Networked multi-agent scheduling computation communication systems, methods, and computer devices

Abstract

The application relates to a network multi-agent scheduling calculation communication system. The system comprises a calculation module and a communication module; in the calculation module, each agent performs calculation on a to-be-executed task through local optimization to obtain a conflict-free task scheduling; when global convergence confirmation is performed, if a conflict is detected, the conflict-free task scheduling is transmitted to other agents through communication; in the communication module, an agent transmits a timetable to other agents, task conflicts generated in the global convergence stage are eliminated by adopting a consistency elimination principle, and a network multi-agent scheduling scheme is obtained. The application can reduce the number of iterations and obviously improve the scheduling efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of agent scheduling, in particular to a network multi-agent scheduling calculation and communication system, method and computer device. BACKGROUND

[0002] As an important application form of multi-agent theory, swarms endow multi-agent with new characteristics such as flexibility, locality and autonomy, which makes it have high value in engineering and can be manifested as a group of vehicles, unmanned aerial vehicles, robots or other intelligent agents with individual dynamics. According to the difference of individual cost and function, the swarm can take group behavior (low cost, single function) or distributed precise scheduling (high cost, complete function). For the former, swarm algorithms based on threshold response can realize behaviors such as aggregation and dispersion. For the latter, multi-agent scheduling algorithms are needed to realize precise coordination such as assembly and scheduling.

[0003] Although scheduling is still a process of matching "agent-task" pairs with time series. Most of the swarms are composed of low-cost agents, which puts higher requirements on calculation, communication and storage, which need to be considered in practical applications. Among these factors, due to the rapid development of chips, small-sized, low-power and high-performance computing units are emerging in an endless stream, such as NVIDIA, AMD or Intel series products, so the contradiction is not so prominent. Although great progress has been made in communication technologies such as self-organizing networks, it is still a more challenging factor due to environmental influences, especially in complex terrain occlusion or electromagnetic environments. The communication between agents is inevitably affected by the environment, resulting in different degrees of errors, packet loss, delay and even interruption, thereby affecting the generation time of scheduling and even success or failure. Therefore, how to reduce the dependence on risky communication and enhance the robustness and timeliness of scheduling generation is a problem that needs to be considered. SUMMARY

[0004] Therefore, it is necessary to provide a network multi-agent scheduling calculation and communication system in view of the above technical problems.

[0005] A network multi-agent scheduling calculation and communication system, the system comprising:

[0006] a calculation module and a communication module;

[0007] In the calculation module, each agent calculates the to-be-executed task through local optimization to obtain a conflict-free task scheduling, and when global convergence confirmation is performed, if it is detected that a conflict exists, the conflict-free task scheduling is transmitted to other agents through a communication mode.

[0008] In the communication module, the agent sends a schedule to other agents, removes task conflicts generated in the global convergence stage by using a consistency elimination principle, and obtains a network multi-agent scheduling scheme.

[0009] In one embodiment, the optimization problem of the agent for task scheduling is:

[0010]

[0011] wherein n represents the number of agents, m represents the number of tasks, x ij is a binary variable, when x ij =1, the agent i performs the task j, when x ij =0, it means that the agent i does not perform the task j, s ij is the score of the agent i performing the task j, and is represented as:

[0012]

[0013] wherein p i represents the schedule of the agent i, represents the reward of the agent i performing the task j according to the schedule p i , λ j <1 is a discount factor for the reward of the task j, is the estimated time for the position of the agent i along the schedule p i or the position of the agent to reach the position of the task j, is the initial task reward of the task j;

[0014] The constraint condition of the optimization problem is:

[0015]

[0016]

[0017]

[0018]

[0019] wherein t ij (p i ) represents the time of the agent i performing the task j in the task scheduling p i .

[0020] In one embodiment, the local optimization includes: task exchange and local sampling; the local range score refers to the score of all agents performing tasks in the local range; the task score of other agents in the local range is obtained through the local sampling, so as to perform local estimation to obtain a conflict-free task scheduling.

