A Dynamic Orchestration System of Multiple Workflows using a Heterogeneous Multi-Agent

Abstract

The present invention relates to a system for simultaneously managing multiple workflows defined by a Directed Acyclic Graph (DAG) in a heterogeneous multi-agent environment where software agents and robot control agents are mixed. In particular, the system optimizes the stability and efficiency of the entire system by efficiently allocating limited resources through the allocation of optimal agents based on executable tasks and agent status information. Furthermore, even in the event of unexpected failures or performance degradation of individual agents, the system rapidly and efficiently reallocates optimal replacement agents by synthesizing agent information linked by similarity, cost-effectiveness, and precedence based on real-time agent status information, thereby ensuring the continuous execution of the entire workflow and allowing for flexible responses to abnormal situations, new workflows, or the addition of agents.

Description

Technology Field

[0001] The present invention relates to a system for simultaneously managing multiple workflows defined by a Directed Acyclic Graph (DAG) in a heterogeneous multi-agent environment where software agents and robot control agents are mixed. In particular, the system optimizes the stability and efficiency of the entire system by efficiently allocating limited resources through the allocation of optimal agents based on executable tasks and agent status information. Furthermore, even in the event of unexpected failures or performance degradation of individual agents, the system rapidly and efficiently reallocates optimal replacement agents by synthesizing agent information linked by similarity, cost-effectiveness, and precedence based on real-time agent status information, thereby ensuring the continuous execution of the entire workflow and allowing for flexible responses to abnormal situations, new workflows, or the addition of agents. Background Technology

[0002] In complex environments such as modern smart factories, autonomous warehouses, and swarm robot systems, multiple intelligent agents must cooperate to perform tasks. These tasks appear as complex structures involving precedence and dependencies, and a representative data structure for representing them is the Directed Acyclic Graph (DAG).

[0003] Existing DAG-based workflow management systems (e.g., Apache Airflow, Kubeflow) have focused on executing tasks in a predefined order in static environments. However, in real physical environments involving robot control agents, unexpected state changes (failures, sensor errors, communication delays, overloads) occur frequently. Consequently, traditional static scheduling leads to cascading failures, where the entire workflow is delayed or fails due to issues with specific agents.

[0004] Furthermore, when multiple complex workflows were executed simultaneously, there were limitations in efficiently allocating limited agent resources. In particular, the performance heterogeneity of heterogeneous agents (robots, sensors, software), network latency, and energy constraints exacerbate this complexity.

[0005] <Patent Literature>

[0006] Korean Registered Patent No. 10-1935006 (Registered Dec. 27, 2018) "Following agent included in a multi-agent system and method for controlling the same"

[0007] In the case of the prior art disclosed in the above <Patent Document>, the concept of collecting, storing, and managing agent states is disclosed in the sense of checking state information (output signals, etc.) for each agent (leader / neighbor agent) included in a multi-agent system and deriving control inputs of follower agents based thereon, but this is merely about monitoring and managing agent states and is completely unrelated to flexibly managing multiple workflows under a heterogeneous agent environment. The problem to be solved

[0008] The present invention has been devised to solve the above-mentioned problems,

[0009] The objective of the present invention is to provide a multi-workflow dynamic orchestration system utilizing heterogeneous multi-agents that optimizes the stability and efficiency of the entire system by efficiently allocating limited resources through the allocation of optimal agents by reflecting executable tasks and agent status information, in the simultaneous management of multiple workflows defined by a directed acyclic graph (DAG) in a heterogeneous multi-agent environment where software agents and robot control agents are mixed.

[0010] Another objective of the present invention is to provide a multi-workflow dynamic orchestration system utilizing heterogeneous multi-agents that can flexibly respond to abnormal situations, new workflows, or the addition of agents by synthesizing agent information linked by similarity, cost-effectiveness, and precedence based on real-time status information of agents, thereby enabling the entire workflow to continue executing even in the event of unexpected failure or performance degradation of individual agents, and by rapidly and efficiently reallocating an optimal replacement agent. means of solving the problem

[0011] The present invention is implemented by an embodiment having the following configuration to achieve the aforementioned objective.

[0012] According to one embodiment of the present invention, a multi-workflow dynamic orchestration system utilizing heterogeneous multi-agents according to the present invention comprises: an agent management unit that collects, analyzes, and manages state information of each heterogeneous multi-agent, including a software agent and a robot control agent; a joint scheduler unit that simultaneously manages workflows defined by a plurality of directed acyclic graphs (DAGs) and identifies and manages executable tasks within each workflow; a dynamic task assignment unit that assigns or reallocates an optimal agent by reflecting the agent's state information to the executable tasks; and an execution monitoring unit that monitors the progress status and performance indicators of the task assigned to the agent and provides related information.

