in Table 5.

[0088] Table 5 Comparison of the average time of a single decision of the agent before and after dimensionality reduction

[0089]

[0090] As shown in Table 5, the present invention reduces the complexity and computational complexity of the matching process by simplifying the five-dimensional matching process (drone-worker-car-task point-endurance point) to a three-dimensional matching process ("drone-worker-task point" or "drone-car-endurance point"), thereby effectively reducing the average time it takes for each agent to make a single decision. Furthermore, through analysis of the results obtained from the DidiOrders, Random_1, TaxiTrajectories, and BicycleOrder datasets, it can be seen that as the data scale increases and the distribution of task points becomes more dense, the effect of dimensionality reduction in the matching process on reducing the average time it takes for an agent to make a single decision becomes increasingly significant. This shows that this dimensionality reduction method has good application prospects in real-world scenarios with a large number of agents and task points, and can be effectively extended to large-scale scenarios such as those after accidents.

Claims

1. A method for online autonomous collaborative scheduling of intelligent agents based on Nash game and dynamic coupling, characterized in that: Relying on the communication capabilities of the agents themselves to communicate information and match scheduling within a local area, consensus between agents is reached through the local Nash equilibrium game method, and online scheduling is achieved based on the dynamic coupling between agents. The steps of this scheduling method include the following: Step 1: Calculate the expected benefits of each drone going to the task point within the local communication range and cooperating with workers to perform tasks. And the expected benefits of each drone going to a flight point within the local communication range and cooperating with the car to charge to replenish its own power Judging the same drone Is greater than or equal to If yes, the drone is selected to perform the task, and the "drone-worker-task point" matching scheduling is carried out in the subsequent step. Otherwise, the "drone-car-endurance point" matching scheduling is carried out in the subsequent step. Step 2: Using the improved softmax function, the distance between each agent and the behavior point is converted into the probability p of selecting the behavior point. n , in order to make real-time online self-scheduling behavior decisions for dynamically coupled intelligent agents. Each intelligent agent goes online or offline strictly according to its own online and offline time. Each intelligent agent calculates the current behavior benefit based on the matching type. The benefits of drones and workers participating in the perception task are the number of successful perception task matches within their own communication range, and the benefits of drones and cars participating in battery charging are the sum of the power added to all drones within their own communication range. Step 3: Based on the current behavior benefit of each agent, it plays a local Nash equilibrium game with other agents within its own communication range; Step 4: Repeat steps 2 and 3 until each agent reaches a local Nash equilibrium state, that is, all agents reach a consensus, and at the current moment all agents cannot simply adjust their own behavior to increase their benefits, and execute the scheduling behavior of each agent at the current moment.

2. The agent online autonomous collaborative scheduling method according to claim 1, characterized in that: Expected returns as described in step 1 It is obtained by the following formula: Among them, uLeftT i Represents the i-th UAV uav i Remaining online time after executing the task, U_uptime i and U_downtime i Respectively represent UAV i Online time and offline time, ΔuPowSu,T i UAV i The power consumed by executing the task, Fullpower i UAV i Fully charged.

3. The agent online autonomous collaborative scheduling method according to claim 2, characterized in that: Expected returns as described in step 1 Among them, uLeftC i Represents the i-th UAV uav i Remaining online time after charging, ΔuPowSumC i Represents the i-th UAV uav i Charging replenishes the power.

4. The agent online autonomous collaborative scheduling method according to claim 3, characterized in that: The probability p of selecting the behavior point described in step 2 n : Among them, e (·) Represents an exponential function with the natural constant e as the base, score n Indicates the agent's choice of behavior point uwvLoc n The score is the negative value of the distance, that is, score n =-uwvDis n .

5. The agent online autonomous collaborative scheduling method according to claim 4, characterized in that: The number of successful matches in step 2 is: Reward_task ijk =∑matchSign ijx Among them, uLoc i Represents the i-th UAV uav i Location, wLoc j Represents the jth worker j Location, tLoc x Indicates task task x Position, matchSign ijx UAV i 、worker j and task x Match success or failure indicator, Reward_task ijk It represents the current behavior benefit of each drone going to the task point within the local communication range to cooperate with the workers to perform the task.

6. The agent online autonomous collaborative scheduling method according to claim 5, characterized in that: The total amount of electricity in step 2 is: Reward_charge ijk =∑ΔPow ik y Among them, vLoc k Indicates vehicle k Location, chLoc y Indicates the endurance point charge y Position, ΔPow iky Indicates vehicle k Charge at the endurance point y UAV i Increased power, Reward_charge ijk Represents the current behavior benefit of the intelligent agent, which is that the drone goes to a flight point within the local communication range and cooperates with the car to recharge its own power.

7. The agent online autonomous collaborative scheduling method according to claim 6, characterized in that: The local Nash equilibrium game described in step 3 is to check each agent UWV ijk Based on the selected behavior, adjust the self-scheduling behavior again to make it profitable: Reward_task ijk or Reward_charge ijk Is it improved? If so, the benefit mark is: S ijk =0, otherwise, the income is marked as: S ijk =1; The intelligent agent UWV will not be able to be upgraded ijk The benefit identifier of obtaining other agents within the communication range is: neighbors }; If the S of all agents within the communication range neighbors =1, then the agent is considered to be UWV ijk The local Nash equilibrium state has been reached, otherwise, it is considered that the local Nash equilibrium state has not been reached.

Images