[0021] In one embodiment, the task exchange process comprises:

[0022] Setting the task exchange strategy to:

[0023]

[0024]

[0025]

[0026] wherein, is the strategy of agent i to exchange θ, local score S L is the sum of scores of agent i and agent j, θ ij (k,q) is to exchange task k of agent i and task q of agent j, S L (θ ij (k,q)) is the local score after exchange.

[0027] In one embodiment, the local sampling process is: according to the current scheduling, task information, position information and score function of all agents in the local range, constructing a local auction as:

[0028]

[0029] wherein, h i (·) represents the local estimation function of agent i to other agents in the local range, p k (t) is the local auction function of other agent k at current time t, q j represents the task information of task j, l k represents the position information of agent k.

[0030] A network multi-agent scheduling calculation communication method, the method comprises:

[0031] Each agent performs calculation on the to-be-executed task through local optimization to obtain a conflict-free task scheduling, and when global convergence confirmation is performed, if it is detected that a conflict exists, the conflict-free task scheduling is transmitted to other agents through communication.

[0032] The agent sends a timetable to other agents, adopts a consistency elimination principle to eliminate task conflicts generated in the global convergence stage, and obtains a network multi-agent scheduling scheme.

[0033] A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that the processor implements the steps of the method when executing the computer program:

[0034] Each intelligent agent calculates the task to be executed through local optimization to obtain a conflict-free task schedule. During the global convergence confirmation, if a conflict is detected, the conflict-free task schedule is transmitted to other intelligent agents through communication.

[0035] The agent sends a schedule to other agents, and the consistency elimination principle is used to eliminate task conflicts generated during the global convergence phase, thus obtaining a network multi-agent scheduling scheme.

[0036] The aforementioned network multi-agent scheduling computation-for-communication system, method, and computer equipment propose the idea of ​​computation-for-communication, which reduces the number of communications by appropriately increasing computational load, thereby maintaining or even reducing the schedule generation time. A novel scoring function is designed, which is more optimized and can support the inference of the positions of other agents. A local optimization method is proposed to further improve the optimization and timeliness of scheduling. Better scheduling methods are explored using task swapping and task sampling to avoid local optima problems caused by individual greed. Conflicts between agents are resolved in advance through local estimation, reducing the number of iterations. Simultaneously, the proposed task-related agent selection strategy is adopted to allocate computational resources to key agents. Attached Figure Description

[0037] Figure 1 This is a comparison diagram showing the differences between the communication computing framework and the traditional market framework in one embodiment;

[0038] Figure 2 This is a schematic diagram of a network multi-agent scheduling computation and communication system in one embodiment;

[0039] Figure 3 This is a schematic diagram of the processing flow for calculating the frame replacement in one embodiment. Detailed Implementation

[0040] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0041] like Figure 1 As shown, Figure 1 On the left is a comparison between the traditional market-based framework and... Figure 1 The right side shows the computational communication framework of this invention. From Figure 1The traditional market-based framework on the left can be seen that the total scheduling time includes planning time, communication time and waiting time, and the planning time, communication time and waiting time include calculation time and communication time. In the calculation and communication framework of the present application, due to the introduction of the calculation and communication conflict resolution strategy, although the overall incremental planning time increases, the communication time is greatly reduced, and the scheduling efficiency is improved.

[0042] In one embodiment, as shown in Figure 2 a network multi-agent scheduling calculation and communication system is provided, as follows:

[0043] The calculation module 100 and the communication module 200.

[0044] In the calculation module 100, each agent calculates the to-be-executed task through local optimization to obtain a conflict-free task scheduling. When global convergence confirmation is performed, if a conflict is detected, the conflict-free task scheduling is transmitted to other agents through communication.

[0045] In the communication module 200, the agent sends a schedule to other agents, adopts a consistency elimination principle to eliminate task conflicts generated in the global convergence stage, and obtains a network multi-agent scheduling scheme.