[0013] According to another embodiment of the present invention, the agent management unit of the present invention is characterized by comprising: an agent state collection module that collects and provides state information of each heterogeneous multi-agent, including a software agent and a robot control agent; an agent fault diagnosis module that diagnoses a state in which performance degradation or failure has occurred for each agent based on the state information of each agent; and an agent state prediction module that analyzes and predicts the possibility of performance degradation or failure of an agent based on the state information of each agent.

[0014] According to another embodiment of the present invention, the agent state collection module in the present invention collects and provides state information including CPU / memory load information of each agent, linked hardware sensor information, linked network state information, availability information, and agent information linked in a sequential relationship; and the agent state prediction module comprises: a first prediction module that predicts the possibility of performance degradation or failure based on the change trends of the CPU / memory state, linked hardware sensor state, and linked network state of each agent collected and managed in real time; a second prediction module that predicts the possibility of performance degradation or failure for an agent exhibiting a state change exceeding a threshold within a group after grouping a plurality of agents operating in similar environments while performing similar functions; a third prediction module that predicts the possibility of performance degradation or failure for a specific agent by comparing it with a reference value set based on existing failure occurrence data within the group in the second prediction module; and a module that synthesizes the information of the first to third prediction modules and reflects the influence of the predicted state information of linked agents preceding and succeeding the specific agent to the agent based on agent information linked in a sequential relationship, thereby performance It is characterized by including a fourth prediction module that predicts the possibility of degradation or failure.

[0015] According to another embodiment of the present invention, the joint scheduler in the present invention comprises a workflow management module that simultaneously manages workflows defined by a plurality of directed acyclic graphs (DAGs), a task selection module that continuously selects tasks that are executable within each managed workflow where a prerequisite is satisfied and provides relevant information, and a task optimization module that optimizes the execution order of the selected tasks. The task optimization module comprises a first optimization module that reflects priority information between the plurality of workflows managed by the workflow management module as a first weight, a second optimization module that reflects information regarding the state condition where a prerequisite is satisfied and subsequent tasks related to executable tasks as a second weight, a third optimization module that reflects the Service Level Agreement (SLA) level for the target workflow or task as a third weight, a fourth optimization module that reflects cost function information for the target workflow or task as a fourth weight, and weight reflection in the first to fourth optimization modules. It is characterized by including a comprehensive execution module that optimizes the task execution order by synthesizing result values.

[0016] According to another embodiment of the present invention, the dynamic task assignment unit of the present invention comprises an agent assignment module that assigns an optimal agent by reflecting the agent's status information to a task that is executable according to the execution order set in the comprehensive execution module, and an agent reassignment module that searches for and reassigns a replacement agent when an abnormality is detected in an agent assigned to and executed for a specific task. The agent assignment module is characterized by comprising: a first assignment module that reflects the state prediction information of each agent presented through the agent state prediction module as a first weight; a second assignment module that reflects distribution information based on load balancing for agents for a plurality of agents as a second weight; a third assignment module that reflects success rate information of a target agent or a group to which the target agent belongs as a third weight based on history information for a plurality of agents that have a history of executing the task; and a fourth assignment module that assigns an optimal agent to the task based on agent information linked in a sequential relationship by synthesizing the information of the first assignment module to the third assignment module.

[0017] According to another embodiment of the present invention, the agent reassignment module of the present invention comprises: a replacement agent search module that searches for a replaceable agent when an abnormality is detected in an agent assigned to and executed for a specific task based on information from the agent fault diagnosis module to the agent state prediction module; a first reassignment module that reflects the similarity between the state and performance of the agents presented through the replacement agent search module and the agent to be replaced as a first weight; a second reassignment module that reflects the ratio of minimizing costs incurred during replacement for each agent presented through the replacement agent search module as a second weight; and a third reassignment module that synthesizes the information from the first reassignment module to the second reassignment module and reassigns an optimal replacement agent based on the current execution information of the agent or agent information linked in a sequential relationship when a specific agent is reassigned as a replacement agent. Effects of the invention

[0018] The present invention can achieve the following effects through the combination and usage relationship of the embodiments described above and the configuration described below.

[0019] The present invention has the effect of optimizing the stability and efficiency of the entire system by efficiently allocating limited resources through the allocation of optimal agents by reflecting executable tasks and agent status information, particularly in managing multiple workflows defined by a directed acyclic graph (DAG) simultaneously in a heterogeneous multi-agent environment where software agents and robot control agents are mixed.

