KR101081932B1 - Apparatus and method of load dispersion in multi-agent system - Google Patents

Abstract

The present invention provides at least one agent that is generated in a host to process a user's request task according to an operation state, an agent platform unit for registering each agent that is generated and operated on a plurality of nodes and managing an operation state; Initiating a multi-agent system comprising a monitor agent unit for calculating a weight according to the operating state of each agent identified through the platform unit and assigning the resource according to the weight to each agent, thereby actively and efficiently limited resources to the agent It can be assigned, and the agent can be migrated based on the operating state of the agent to improve the performance of the system.

Description

Apparatus and method of load dispersion in multi-agent system

The present invention relates to a load balancing method and apparatus thereof of a multi-agent system, and more particularly, to allocate a limited resource actively and efficiently in a multi-agent system in which a plurality of agents exist, and based on an operation state of an agent. The present invention relates to a load balancing method and a device of a multi-agent system capable of migrating agents to improve system performance.

With the development and spread of network technologies such as the Internet, and the development of various application services, various applications have been developed to provide services required by consumers in various ways.

As the ubiquitous environment becomes more common, things or computers work intelligently on behalf of users to provide any service to consumers and users at any time, anywhere, and any device. The interest in the aspect of providing a service that can satisfy the user's needs is increasing.

The agent technology to handle such work performance has features such as intelligence, autonomy, reactivity and sociality so that the user can perform the desired service and interact with the external environment. While recognizing the change of situation, it is being developed to provide complex application service through cooperation with other agents or systems.

For example, a myriad of information on the Internet can be used by agents to find or filter information on your behalf, sorting, searching, and deleting e-mails that you have manually processed, or in online e-commerce. As an agent of a user, you can enter a bid and compete with other agents in the bid to get a product at a lower price.

In addition, a system capable of performing work through cooperation among several agents in a distributed environment is called a multi-agent system, and a federation service of agents performing work through cooperation between several agents may be provided.

Such a multi-agent system will be built by extending the standard model of the agent platform part proposed by FIPA (Foundation for Intelligent Physical Agent).

1 is a view for explaining a standard model of the Airgent platform proposed by FIPA.

Referring to FIG. 1, the agent platform unit manages a plurality of agents 2 performing various services and overall agent life cycles such as creation, registration, removal, pause, and recovery of agents in the agent platform unit. Agent Management System (AMS) (3), and DF (Directory Facilitator) (4) that processes the service information of each agent to provide each agent with information about the capabilities or services provided by agents in the agent platform unit; And a message transport system (MTS) 5 for communicating information or messages with other agent platform units.

Agents are performed on behalf of a user's desired service. Unlike an existing application, an agent has characteristics such as knowledge, reasoning ability, learning, and planning ability to interact with the external environment to solve the user's desired work on behalf of the user. It refers to a system or a subsystem that recognizes a change in a situation according to an action and solves a problem in a goal-oriented manner by cooperating with another system or other agents.

In addition, a large number of agents (2) are required to process complex application services, and the agents (2) are efficiently managed and bundled and operated in one package to exchange data or messages with other agents. It is necessary. The area in which the group of agents 2 providing similar application services work is called an agent platform unit.

The agent platform unit has autonomy, knowledge base, and inference ability to actively perform tasks by judging or operating on its own without direct instruction or interference with users or other agents. Intelligence to grasp new intentions through planning or learning by knowing the intention of the user, and to actually perform the work without the user's desired task in the current host or node. Mobility, which improves efficiency and reduces network load by moving to a processing node, and sociability of exchanging messages between agents when one agent needs help to perform tasks that one agent cannot handle. (Social Ability) and the like.

Meanwhile, a domain refers to a group in which a plurality of agent platform units are gathered to perform a single space or complex application service, and constitute a multi-agent system. The multi-agent system includes a main agent platform unit that manages agent platform units of a corresponding domain and a sub agent platform unit that performs each application service.