[0046] Specifically,

[0047] In the above network multi-agent scheduling calculation and communication system, the idea of calculation and communication is proposed, the number of communications is reduced by appropriately increasing the amount of calculation, so as to maintain or even reduce the scheduling generation time. A new scoring function is designed, which is more optimized and can provide support for inferring the positions of other agents. A local optimization method is proposed to further improve the optimization and timeliness of the scheduling. The task exchange and task sampling method is used to explore a better scheduling method to avoid the local optimal problem caused by individual greed. The conflict between agents is solved in advance through local estimation, the number of iterations is reduced, and the proposed task-related agent selection strategy is used to spend the calculation amount on key agents.

[0048] For the scheduling calculation of network multi-agent, it is generally divided into two stages, namely the calculation stage and the communication stage. In the calculation stage, the agents all pursue the common goal of score maximization, but due to the limited ability of the agents, the scores of the agents all have upper bounds. In the communication stage, each agent all complies with the consistency rule of the highest bidder, that is, the task belongs to the agent with the highest score, and the task conflict can be eliminated by the score. Therefore, as long as there is no deadlock problem caused by task score oscillation, the market-based method can certainly converge. The reasons affecting the convergence speed mainly include three aspects: first, the number of task conflicts, which is related to the distribution of agents or tasks and the performance of the algorithm, for example, individual greedy algorithm is prone to task conflicts; second, the monotonicity of task score, if the score is not monotonic, the preferred task of the agent may be lost due to low score, and it is necessary to reselect the task for iteration, and even the task selection cycle may be trapped in a stalemate; third, the update of task score, after the task is deleted or added, once the task score is updated, the attributes of the task may change, and it is necessary to iterate again to eliminate the task conflict.

[0049] In summary, the core of improving the scheduling efficiency of network multi-agent lies in how to accelerate the convergence, that is, to reduce the number of iterations. Therefore, in the present application, the calculation amount of other agents is increased to reduce the task conflict, thereby reducing the iteration, although the calculation time is prolonged, the number of iterations is reduced, thereby reducing the communication round, and the increase of the calculation amount is mainly due to the local optimization including other agents, which is expected to reduce the number of communications and keep the total scheduling generation time unchanged or even reduced.

[0050] In one embodiment, the calculation-communication system of the present application, similar to the traditional calculation, retains two iteration stages. In the calculation stage, the agent removes the tasks not belonging to itself according to the attributes of the tasks after the conflict is solved, and then calculates the conflict-free tasks by local optimization and adds them to the agent scheduling. Next, the convergence is confirmed, if there is a conflict, the scheduling is transmitted to other agents, and the conflict is solved in the communication stage, otherwise the final scheduling is obtained. In the communication stage, the agent sends its schedule to other agents in the form of task winner, task winning bid and time stamp, and receives the schedule of other agents, and then eliminates the task conflict according to the consistency rule. A commonly used consistency rule is that the higher the price, the higher the agent wins the task. For example, as shown in FIG. 1, the agent i (iteration N times) first removes the tasks to obtain the locally optimal task execution scheme, and then performs the inclusion stage to check the convergence, if the condition is met, the final solution is output, if not, as the sender, the schedule of the agent itself needs to be sent to other agents, and the global task conflict is eliminated, and then the eliminated tasks are sent to the agent i again for N+1 times of iteration. Figure 3 ​

[0051] In one embodiment, the optimization problem for task scheduling by the agent is:

[0052]

[0053] where n represents the number of agents, m represents the number of tasks, x ij is a binary variable, when x ij = 1, agent i performs task j, when x ij = 0, it means that agent i does not perform task j, s ij is the score of agent i performing task j, which is represented as:

[0054]

[0055] where p i represents the schedule of agent i, represents the reward of agent i performing task j according to schedule p i , λ j < 1 is the discount factor for the reward of task j, is the estimated time for agent i to reach the location of task j along the schedule p i or the position of the agent, is the initial task reward of task j;

[0056] The constraint conditions of the optimization problem are:

[0057]

[0058]

[0059]

[0060]

[0061] where t ij (p i ) represents the time of agent i performing task j in task scheduling p i .