[0020] The present invention has the effect of enabling the entire workflow to continue executing even in the event of unexpected failure or performance degradation of individual agents, by synthesizing agent information linked by similarity, cost-effectiveness, and precedence based on real-time status information of the agents to quickly and efficiently reallocate the optimal replacement agent, thereby allowing for flexible response to abnormal situations, new workflows, or the addition of agents. Brief explanation of the drawing

[0021] FIG. 1 is a system configuration diagram according to an embodiment of the present invention. Figure 2 is a detailed configuration diagram of the Agent Management Department. Figure 3 is a detailed configuration diagram of the joint scheduler section. Figure 4 is a detailed configuration diagram of the dynamic task assignment unit. Specific details for implementing the invention

[0022] Preferred embodiments of a multi-workflow dynamic orchestration system utilizing heterogeneous multi-agents according to the present invention will be described in detail below with reference to the accompanying drawings. It should be noted that identical components in the drawings are represented by the same reference numerals wherever possible. Unless otherwise specifically defined, all terms in this specification have the same general meaning as understood by a person skilled in the art to which the present invention pertains, and in the event of a conflict with the meaning of a term used in this specification, the definition used in this specification shall prevail. Throughout the specification, when a part is described as "comprising" a certain component, this means that, unless specifically stated otherwise, it does not exclude other components but may include additional components. Furthermore, terms such as "...part," "...module," etc., described in the specification refer to a unit that processes at least one function or operation, and this may be implemented in hardware, software, or a combination of hardware and software.

[0024] Referring to FIG. 1 and the like, a multi-workflow dynamic orchestration system utilizing heterogeneous multi-agents according to one embodiment of the present invention is characterized by comprising: an agent management unit (10) that collects, analyzes, and manages status information of each heterogeneous multi-agent, including a software agent and a robot control agent; a joint scheduler unit (20) that simultaneously manages workflows defined by a plurality of directed acyclic graphs (DAGs) and identifies and manages tasks that can be executed within each workflow; a dynamic task assignment unit (30) that assigns or reassigns an optimal agent by reflecting the status information of the agent to the tasks that can be executed; and an execution monitoring unit (40) that monitors the progress status and performance indicators of the tasks assigned to the agent and provides related information. As previously explained, the system of the present invention is a system that simultaneously manages multiple workflows defined by a Directed Acyclic Graph (DAG) in a heterogeneous multi-agent environment where software agents and robot control agents are mixed. In particular, it is characterized by optimizing the stability and efficiency of the entire system by efficiently allocating limited resources through the allocation of optimal agents by reflecting executable tasks and agent status information. Furthermore, even in the event of unexpected failures or performance degradation of individual agents, it rapidly and efficiently reallocates optimal replacement agents by synthesizing agent information linked by similarity, cost-effectiveness, and precedence based on real-time agent status information, thereby enabling the continuous execution of the entire workflow and allowing for flexible responses to abnormal situations, new workflows, or the addition of agents. The main components and functions for this purpose will be described in detail below.

[0025] The above agent management unit (10) is configured to collect, analyze, and manage status information of each heterogeneous multi-agent, including a software agent and a robot control agent. To this end, the agent management unit (10) may more specifically include an agent status collection module (110) that collects and provides status information of each heterogeneous multi-agent, including a software agent and a robot control agent; an agent fault diagnosis module (120) that diagnoses a state in which performance degradation or failure has occurred for each agent based on the status information of each agent; and an agent status prediction module (130) that analyzes and predicts the possibility of performance degradation or failure of an agent based on the status information of each agent.

[0026] The agent state collection module (110) is configured to collect and provide state information for each of the heterogeneous multi-agents, including software agents and robot control agents. Specifically, the agent state collection module (110) can collect and provide state information including CPU / memory load information, linked hardware sensor information, linked network status information, availability information, and information on agents linked in a sequential relationship for each agent. That is, the agent state collection module (110) can collect and provide relevant state information including information on the performance, current status, and linked hardware sensors of each agent, as well as information on the connection relationships with other agents linked in a sequential relationship with the agent.

[0027] The agent fault diagnosis module (120) described above is configured to diagnose a state in which performance degradation or failure has occurred for each agent based on the status information of each agent. That is, the agent fault diagnosis module (120) diagnoses the performance degradation or failure currently occurring in the agent, and this is distinguished from the agent state prediction module (130) described later, which predicts and analyzes the possibility of performance degradation or failure of the agent.