The characteristics of the multi-agent system seek to exchange and share information through agent-to-agent cooperation, to maximize the performance and utilization of agents, and to cooperate with other agent platform departments for complex application services that cannot be solved by independent application sets. When a new application service is needed, a new agent is added to the system to provide the service.

2 is a diagram for explaining a multi-agent system.

Referring to FIG. 2, in the multi-agent system, a plurality of agent platform units 10, 20, and 30 exchange information or messages with other agent platform units through MTSs 15, 25, and 35. 10, 20, 30) to provide a complex application service that can not handle.

At this time, the information exchange operation is performed by using an Agent Communication Language (ACL) message structure, which is a communication language between the agent platform units 10, 20, and 30.

Currently, various multi-agent systems such as 'Telescript', 'Aglets', 'Jade', 'Madkit', 'Agent Gateway', and 'Mobile Spaces' have been developed and proposed. Simple scheduling algorithms such as sharing are applied.

In addition, the multi-agent system based on the 'Max-min' algorithm used in the grid computing environment has been proposed a scheduling algorithm based on the multi-agent-oriented priority, the 'Max-min' algorithm for fairness and efficiency Priority is given to agents with fewer clone agents and smaller computational needs.

However, such an algorithm does not take into account the agent life cycle that represents the state of the agent, which can cause waste of resources by allocating CPU resources to agents that are not running.

For example, if multiple agents exist on a node, it can be assumed that the execution of the agent is scheduled by process sharing, in which the newly created agent is already running on the same node. This affects preoccupied resources. Therefore, the newly created agent incurs an additional time overhead in reducing node performance and completing each agent's work.

Meanwhile, the Earliest Deadline First (EDF) method is used as a scheduling method suitable for time-constrained applications during agent scheduling. In the EDF method, all resources are allocated to the agent having the shortest deadline.

However, EDF is not suitable for multi-agent systems because the allocation of resources is not fair. For example, suppose that two agents with nearly the same deadline time are running on one node in the EDF method, all resources will be allocated only to the agent with the shortest deadline, and no resources will be allocated to the other agent.

Thus, a problem arises that agents with some long deadlines are unable to complete their work within the deadlines.

Therefore, a method for efficiently and actively allocating resources in a multi-agent system should be proposed.

The present invention has been proposed to solve the above problems, and provides a method and apparatus for load balancing of a multi-agent system capable of allocating limited resources actively and efficiently in a multi-agent system having a plurality of agents. There is a purpose.

In addition, the present invention provides a method and apparatus for load balancing a multi-agent system capable of improving system performance by effectively allocating an agent by migrating an agent based on an operation state of an agent present on a plurality of nodes. The purpose is.

Multi-agent system according to an aspect of the present invention, at least one agent that is generated in the host to process the user's request task according to the operation state, and each agent is created and operated on a plurality of nodes, the operation state And a monitor agent unit configured to calculate a weight according to an operation state of each agent identified through the agent platform unit, and to allocate resources according to the weight to each agent.

The monitor agent unit calculates the weight based on the number of agents in an activated state among the agents on each node, and when the round trip time (RTT) value for each agent is equal to or smaller than the average RTT value of the node, the average RTT It is preferable to perform load balancing by migrating one of the agents belonging to a node having a small value to a node having a high average RTT value.

The monitor agent unit may repeatedly migrate the agent of the node having a low weight average until the weight average of each node is equal to the weight average of all nodes.

The multi-agent system according to another aspect of the present invention, the agent agent to handle the load balancing process by requesting the agent migration to the agent management system (AMS) agent, and the monitor agent unit calculates the weight by defining the operation status of the generated agent Agent Life Cycle, and the Scheduler of Behaviors for requesting resource allocation to each agent based on the weight calculated by the monitor agent.

In the load balancing method of a multi-agent system according to another aspect of the present invention, an agent platform unit registers an agent generated to work according to a user's request and manages an operation state, and a monitor agent unit performs an average RTT of each node. Determining load balancing based on the value and the RTT value of each agent, sorting the RTT values of each node in ascending order, and migrating one agent of the first node to the last node, at each node. Calculating a weight based on the number of agents in an active state, allocating resources to each agent based on the amount of resources allocated to the node and the weight, and based on the weight of each agent and a weighted average of the nodes Determining load balancing, the weights, and each resource according to the amount of resources of the node. Allocating resources to the agent.

