An ai research system and method for dynamic critical path multi-objective optimization

Abstract

The application discloses an AI discussion system and method for dynamic critical path multi-objective optimization, and the system comprises: an infrastructure layer for providing underlying physical architecture and basic technical support; a technology layer for supporting and realizing core technical components and development frameworks of the running of the discussion system, including: an intelligent agent module, a database, a SpringBoot framework and an algorithm; the intelligent agent module is used for realizing a special intelligent agent construction function system including dynamic task decomposition and optimization decision; a tool layer is used for constructing a unified tool management platform and a knowledge base, realizing registration scheduling and collaborative calling of cross-system resources; a business layer is used for realizing core business functions including expert profile management and whole-cycle management and control of discussion meetings; and an application layer divides a user system into field expert users and system operation and maintenance administrators, and respectively corresponds to functional requirements of participating in discussion and system operation and maintenance. The application is based on an AI Agent architecture, and can realize complex system task decomposition and resource scheduling.

Description

Technical Field

[0001] This invention relates to the field of dynamic multi-objective collaborative optimization technology, and in particular to an AI discussion system and method for dynamic critical path multi-objective optimization. Background Technology

[0002] Addressing the dynamic path planning and multi-objective optimization problems in collaborative decision-making scenarios involving group experts within a High-Level Integrated Workshop (HWME) environment, this technical field encompasses reinforcement learning-based algorithms for real-time adjustment of dynamic critical path weights, methods for generating Pareto optimal solution sets under multiple constraints, intelligent routing mechanisms for heterogeneous toolchains, and a hierarchical knowledge base retrieval system oriented towards long short-term memory fusion. It focuses on resolving technical bottlenecks in traditional workshop systems, such as rigid task execution paths, conflicts in multi-objective resource allocation, and sluggish real-time environment response. Its core technological innovations lie in enhancing the adaptability of task decomposition and resource scheduling in dynamic environments. This includes dynamic reconstruction of critical paths through multi-dimensional state space modeling, multi-objective collaborative optimization based on improved evolutionary algorithms, and intelligent fault-tolerant scheduling under abnormal toolchain conditions. It belongs to the interdisciplinary field of intelligent decision support systems and complex system control theory, distinguishing itself from general-purpose artificial intelligence algorithms or static task planning techniques.

[0003] The High-Level Integrated Workshop (HWME), a decision-making methodology for complex mega-systems proposed by Qian Xuesen, has long relied on rule-based task decomposition and static critical path planning for its technical implementation. Mainstream domestic and international solutions, such as the CoT-CPM hybrid framework proposed by MIT in 2022, use static CPM to plan paths after generating a task tree through chain thinking (CoT). While this solves the problem of structured task decomposition, it suffers from dynamic adaptability defects: its preset fixed weight coefficients lead to excessively long path adjustment response times and cannot cope with real-time changes in expert opinions. IBM's toolchain management system released in 2023, while supporting multiple tool registrations, lacks load awareness and dynamic routing capabilities, resulting in excessively high tool switching failure rates in high-concurrency scenarios such as ship collaborative design. Regarding multi-objective optimization, the NSGA-II improved scheme proposed by the EU's Horizon 2025 project increases the solution set adoption rate to 68% through algorithm parameter tuning, but still suffers from problems such as the disconnect between human and machine collaboration and the inability to provide real-time feedback of expert experience to the optimization process. The HWME 3.0 system (CN115829132A) of the National University of Defense Technology in China adopts an expert weighted voting mechanism to achieve dynamic adjustment of tasks, but its response speed and resource utilization rate are far from meeting the requirements of actual combat.