[0028] The agent state prediction module (130) is configured to analyze and predict the possibility of performance degradation or failure of an agent based on the state information of each agent. To this end, the agent state prediction module (130) more specifically comprises: a first prediction module (131) that predicts the possibility of performance degradation or failure based on the trend of change in the CPU / memory status, linked hardware sensor status, and linked network status of each agent collected and managed in real time; a second prediction module (132) that predicts the possibility of performance degradation or failure for an agent that shows a state change exceeding a threshold within a group after grouping multiple agents that perform similar functions and operate in similar environments; a third prediction module (133) that predicts the possibility of performance degradation or failure for a specific agent by comparing it with a reference value set based on existing failure occurrence data within the group in the second prediction module (132); and a third prediction module (133) that synthesizes the information from the first prediction module (131) to the third prediction module (133) and, based on agent information linked in a sequential relationship, predicts the possibility of failure of the agents linked to the preceding and succeeding ends of a specific agent. It may include a fourth prediction module (134) that predicts the possibility of performance degradation or failure by reflecting the influence of predicted state information to the agent.

[0029] The first prediction module (131) is configured to predict the possibility of performance degradation or failure based on the trend of change in the CPU / memory status, associated hardware sensor status, and associated network status of each agent collected and managed in real time. The first prediction module (131) can predict the possibility of performance degradation or failure for a specific agent by analyzing whether a change in the trend of state change occurs in one or more of the information, such as the CPU / memory status, associated hardware sensor status, and associated network status of the agent, and whether a certain inflection point appears in such change, based on the status information continuously collected and managed for the specific agent. That is, the first prediction module (131) performs a prediction function for each specific agent based on the occurrence or possibility of occurrence of an inflection point in the individual status information of the agent, particularly the trend information of state change.

[0030] The second prediction module (132) is configured to group multiple agents that operate in similar environments while performing similar functions, and then predict the possibility of performance degradation or failure for an agent that exhibits a state change outside a threshold within the group. Unlike the first prediction module (131) described above, which is based on individual state information of each specific agent, the second prediction module (132) groups multiple agents that operate in similar functions or environments to a specific agent, and then predicts the possibility of performance degradation or failure through relative comparison within the group. This allows for the prediction of the possibility of failure by determining relative performance degradation, etc., even if an inflection point in the state change does not appear.

[0031] The third prediction module (133) is configured to predict the possibility of performance degradation or failure for a specific agent by comparing it with a reference value set based on existing failure occurrence data within a group in the second prediction module (132). That is, based on the analysis results of historical data related to existing failure occurrences for multiple agents operating in similar functions or similar environments grouped through the second prediction module (132), the possibility of performance degradation or failure for a specific agent is predicted by comparing the reference value information regarding the possibility of performance degradation or failure derived from the analysis results with the current status information of the specific agent. This allows for the prediction of the possibility of performance degradation or failure through a comparison with a kind of absolute value (reference value), in contrast to the relative comparison in the second prediction module (132).

[0032] The above-mentioned fourth prediction module (134) is configured to predict the possibility of performance degradation or failure by synthesizing the information of the first prediction module (131) to the third prediction module (133) and reflecting the influence of the predicted state information of the agents connected to the preceding and succeeding ends of a specific agent on the agent based on agent information linked in a sequential relationship. That is, the above-mentioned fourth prediction module (134) does not simply synthesize and judge only the information of the first prediction module (131) to the third prediction module (133), but rather reflects the predicted state information of the agents connected to the preceding and succeeding ends of a specific agent based on the connection relationship of multiple agents, that is, agent information linked in a sequential relationship, thereby comprehensively considering even the state change of the agents connected to the preceding or succeeding ends to predict the possibility of performance degradation or failure for the specific agent. In this way, data regarding agents identified and predicted through the agent management unit (10) can be recorded and managed in a separate 'agent state table'.

[0033] The above joint scheduler unit (20) is configured to simultaneously manage workflows defined by a plurality of directed acyclic graphs (DAGs) and to identify and manage executable tasks within each workflow. To this end, the joint scheduler unit (20) may more specifically include a workflow management module (210) that simultaneously manages workflows defined by a plurality of directed acyclic graphs (DAGs), a task selection module (220) that continuously selects executable tasks within each managed workflow where preconditions are satisfied and provides relevant information, and a task optimization module (230) that optimizes the execution order of the selected tasks.