In the load balancing method of the multi-agent system, the monitor agent unit recalculates the weights when an operation state of at least one agent transitions or an agent is generated / terminated on a node, and the weighted average of each node. And repeatedly migrating agents of the node whose weight average is low until it is equal to the weight average of all the nodes.

According to the present invention as described above, in a multi-agent system in which a plurality of agents exist, it is possible to allocate active resources efficiently and efficiently. Can improve.

Therefore, the average round trip time (RTT) of messages exchanged between agents varies according to the state of the agent, and when the state change rate of the agent is relatively high, the average RTT increases, but according to the present invention, it is calculated through the monitor agent unit. By allocating CPU resources to agents based on weights, we can see that the performance of the system improves when the agent periodically changes its status and sends messages.

DETAILED DESCRIPTION Hereinafter, a load balancing method and apparatus thereof of a multi-agent system according to a preferred embodiment of the present invention will be described in detail with reference to the accompanying drawings. .

3 is a diagram illustrating a multi-agent system according to the present invention.

Referring to FIG. 3, a monitor agent unit for requesting resource allocation by calculating a weight of each agent A based on an agent A operating in a plurality of nodes, an agent platform unit, and an operating state of the agent A (MAg).

The agent A processes a job generated and requested according to a user's request to provide a service, and the agent platform unit handles registration of the agent A, message exchange, and operation state management.

The monitor agent MAg requests resource allocation based on the weight after calculating the weight according to the operation state of the agent A for each node identified through the agent platform.

4 is a diagram for explaining an operation state of an agent.

Referring to FIG. 4, the operating state of the agent A, that is, the life cycle, may be largely classified into initiated, active, suspended, waiting, and transit.

That is, after the agent A is created and activated by an invoke in the initiation state, the operation A is started by the user's request, and the operation state is changed by the operation.

Since the technical contents to which the operation state of the agent A changes are well known contents, detailed description thereof will be omitted.

The monitor agent unit MAg measures the message load of each agent A exchanged through the agent platform unit.

In addition, the monitor agent MAg calculates a weight based on the operation state of each agent A so that an efficient resource is allocated to each agent A through the agent platform unit.

 In this case, the monitor agent MAg may determine whether the agent A is in an activated state or inactivated state rather than five operating states managed by the agent platform unit in order to efficiently calculate a weight based on the operating state of each agent A. FIG. The weight is calculated based on whether or not.

The monitor agent MAg determines the amount of resources (ie, CPU resource amount) required by the agent A based on the weight, and the multi-agent system determines the amount of resources determined by the monitor agent MAg. Assign to the agent (A).

That is, the multi-agent system load balances resources by migrating the agent A for efficient allocation of resources through the amount of resources calculated based on the weight of each agent A calculated by the monitor agent unit MAg. Let this be done.

First, the monitor agent unit MAg measures the load according to the resources allocated to each agent A. FIG.

The monitor agent MAg distributes the load through an agent migration determined based on a round trip time (RTT) determined by the load of each agent A and a weight of the agent A.

At this time, the weight of each agent A measured by the monitor agent unit MAg is determined according to the number of agents A running in a specific node.

Referring to FIG. 3, the monitor agent unit MAg measures the load of the agent A (Agik) existing in the node Nodek to which the resource Pk is allocated. In other words, the monitor agent MAg measures the RTT for determining the overload and calculates a weight for equally allocating resources. When the operating state of the agent A transitions or when a new agent A is created / terminated on the node, the weight is recalculated to reallocate resources.

All agent A (Agik) and monitor agent unit MAg on the node are registered with the agent platform unit (e.g., AP). The resource amount of the node Nodek represents the number of instructions that can be executed per unit time in the node.

The monitor agent MAg measures the RTT as shown in Equation 1 based on the time received after periodically transmitting an acknowledgment (ACK) message to each agent (A).