[0004] Existing technologies are limited by three major bottlenecks: First, the contradiction between static path planning and dynamic decision-making requirements; traditional CPM relies on fixed-time estimation and sequential logic, which cannot adapt to resource fluctuations through online learning. Second, the lack of coordination between multi-objective optimization and group decision-making; Pareto solution generation and expert preference calibration suffer from temporal misalignment. Third, rigid toolchain binding and insufficient compatibility with heterogeneous systems make it difficult to guarantee task continuity under sudden anomalies. These problems stem from insufficient accuracy in dynamic environment modeling, the lack of closed-loop human-computer interaction design, and limitations in system architecture scalability, hindering the engineering implementation of HWME theory in modern agile decision-making scenarios.

[0005] Therefore, in response to the above bottlenecks, this invention aims to overcome the core shortcomings of existing integrated discussion systems in dynamic decision-making scenarios: (1) The traditional static critical path method (CPM) cannot respond in real time to changes in expert opinions and resource fluctuations due to fixed weight coefficients, resulting in a delay of more than 28 minutes in adjusting the task path, and the execution link becomes rigid when the priority is suddenly reset. (2) In the process of multi-objective optimization, the Pareto solution set is disconnected from the expert decision-making requirements. The existing NSGA-II improvement scheme lacks a human-machine collaborative verification mechanism, and the solution set adoption rate is less than 68%, making it difficult to balance the conflicts of multiple constraints such as time, cost, and quality. (3) Toolchain management relies on preset binding rules, with an abnormal switching failure rate as high as 15%, and the data flow efficiency across tools is low; (4) Knowledge base retrieval is limited by a single storage mode, resulting in low semantic matching accuracy and a knowledge retrieval delay of over 5 seconds during dynamic decision-making. By integrating reinforcement learning dynamic path optimization, human-machine collaborative multi-objective decision-making, intelligent tool routing, and a hierarchical knowledge base architecture, this invention significantly improves the real-time performance of task adjustments, the availability of multi-objective solution sets, the robustness of the toolchain, and the efficiency of knowledge retrieval, overcoming the challenges of agile decision-making and efficient resource utilization in complex scenarios. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide an AI discussion system and method for dynamic critical path multi-objective optimization, which addresses the deficiencies in the prior art.

[0007] The technical solution adopted by this invention to solve its technical problem is: This invention provides an AI discussion system for dynamic critical path multi-objective optimization, the system comprising: The infrastructure layer provides the underlying physical architecture and basic technical support, including: network infrastructure, server arrays, storage systems, distributed computing clusters, and hardware devices; The technical layer comprises the core technical components and development framework that support and enable the operation of the seminar system, including: an agent module, a database, the Spring Boot framework, and algorithms; the agent module is used to implement a specialized agent construction function system, including dynamic task decomposition and optimization decision-making. The tool layer is used to build a unified tool management platform and knowledge base to realize the registration, scheduling and collaborative invocation of cross-system resources; The business layer is used to implement core business functions, including expert file management and full-cycle control of seminars and conferences. The application layer is used to define core permissions for the user system, including: domain expert users and system operation and maintenance administrators, which correspond to the functional requirements of participating in discussions and system operation and maintenance, respectively.

[0008] Furthermore, the method for dynamic task decomposition in the intelligent agent module of the present invention includes: A dual-channel architecture combining Chained Thinking (CoT) and reinforcement learning is adopted. CoT decomposes fuzzy requirements into structured subtask trees, including three types of dependencies: sequential flow, parallel flow, and feedback loop, ultimately generating an initial task network. Simultaneously, a PPO reinforcement learning model is constructed, which uses a six-dimensional state vector based on task latency rate, CPU / memory utilization, solution quality gradient, expert participation entropy, dynamic event trigger flags, and consensus change rate to perceive environmental changes in real time and dynamically adjust the critical path through a continuous action space.

[0009] Furthermore, the method for making optimization decisions in the intelligent agent module of the present invention includes: We construct a three-dimensional objective function of time, resources, and quality, and use constraint violation calculation to screen feasible solutions. We improve the NSGA-II algorithm by introducing an adaptive crossover rate and elite retention strategy to generate a Pareto optimal solution set and embed an expert correction verification loop.