[0034] The workflow management module (210) is configured to simultaneously manage workflows defined by multiple directed acyclic graphs (DAGs). In complex environments such as modern smart factories, logistics warehouses, and swarm drone systems, multiple workflows defined by directed acyclic graphs (DAGs) are executed simultaneously. Since it is important to manage these multiple workflows by efficiently utilizing limited agent resources, the workflow management module (210) can collect, analyze, and provide relevant data (including linkage relationships, sequence relationships, etc. between workflows) to simultaneously manage workflows defined by multiple directed acyclic graphs (DAGs).

[0035] The task selection module (220) is configured to continuously select tasks that are executable when preconditions are satisfied within each managed workflow and provide related information. That is, based on real-time collected and provided information regarding workflows defined as multiple directed acyclic graphs (DAGs) managed through the workflow management module (210), the task selection module (220) can continuously select tasks that become executable when preconditions are satisfied (i.e., the execution of a precondition task is completed) within each workflow and provide related information (to the task optimization module (etc. to be described later).

[0036] The task optimization module (230) is configured to optimize the execution order of selected tasks. To this end, the task optimization module (230) more specifically comprises: a first optimization module (231) that reflects priority information among multiple workflows managed by the workflow management module (210) as a first weight; a second optimization module (232) that reflects information regarding a state condition in which a preceding condition is satisfied and subsequent tasks related to an executable task as a second weight; a third optimization module (233) that reflects the Service Level Agreement (SLA) level for a target workflow or task as a third weight; a fourth optimization module (234) that reflects cost function information for a target workflow or task as a fourth weight; and a method for optimizing the task execution order by synthesizing the result values ​​of the weight reflections in the first optimization module (231) to the fourth optimization module (234). It may include a comprehensive execution module (235).

[0037] The first optimization module (231) is configured to reflect priority information among multiple workflows managed by the workflow management module (210) as a first weight. Since there exists a linkage and precedence relationship among the multiple workflows managed by the workflow management module (210), such that even if workflows are executed simultaneously in a complex environment, a specific workflow must precede the subsequent workflow linked thereto, the first optimization module (231) can calculate and provide execution order information for multiple tasks by reflecting priority information among multiple workflows managed by the workflow management module (210) as a first weight when optimizing the execution order of multiple tasks.

[0038] The above second optimization module (232) is configured to reflect information regarding the state condition in which a preceding condition is satisfied and information regarding subsequent tasks as a second weight in relation to an executable task. That is, regarding tasks that have become executable, information such as the time and waiting time when the task became executable (i.e., when a preceding condition is satisfied) as a condition for the state in which the task has become executable, as well as information regarding subsequent tasks linked to the task, such as the number of subsequent tasks and other linked workflow information, are reflected as a second weight in optimizing the execution order, thereby calculating and providing information on the execution order of multiple tasks.

[0039] The above third optimization module (233) is configured to reflect the Service Level Agreement (SLA) level for the target workflow or task as a third weight. That is, in optimizing the execution order of multiple tasks, the above third optimization module (233) can calculate and provide information on the execution order of multiple tasks by additionally reflecting the Service Level Agreement (SLA) level for the target task or the workflow containing the target task—that is, the information on the quality (execution capability) level of the corresponding task that must be satisfied—as a third weight.

[0040] The above-mentioned fourth optimization module (234) is configured to reflect cost function information for a target workflow or task as a fourth weight. That is, in the fourth optimization module (234), when optimizing the execution order of multiple tasks, the expected cost information in the case where a target task or a workflow including a target task is executed or cannot be executed is measured through a cost function and reflected as a fourth weight. At this time, not only the task but also the cost information for subsequent tasks linked thereto is reflected together to calculate and provide execution order information for multiple tasks.

[0041] The above-described comprehensive execution module (235) is configured to optimize the task execution order by synthesizing the result values ​​reflecting weights from the first optimization module (231) to the fourth optimization module (234). It can derive and provide a finally optimized execution order for a plurality of tasks based on the result values ​​by comprehensively reflecting the execution order information calculated by reflecting each weight (first weight to fourth weight) derived from each of the first optimization module (231) to the fourth optimization module (234) described above.

[0042] The dynamic task assignment unit (30) is configured to assign or reassign an optimal agent by reflecting the agent's status information to an executable task. To this end, the dynamic task assignment unit (30) may more specifically include an agent assignment module (310) that assigns an optimal agent by reflecting the agent's status information to an executable task according to the execution order set in the comprehensive execution module (235), and an agent reassignment module (320) that searches for and reassigns a replacement agent when an abnormality is detected in an agent assigned to and executed on a specific task.