는 각 에이전트(A)의 RTT 값이고,

는 모니터 에이전트부(MAg)가 각 에이전트(A)로 확인 메시지를 전송한 송신 시간 값이고,

는 각 에이전트(A)로부터 메시지가 수신되는 수신 시간 값이다.here

Is the RTT value of each agent (A),

Is a transmission time value at which the monitor agent unit MAg transmits a confirmation message to each agent A,

Is a reception time value at which a message is received from each agent (A).

The monitor agent MAg calculates an average RTT value of each node Nodek as shown in Equation 2 below.

는 노드의 평균 RTT 값이고,

는 해당 노드의 각 에이전트(A)별 RTT 값이고, 'N'은 해당 노드의 에이전트(A) 개수이다.here

Is the average RTT value of the node,

Is the RTT value for each agent (A) of the node, and 'N' is the number of agents (A) of the node.

와,

를 비교한 이후에 부하 분산을 결정한다.The monitor agent MAg compares the average RTT value of each node with the RTT value of the corresponding agent A, that is,

Wow,

After comparing the load distributions are determined.

For example, the load A may be distributed by migrating the agent A having a large RTT value to a node having a relatively small average RTT value.

On the other hand, the monitor agent MAg calculates a weight according to the operation state of each agent A included in each node.

At this time, the monitor agent MAg calculates the weight by dividing the operation state of each agent A into an activated state or an inactive state. In other words, the monitor agent MAg assigns a weight only to the agent A that is in an activated state.

The monitor agent MAg calculates the weight of each agent A as shown in Equation 4 below.

는 에이전트(A)의 가중치이고,

는 해당 노드에 활성화 상태인 에이전트(A)의 개수이다.here

Is the weight of agent A,

Is the number of agents A that are active in the node.

In addition, when the operation state of the agent A is changed or the agent A is generated / terminated on the node, the monitor agent unit MAg reallocates resources according to the recalculated weight after recalculating the weight. do.

Hereinafter, a process in which the monitor agent MAg calculates a weight and allocates resources in a multi-agent system will be described.

5 is a view for explaining the resource allocation according to the present invention in more detail.

FIG. 5 illustrates a situation where a plurality of agents A are generated in each node of a multi-agent system, in which an agent A (blue in the drawing) and an agent A (white in the drawing) in an active state coexist. For example, the operation state of the agent A is changed or generated.

In the first node, three agents A are activated, the resource amount is 0.6, and in the second node, two of the four agents A are inactive, two are activated, and the resource amount is 0.8.

In the third node, there is one agent A in the activated state, the agent A in one inactive state, the resource amount is 1.4, the agent A in the two active states in the fourth node, There is one inactive agent A, and the resource amount is 0.9.

The monitor agent MAg calculates the weight of each agent A based on Equation 4, as shown in Table 1 below.

As described in Table 1 above, the monitor agent MAg according to the present invention calculates a weight of a corresponding node based on the number of agents A in an active state, and allocates resources to each agent A. FIG. .

For example, since there is only one agent A in the third node, the monitor agent MAg according to the present invention has a weight of '1', so that 1.4, the resource amount of the third node, is in the active state. When the resource A is assigned to the agent A, but the resources are allocated to the agent A by an existing method, for example, a processor sharing method, the operating state of the agent A is not taken into account, so that the agent A also has an inactive state. By allocating resources, 0.7 resources are allocated to the agent A in the activated state.

Therefore, the agent A that is active may be allocated more resources than when the resource A is allocated to the agent A by the existing processor sharing method, and the number of the agent A that is inactive on the node increases. As a result, resources can be allocated to the agent A that is activated more efficiently.

On the other hand, as illustrated in FIG. 5, when the operating state of the agent A changes or when the agent A is generated / terminated, the weight of the agent A is re-allocated to reallocate the resource of each agent A. FIG.

The monitor agent unit MAg re-allocates resources by recalculating weights in real time according to changes in the operation state and the number of agents A managed by the agent platform unit.