[0062] Specifically, the multi-agent scheduling is to allocate m tasks to n agents to form their respective schedules. The schedule pi of each agent i is the execution order and time of the allocated tasks, which satisfies the constraints such as the start time and deadline of the tasks. It can be generally expressed as a constrained optimization problem. In the above optimization problem, the parameter s ijThe improvement is made to introduce the influence of task j on the schedule of agent i, which drives the agent to explore new tasks based on the previous task, regardless of whether the new task is near other agents, which can lead to blind outward exploration. Unlike the common scoring function, the latter is introduced to drive the agent to select tasks near its location, so that the agent tries to avoid competing with other agents near the task, which helps to reduce conflicts between agents. Each agent tries to select a task that is closer to itself, which can avoid the situation where agents near the task wait for other agents far away to perform time-limited tasks, resulting in task failure. In addition, the introduced The distance information between the agent and the task is also implied.

[0063] In one embodiment, the local optimization includes: task exchange and local sampling, the local range score is increased by task exchange; the local range score refers to the score of all agents performing tasks in the local range; the task score of other agents in the local range is obtained by local sampling, so as to obtain a conflict-free task scheduling by local estimation.

[0064] Specifically, the local optimization means that the agent considers the scheduling of other agents in the local range, rather than only considering its own scheduling like individual optimization. This allows the agent to avoid possible conflicts with other agents in the local range in advance, and tries to escape the local optimum into which the individual optimization falls. The local optimization mainly includes two steps: the first step is task exchange, which improves the performance of the scheduling by exchanging the assigned tasks between agents; the second step is local sampling, which samples various feasible schedulings to select the optimal scheduling, further improving the scheduling performance. In these two steps, since the scheduling of other agents is calculated, the task conflict will be eliminated.

[0065] In one embodiment, the task exchange process includes:

[0066] The task exchange strategy is set as:

[0067]

[0068]

[0069]

[0070] wherein, is the strategy of agent i to exchange θ, the local score S L is the sum of the scores of agent i and agent j, θ ij (k, q) is to exchange task k of agent i with task q of agent j, S L (θ ij (k, q)) is the local score after exchange.

[0071] Specifically, in the task exchange step, the local scope is the two agents exchanging tasks, and in the local sampling, the local scope is the multiple agents related to the task. Therefore, the performance of the local optimization is evaluated by the local score, i.e., the sum of the scores of all agents in the specified local scope. The local score S L is:

[0072]

[0073] where L is all agents in the local scope, which is different in the task exchange and local sampling stages.

[0074] For the task exchange process, agent i attempts to exchange tasks with other agent j to improve the local optimization, which can be reflected by the local score of the two agents. Since there can be many feasible exchanges, the optimal exchange is selected that increases the local score the most.

[0075] Specifically, since a large number of traversals are required for task exchange, a heuristic algorithm can be used to solve the above task exchange strategy to improve computational efficiency.

[0076] In one embodiment, the process of local sampling is:

[0077] According to the current scheduling of all agents in the local scope, task information, location information, and scoring function, a local auction is constructed as:

[0078]

[0079] where h i (·) represents the local estimation function of agent i on other agents in the local scope, p k (t) is the local auction function of other agent k at the current time t, q j represents the task information of task j, l k represents the location information of agent k.

[0080] Specifically, in each sampling, agent i calculates the task score of all related agents j, estimates their task selection and result progress, thereby avoiding task conflicts with other agents in advance and reducing the required communication between agents. This process of estimating the scheduling of other agents is the local estimation h, which is mainly based on the current scheduling of all agents k in the local scope, the information of all tasks j q j , the location of agent k l k , and the scoring function s ij to construct a local auction.

[0081] In one of the embodiments, a network multi-agent scheduling calculation communication method is further provided, and the specific steps are as follows:

[0082] Each agent performs calculation on the to-be-executed tasks through local optimization to obtain a conflict-free task scheduling, and when global convergence confirmation is performed, if a conflict is detected, the conflict-free task scheduling is transmitted to other agents through communication.