[0043] The agent assignment module (310) is configured to assign an optimal agent to a task that is executable according to the execution order set in the comprehensive execution module (235) by reflecting the agent's status information. To this end, the agent assignment module (310) more specifically comprises: a first assignment module (311) that reflects the status prediction information of each agent presented through the agent status prediction module (130) as a first weight; a second assignment module (312) that reflects distribution information based on load balancing for agents targeting a plurality of agents as a second weight; a third assignment module (313) that reflects the success rate information of a target agent or the group to which the target agent belongs as a third weight based on history information of a plurality of agents that have a history of executing the task; and an agent that assigns an optimal agent to the task based on agent information linked in a sequential relationship by synthesizing the information of the first assignment module (311) to the third assignment module (313). It may include a fourth allocation module (314).

[0044] The first allocation module (311) is configured to reflect the state prediction information of each agent presented through the agent state prediction module (130) as a first weight. That is, the first allocation module (311) reflects not only the current state information of each agent identified through the agent state collection module (110) but also the state prediction information of each agent presented through the agent state prediction module (130) as a first weight, thereby ensuring that the current and predicted state information of the agent to perform the corresponding task is reflected with the highest priority.

[0045] The second allocation module (312) is configured to reflect distribution information based on load balancing for agents for multiple agents as a second weight. That is, since the system of the present invention enables efficient operation by efficiently distributing limited agent resources when multiple complex workflows are executed simultaneously, the second allocation module (312) reflects not only the status information of multiple agents but also distribution information based on load balancing for multiple agents as a second weight, thereby allowing for consideration of the efficient distribution of limited agent resources in the execution of multiple tasks.

[0046] The third assignment module (313) is configured to reflect the success rate information of a target agent or the group to which the target agent belongs as a third weight based on historical information of multiple agents that have a history of executing the task. Since the success of the task is the most important factor when a specific agent is assigned to and executes a specific task, the third assignment module (313) specifically reflects the success rate information analyzed from the target agent (if there is historical information of the target agent) or the group to which the target agent belongs (for example, the group in the second prediction module (132)) based on historical data information, based on the past history information of multiple agents that have a history of executing the task, thereby increasing the probability of success of the task execution during assignment.

[0047] The above-mentioned fourth assignment module (314) is configured to integrate information from the above-mentioned first assignment module (311) to third assignment module (313) and to assign an optimal agent to the corresponding task based on agent information linked in a sequential relationship. That is, the above-mentioned fourth assignment module (314) does not simply integrate and determine only the information from the above-mentioned first assignment module (311) to third assignment module (313), but rather reflects the predicted state information for the agents (or tasks) linked to the preceding and succeeding ends of a specific agent based on the connection relationships of multiple agents, that is, agent (or task) information linked in a sequential relationship. By comprehensively considering the state, performance capability, and expected state changes of the agents linked to the preceding or succeeding ends, the optimal agent is assigned to the target task, thereby enabling rapid, efficient, and successful execution.

[0048] The agent reassignment module (320) is configured to search for and reassign a replacement agent when an abnormality is detected in an agent assigned to and executed for a specific task. To this end, the agent reassignment module (320) more specifically comprises: a replacement agent search module (321) that searches for a replacement agent when an abnormality is detected in an agent assigned to and executed for a specific task based on information from the agent fault diagnosis module (120) to the agent state prediction module (130); a first reassignment module (322) that reflects the similarity between the status and performance of the agents presented through the replacement agent search module (321) and the replacement target agent as a first weight; a second reassignment module (323) that reflects the cost minimization ratio for each agent presented through the replacement agent search module (321) as a second weight; and the first reassignment module (322) to It may include a third reassignment module (324) that synthesizes information from the second reassignment module (323) and reassigns an optimal replacement agent based on the current execution information of the agent or agent information linked in a sequential relationship when a specific agent is reassigned as a replacement agent.

[0049] The above-mentioned replacement agent search module (321) is configured to search for a replacement agent when an abnormality is detected in an agent assigned to and executed for a specific task based on information from the agent failure diagnosis module (120) to the agent status prediction module (130). That is, among a plurality of agents under management, it can search for and provide a candidate group of agents that can be replaced and deployed based on information such as current status and available performance.

[0050] The first reallocation module (322) is configured to reflect the similarity between the state and performance of the agents presented through the replacement agent search module (321) and the agent to be replaced as a first weight. The state and performance information of each candidate group of replaceable agents presented through the replacement agent search module (321) and the performance similarity of the agent to be replaced are reflected as the first weight with the highest priority. At this time, the first reallocation module (322) reflects not only the current state information of each agent but also the state prediction information of each agent presented through the agent state prediction module (130), so that the current and predicted state information of the agent to be replaced in the task is reflected with the highest priority.