Table 2 below is for explaining that the monitor agent (MAg) reallocated weights and resources in real time as illustrated in FIG.

As described in Table 2, when the operating state of the agent A is changed or when the agent A is generated / terminated, the monitor agent unit MAg recalculates the weight in real time to each agent A. Reallocate resources.

That is, when the transition and generation / end of the operation state of the agent A occurs and the monitor agent unit MAg re-allocates resources by recalculating weights in real time, it can be seen that resource allocation is efficiently performed. .

Meanwhile, in a multi-agent system, when a message overload occurs in a specific agent A or when the agent A in an active state is concentrated in a specific node, that is, when an overload occurs, the load should be distributed.

The monitor agent MAg determines load balancing based on the average RTT value () of each node and the RTT value () for each agent (A) of the node.

The monitor agent MAg determines load balancing as shown in Equation 5 below.

That is, the monitor agent MAg loads when the RTT value of each agent A is equal to or smaller than the average RTT value of the node, that is, when the message A is overloaded or the node is overloaded. Determine the variance.

At this time, the monitor agent unit MAg sorts the RTT values Ck (t) of the nodes in ascending order and instructs the migration of one agent A of the first node to the last node in the sorted state. That is, the monitor agent unit MAg migrates one agent A among the agents A belonging to a node having a small average RTT value to a node having a large average RTT value so that the load is distributed.

On the other hand, when the agent A in an active state is concentrated in a specific node, the monitor agent MAg calculates an average of weights assigned to the agent A of the corresponding node as shown in Equation 6 below.

'W' is the weight of the agent (A) belonging to the node, 'Q' is the average of the weight.

In addition, since the weight assigned to the agent A of each node is calculated by the number of agents A in the activated state as shown in Equation 4, the weight W and the average Q of the weights have the same value.

The monitor agent MAg determines whether there is a resource to be allocated to the node based on the average of the calculated weights. In other words, the monitor agent MAg sorts the average of the weights in ascending order.

The monitor agent unit MAg has a large weight of the agent A, so that there are fewer active agents A running on the node to which the agent A belongs, so that the agent A has more resources in the node. I think it can be assigned to.

When the weighted average of one node is smaller than the weighted average of all nodes, the monitor agent MAG migrates the agent A of the node to the node having the largest weighted average in the multi-agent system, and load balances them. The process is repeated until the weighted averages of all nodes are equal (similar), as shown in Equation 6 below.

Tables 3a and 3b below are for explaining the load balancing process.

Table 3a shows that when the monitor agent MAg sorts the weighted average of each node, and the weighted average of the first node is less than the averaged weighted average of all nodes, the agent A of the first node is transferred to another node. Determine the load balancing to migrate and migrate to the node with the large weighted average. Therefore, as shown in Table 3b, when the agent A of the first node migrates to the third node to distribute the load, it can be seen that the weights of all nodes are equal and the load is distributed.

6 is a view for briefly explaining an agent migration process in a multi-agent system of the present invention.

Referring to FIG. 6, the multi-agent system according to the present invention illustrates a case of combining with a mobile agent (A) system to support agent migration, and the agent (A) in the multi-agent system is transferred from one node to another node. You can migrate.

At this time, the destination node to which the agent A will migrate includes the agent A in one or more activated states, and the multi-agent system performs agent migration for the purpose of load balancing or fault tolerance.

The migration request agent (A), which needs to migrate the node, sends a request message to the agent platform unit, that is, the AMS agent (A), and the AMS agent (A) sends a time stamp and authentication message to the migration request agent (A). do.

The migration request agent (A) sends a migration request message to the dynamic library agent (A) of the destination node with an authentication message.

The dynamic library agent A receives the agent A code after verifying the authentication and validity of the authentication message.

The migration request agent (A) moves itself to the destination node or moves the clone agent (A), and the migrated agent (A) is executed by the dynamic library agent (A). At this time, when the clone agent A moves, it registers with the agent platform.

7 is a view for explaining the structure of the agent platform unit.