[0083] The agent transmits a schedule to other agents, eliminates the task conflict generated in the global convergence stage by using a consistency elimination principle, and obtains a network multi-agent scheduling scheme.

[0084] In one of the embodiments, a computer device is provided, which includes a memory and a processor, the memory stores a computer program, and the processor implements the steps of the method in the above-mentioned embodiments when executing the computer program.

[0085] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be completed by a computer program instructing related hardware, and the computer program can be stored in a non-volatile computer readable storage medium. When the computer program is executed, the processes of the above-mentioned embodiments can be included. Any reference to the memory, storage, database or other medium used in each embodiment provided in the present application can include non-volatile and / or volatile memory. The non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. The volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM) and memory bus dynamic RAM (RDRAM) and the like.

[0086] Each technical feature of the above embodiments can be combined arbitrarily, and for the sake of brevity, all possible combinations of the technical features in the above embodiments are not described, however, as long as the combination of the technical features does not exist, it should be considered as the scope of the present application.

[0087] The above-described embodiments are merely illustrative of several embodiments of the present application, which are described in more detail and in a specific manner, but should not be construed as limiting the scope of the patent. It should be noted that for those skilled in the art, several modifications and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application should be subject to the appended claims.

Claims

1. A network multi-agent scheduling computing and communication system, characterized in that, The system includes: Computing module and communication module; In the computing module, each agent calculates the task to be executed through local optimization to obtain a conflict-free task schedule. During the global convergence confirmation, if a conflict is detected, the conflict-free task schedule is transmitted to other agents through communication. In the communication module, agents send schedules to other agents, and the consistency elimination principle is used to eliminate task conflicts generated during the global convergence phase, resulting in a network multi-agent scheduling scheme. The schedule is sent during the communication phase, where agents send their own schedules to other agents in the form of task winner, task winning price, and timestamp, and receive schedules from other agents. The consistency elimination principle is the highest bidder wins. The optimization problem for task scheduling by the intelligent agent is as follows: ； Where n represents the number of agents and m represents the number of tasks. For binary variables, when At that time, intelligent agent Execute the task ,when When, it indicates an intelligent agent Do not perform task , It is an intelligent agent Execute the task The score is expressed as: ； in, Represents intelligent agents The schedule Represents intelligent agents According to the timetable Execute the task The return This is the mission Discount factor of returns For intelligent agents According to the timetable Or the location of the agent reaches the task Estimated time of location For the task The initial task reward; The constraints of the optimization problem are: ； ； ； ； in, Represents intelligent agents In task scheduling Performing tasks Time; The local optimization includes: task switching and local sampling; The task exchange increases the local scope score; the local scope score refers to the score of all agents performing tasks within the local scope. The task scores of other agents within the local area are obtained through the local sampling, thereby performing local estimation to obtain conflict-free task scheduling. The task exchange process includes: Set the task switching strategy as follows: ； ； ； in, It is an intelligent agent Choosing a strategy of swapping θ, local scores It is an intelligent agent With intelligent agents The sum of fractions, It is to make intelligent agents Task With intelligent agents Task exchange, These are the local fractions after the exchange; The local sampling process is as follows: Based on the current scheduling, task information, location information, and scoring function of all agents within the local area, the local auction is constructed as follows: ； in, This represents the local estimation function of agent i for other agents within a local range. The local auction function of other agents k at time t. Indicates task Task information, This represents the position information of agent k.

2. A network multi-agent scheduling computation-communication method, executed in the network multi-agent scheduling computation-communication system of claim 1, characterized in that, The method includes: Each intelligent agent calculates the task to be executed through local optimization to obtain a conflict-free task schedule. During the global convergence confirmation, if a conflict is detected, the conflict-free task schedule is transmitted to other intelligent agents through communication. The agent sends a schedule to other agents, and the consistency elimination principle is used to eliminate task conflicts generated during the global convergence phase, thus obtaining a network multi-agent scheduling scheme.

3. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method of claim 2.

Images