[0051] The second reallocation module (323) is configured to reflect the cost minimization ratio that occurs when replacing each of the agents presented through the replacement agent search module (321) as a second weight. That is, in the second reallocation module (323), when reallocating the optimal replacement agent, the cost information expected when replacing with or not replacing with the corresponding agent is measured through a cost function for each of the agents presented through the replacement agent search module (321), and the information is reflected as a second weight. At this time, not only the corresponding agent but also the cost information linked to subsequent tasks (or agents) linked thereto can be reflected together.

[0052] The third reassignment module (324) is configured to synthesize information from the first reassignment module (322) to the second reassignment module (323) and to reassign an optimal replacement agent based on the current execution information of the agent or information of agents linked in a sequential relationship when a specific agent is reassigned as a replacement agent. That is, the third reassignment module (324) does not simply synthesize and judge only the information from the first reassignment module (322) to the second reassignment module (323), but reflects not only the current execution information of the agent to be replaced but also the connection relationship of multiple agents, that is, information of agents (or tasks) linked in a sequential relationship, and even reflects predicted state information for the agents (or tasks) linked to the preceding and succeeding ends of the agent. By comprehensively considering the state, performance capability, and expected state change of the linked agents at the preceding or succeeding ends, the optimal replacement agent can be reassigned to the target task, thereby enabling rapid, efficient, and successful execution.

[0053] The execution monitoring unit (40) is configured to monitor the progress status and performance indicators of tasks assigned to agents and provide related information. It performs the function of monitoring in real time the progress status of tasks, progress results (success / failure information, etc.), and various occurrence events such as network failures for tasks that are assigned or reallocated to the optimal agent through the joint scheduler unit (20) and dynamic task assignment unit (30) described above, and collects and provides related information.

[0054] In this way, through the system of the present invention, multiple workflows defined by a Directed Acyclic Graph (DAG) can be managed quickly and efficiently in a heterogeneous multi-agent environment where software agents and robot control agents are mixed. In particular, by assigning the optimal agent based on executable tasks and agent status information, limited resources are efficiently allocated to optimize the stability and efficiency of the entire system. Furthermore, even in the event of unexpected failure or performance degradation of individual agents, the system can quickly and efficiently reallocate the optimal replacement agent by synthesizing agent information linked by similarity, cost-effectiveness, and precedence based on the real-time status information of the agents, thereby ensuring that the entire workflow continues to execute and enabling flexible responses to abnormal situations, new workflows, or the addition of agents.

[0056] Although the applicant has described various embodiments of the present invention above, such embodiments are merely examples of implementing the technical concept of the present invention, and any modification or alteration that implements the technical concept of the present invention should be interpreted as falling within the scope of the present invention. Explanation of the symbols

[0057] 10: Agent Management Department 110: Agent Status Collection Module 120: Agent Fault Diagnosis Module 130: Agent Status Prediction Module 131: 1st Prediction Module 132: 2nd Prediction Module 133: 3rd Prediction Module 134: 4th Prediction Module 20: Joint Scheduler Unit 210: Workflow Management Module 220: Task Selection Module 230: Task Optimization Module 231: 1st Optimization Module 232: 2nd Optimization Module 233: 3rd Optimization Module 234: 4th Optimization Module 235: Comprehensive Execution Module 30: Dynamic Task Assignment Unit 310: Agent Assignment Module 311: 1st Allocation Module 312: 2nd Allocation Module 313: 3rd Allocation Module 314: 4th Allocation Module 320: Agent Reassignment Module 321: Replacement Agent Search Module 322: 1st Reallocation Module 323: 2nd Reallocation Module 324: 3rd Reallocation Module 40: Execution Monitoring Section