In FIG. 7, the agent platform unit includes the "Hierarchical P2P Networking Two-Level Compression Scheme for Multi-agent System Supporting Context-Aware Applications," published in November 2008, and "A New Agent Platform Architecture Supporting the published in November 2007". Agent Group Paradigm for Multi-Agent Systems, "and the like, so detailed description thereof will be omitted.

The agent platform unit consists of one agent platform core and a message transfer protocol (MTP).

The agent platform core consists of a main container, an agent (A) container, and an agent (A) description table. The message transfer protocol supports event-based communication.

The agent platform unit includes reference information of the agent A and functions used for managing the internal agent A. FIG.

Then, all ACL messages of the multi-agent system are exchanged through the agent platform unit, and the agent platform unit receiving the ACL message transmits to the destination agent A using the registered reference information. The AMS and the DF are registered with the agent platform as the agent A, respectively.

That is, the agent platform unit according to the present invention implements the agent platform unit core module as the agent (A) to have a higher robustness and expandability.

The agent platform core uses an MTS library for communication and consists of one main container and several agent (A) containers. The agent A container manages an agent A description having information on connection and management of the agent A, such as state information for life cycle management.

8 is a view for explaining the structure of a multi-agent system according to the present invention.

Referring to FIG. 8, the multi-agent system according to the present invention provides core functions for the creation and operation of the agent A as a dynamic library.

In addition, inter-agent A communication ensures interoperability using MTP compliant with HTTP (hypertext transport protocol) standard and ACL for integrity of message.

Agent migration processes the load balancing process by requesting the migration of agent A to AMS agent A during load balancing process, and the Agent Life Cycle allows the monitor agent unit (MAg) to calculate the weight to determine resource allocation. Define the operating state of the agent (A).

The Scheduler of Behaviors requests to allocate the resource according to the weight to the agent A based on the weight calculated by the monitor agent MAg.
The multi-agent system of the present invention requests an agent migration to an agent management system (AMS) agent so as to calculate a weight by defining an agent migration unit for processing a load balancing process and an operation state of the generated agent. And an agent activity scheduler for requesting resource allocation to each agent based on the agent life cycle information unit and the weight calculated by the monitor agent unit.

Descriptions of other unexplained modules are well known descriptions, and thus detailed descriptions thereof will be omitted.

9 is a flowchart illustrating a load balancing method of a multi-agent system according to the present invention.

Referring to FIG. 9, the agent platform unit registers an agent A generated to work according to a user's request from a host (for example, a personal computer) and manages an operation state (S100).

In other words, the agent platform unit manages generation / end and operation states of the agent A for each node.

The monitor agent MAg calculates an RTT based on a time received after periodically sending an acknowledgment (ACK) message to each agent (A), thereby calculating an average RTT value of each node and an RTT for each agent (A). The load balancing is determined based on the value (S110).

The monitor agent MAg calculates a weight based on the number of agents A that are activated at each node (S120).

On the other hand, the monitor agent unit MAg recalculates the weight when the operating state of the agent A changes or when the agent A is generated / terminated on the node.

The monitor agent unit MAg allocates a resource to each agent A based on the resource amount and the weight allocated to the node (S130).

The monitor agent MAg determines load balancing based on the average RTT value of each node and the RTT value of each agent A of the node.

That is, the monitor agent MAg loads when the RTT value of each agent A is equal to or smaller than the average RTT value of the node, that is, when the message A is overloaded or the node is overloaded. Determine the variance.

At this time, the monitor agent unit MAg sorts the RTT values Ck (t) of the nodes in ascending order and instructs the migration of one agent A of the first node to the last node in the sorted state. That is, the monitor agent unit MAg migrates one agent A among the agents A belonging to a node having a small average RTT value to a node having a large average RTT value so that the load is distributed.

Then, the load balancing is determined based on the weight of the agent A of the node and the weighted average of the corresponding node (S140).

The monitor agent MAg sorts the average of the weights of all the nodes in ascending order, and when the weighted average of one node is smaller than the weighted average of all the nodes, the agent A of the node has the largest weight in the multi-agent system. Migration to a node having an average (S150).