Claims

Claim 1 delete Claim 2 delete Claim 3 An agent management unit that collects, analyzes, and manages status information of each heterogeneous multi-agent, including software agents and robot control agents; a joint scheduler unit that simultaneously manages workflows defined by multiple directed acyclic graphs (DAGs) and identifies and manages executable tasks within each workflow; a dynamic task assignment unit that assigns or reallocates the optimal agent by reflecting the agent's status information in the executable tasks; and an execution monitoring unit that monitors the progress status and performance indicators of tasks assigned to agents and provides related information.The agent management unit includes an agent state collection module that collects and provides state information for each of heterogeneous multi-agents, including software agents and robot control agents; an agent fault diagnosis module that diagnoses a state in which performance degradation or failure has occurred for each agent based on the state information of each agent; and an agent state prediction module that analyzes and predicts the possibility of performance degradation or failure of an agent based on the state information of each agent. The agent state collection module collects and provides state information including CPU / memory load information, linked hardware sensor information, linked network status information, availability information, and agent information linked in a sequential relationship for each agent. The agent state prediction module includes a first prediction module that predicts the possibility of performance degradation or failure based on the change trends of the CPU / memory status, linked hardware sensor status, and linked network status of each agent collected and managed in real time; a second prediction module that predicts the possibility of performance degradation or failure for an agent that exhibits a state change exceeding a threshold within a group after grouping multiple agents that perform similar functions and operate in similar environments; and the existing within the group in the second prediction module A multi-workflow dynamic orchestration system utilizing heterogeneous multi-agents, characterized by comprising: a third prediction module that predicts the possibility of performance degradation or failure for a specific agent in comparison to a reference value set based on failure occurrence-related data; and a fourth prediction module that predicts the possibility of performance degradation or failure by synthesizing the information of the first to third prediction modules and reflecting the influence of the predicted state information of agents linked upstream and downstream of the specific agent to the agent based on agent information linked in a sequential relationship. Claim 4 An agent management unit that collects, analyzes, and manages status information of each heterogeneous multi-agent, including software agents and robot control agents; a joint scheduler unit that simultaneously manages workflows defined by multiple directed acyclic graphs (DAGs) and identifies and manages executable tasks within each workflow; a dynamic task assignment unit that assigns or reallocates the optimal agent by reflecting the agent's status information in the executable tasks; and an execution monitoring unit that monitors the progress status and performance indicators of tasks assigned to agents and provides related information.The agent management unit includes an agent state collection module that collects and provides state information for each of heterogeneous multi-agents, including software agents and robot control agents; an agent fault diagnosis module that diagnoses a state in which performance degradation or failure has occurred for each agent based on the state information of each agent; and an agent state prediction module that analyzes and predicts the possibility of performance degradation or failure of an agent based on the state information of each agent. The joint scheduler unit includes a workflow management module that simultaneously manages workflows defined by a plurality of directed acyclic graphs (DAGs); a task selection module that continuously selects tasks that are executable within each managed workflow where preconditions are satisfied and provides related information; and a task optimization module that optimizes the execution order of the selected tasks. The task optimization module includes a first optimization module that reflects priority information between a plurality of workflows managed by the workflow management module as a first weight; and a second weight that reflects information regarding the state condition in which preconditions are satisfied and information regarding subsequent tasks related to executable tasks. A multi-workflow dynamic orchestration system utilizing heterogeneous multi-agents, characterized by comprising: a second optimization module; a third optimization module that reflects the Service Level Agreement (SLA) level for a target workflow or task as a third weight; a fourth optimization module that reflects cost function information for a target workflow or task as a fourth weight; and a comprehensive execution module that optimizes the task execution order by synthesizing the result values ​​of weight reflection from the first to fourth optimization modules. Claim 5 In claim 4, the dynamic task assignment unit comprises an agent assignment module that assigns an optimal agent by reflecting the agent's status information to an executable task according to the execution order set by the comprehensive execution module, and an agent reassignment module that searches for and reassigns a replacement agent when an anomaly is detected in an agent assigned to and executed on a specific task; wherein the agent assignment module comprises a first assignment module that reflects the status prediction information of each agent presented through the agent status prediction module as a first weight, a second assignment module that reflects distribution information based on load balancing for agents targeting a plurality of agents as a second weight, a third assignment module that reflects the success rate information of a target agent or the group to which the target agent belongs as a third weight based on history information for a plurality of agents that have an execution history of the corresponding task, and a fourth assignment module that integrates the information of the first to third assignment modules and assigns an optimal agent to the corresponding task based on agent information linked in a sequential relationship. Multi-workflow dynamic orchestration system utilizing Claim 6 In claim 5, the agent reassignment module comprises: a replacement agent search module that searches for a replaceable agent when an abnormality is detected in an agent assigned to and executed for a specific task based on information from the agent fault diagnosis module or the agent state prediction module; a first reassignment module that reflects the similarity between the state and performance of the agents presented through the replacement agent search module and the agent to be replaced as a first weight; a second reassignment module that reflects the ratio of minimizing costs incurred during replacement for each agent presented through the replacement agent search module as a second weight; and a third reassignment module that synthesizes the information from the first reassignment module to the second reassignment module and reassigns an optimal replacement agent based on the current execution information of the agent or agent information linked in a sequential relationship when a specific agent is reassigned as a replacement agent.

Images