10 and 11 are graphs of the result of measuring a message transmission state occurring in the multi-agent system according to the present invention.

10 and 11 show the average RTT of the multi-agent system to which the present invention (propose scheme) and the existing process (process sharing) are applied, and a 2.4 GHz Pentium in which an agent platform unit and a transmit / receive agent A are installed. Processors, 2 gigabytes of memory, 10 personal computers with 10 / 100Mbps specifications, distributed agents (A) send and receive messages of 100Kb (Figure 10) and 500Kb (Figure 11) periodically during state changes. Receive.

Then, the state change rate s per minute was set such that the agent A caused a state change every 30 seconds or 60 seconds.

10 and 11, when the message transmission cycle is relatively fast, it can be seen that the average RTT according to the present invention is smaller than the conventional scheme, and the multi-agent system according to the present invention has a relatively high workload. This means that it is more useful than the existing technique in the multi-agent system.

10 (b) and 11 (b) are cases where the agent (A) state change rate is higher than (a), and according to the present invention, when the change rate is high, a significant performance improvement is achieved in comparison with the low change rate. The higher the rate of change of the state of the agent A, the more effective it is.

Although the present invention has been described in detail only with respect to the described embodiments, it will be apparent to those skilled in the art that various modifications and changes are possible within the technical spirit of the present invention, and such modifications and modifications belong to the appended claims.

1 is a view for explaining a standard model of the Airgent platform proposed by FIPA.

2 is a diagram for explaining a multi-agent system.

3 is a view for explaining a multi-agent system according to the present invention.

4 is a diagram for explaining an operation state of an agent.

5 is a diagram for explaining resource allocation according to the present invention in more detail.

6 is a view for briefly explaining an agent migration process in a multi-agent system of the present invention.

7 is a view for explaining the structure of the agent platform unit.

8 is a view for explaining the structure of a multi-agent system according to the present invention.

9 is a flowchart illustrating a load balancing method of a multi-agent system according to the present invention.

10 and 11 are graphs of the result of measuring a message transmission state occurring in the multi-agent system according to the present invention.

Claims (8)

In a multi-agent system, Agent platform unit for managing the operation state by registering each agent created and operated on a plurality of nodes and And a monitor agent unit for calculating a weight according to an operation state of each agent identified through the agent platform unit and allocating a resource according to the weight to each agent. The agent may be at least one or more generated by the host to process the user's request operation according to the operation state, The monitor agent unit, And calculating the weight based on the number of agents in an activated state among the agents on each node. delete The method of claim 1, The monitor agent unit, When the round trip time (RTT) value of each agent is equal to or smaller than the average RTT value of the node, the load balancing is performed by migrating one agent among the agents belonging to the node having the smallest average RTT value to a node having an average RTT value. Multi-agent system, characterized in that. The method of claim 1, The monitor agent unit, And repeatedly migrating the agent of the node having a low weighted average until the weighted average of each node is equal to the weighted average of all nodes. In a multi-agent system, An agent migration unit for requesting agent migration to an agent management system (AMS) agent to handle a load balancing process; An agent life cycle information unit which defines an operating state of the generated agent and allows the monitor agent unit to calculate a weight; And And an agent activity scheduler for requesting resource allocation to each agent based on the weight calculated by the monitor agent unit. In the load balancing method of a multi-agent system, Registering, by the agent platform unit, an agent generated to work according to a user's request, and managing an operation state; Determining, by the monitor agent, load balancing based on an average RTT value of each node in which the agent operates and an RTT value for each agent; Calculating, by the monitor agent unit, a weight based on the number of agents in which the operating state of the agent is activated; The monitor agent allocating resources to each agent according to the weight and the amount of resources of the node where each agent operates; And a monitor agent determining load balancing based on the weight of each agent and the weighted average of the corresponding nodes. The method of claim 6, The calculating of the weights may include recalculating the weights when an operation state of at least one agent is transitioned or when an agent is generated or terminated on a node. delete

Images