[0010] Furthermore, the knowledge base of this invention adopts dynamic management technology and designs a hierarchical storage architecture for long short-term memory. Long-term memory uses a relational database to store expert information, a graph database for solution association networks, and a vector database for semantic indexing of research reports, and performs categorized storage. Short-term memory caches context states through Redis and introduces an expert participation entropy value quantification model to dynamically adjust knowledge retrieval weights.

[0011] Furthermore, the tool layer of this invention adopts a tool intelligent routing engine, which realizes dynamic registration and capability description of tools based on standardized API interfaces, develops a task status sensor to monitor tool execution latency and resource utilization, and triggers tool switching decisions in combination with task priority queues and sliding window mechanisms.

[0012] This invention provides an AI-based method for dynamic critical path multi-objective optimization, comprising the following steps: Step 1: Construct an AI discussion system for dynamic critical path multi-objective optimization based on a multi-layered framework; build a knowledge base using a multimodal knowledge classification system, a distributed storage engine, and a semantically enhanced retrieval algorithm; Step 2: Input the discussion task, perceive the dynamic events of the discussion task, and extract constraints and semantic keywords; adopt a two-stage task optimization mechanism: the first stage generates the task topology structure through thinking chain technology and sets up an expert intervention verification mechanism to ensure logical completeness; the second stage combines the near-end strategy optimization PPO algorithm and the critical path method CPM to establish a dynamic critical path evolution model, realize dynamic task classification and path reconstruction. Step 3: Integrate the system, model and external tools to realize tool calls driven by the process engine, and perform function matching, parameter generation and service docking during the interaction process.

[0013] Furthermore, the implementation method of step 1 of the present invention includes: Step 1.1, Multimodal Knowledge Classification System: A hierarchical storage architecture is adopted, dividing the memory system into long-term memory and short-term memory; long-term memory realizes the long-term storage and semantic retrieval of structured knowledge such as event consensus set, algorithm rule base, and experience case base through vector database and semantic indexing technology; short-term memory builds a real-time status tracking module based on memory database, and integrates time-series marking mechanism to manage dynamic data such as task execution context, meeting process snapshot and user interaction log; Step 1.2, Distributed Storage Engine: Extracts and maps long-term memory, storing it in an ordered and hierarchical manner in a structured way; short-term memory is stored in an in-memory database, and whether to persist it is selected according to business needs; based on the type of knowledge data, relational database, graph database, vector database, and cache are selected for storage respectively. Step 1.3, Semantic Enhancement Retrieval Algorithm, including a three-layer processing architecture: Indexing mechanism, which builds an index structure for core data fields; Semantic matching, which completes the semantic association mapping between query statements and knowledge base content based on natural language processing and semantic understanding algorithms; Caching mechanism, which uses a memory-level cache pool to store frequently accessed knowledge nodes, thereby optimizing query performance by reducing the backend storage I / O load.

[0014] Furthermore, the implementation method of step 2 of the present invention includes: Step 2.1, Dynamic Event Perception Stage of Discussion Task: For the discussion task, the problem task is defined based on the preset discussion topic, the task boundary is intelligently identified according to the knowledge base, and the core constraints and semantic keywords are extracted simultaneously. Step 2.2, Dynamic Task Decomposition Stage: Based on core constraints and semantic keywords, combined with historical data from the knowledge base, multi-level decomposition is implemented using chain reasoning technology to generate an atomized subtask topology structure with executable granularity; the iterative parameters of expert opinions are captured in real time through a dynamic entropy feedback mechanism, and the decomposition path weight coefficients are optimized online accordingly to achieve optimization of the decomposition mode and subtask priority, resulting in an adaptive subtask structure formed through dynamic decomposition. Step 2.3, Review Mechanism: Establish a multimodal verification process and set up expert verification loops in the dynamic task decomposition stage; Step 2.4, RL-CPM Collaborative Optimization Scheduling and Execution: Based on the obtained subtask structure, i.e. task classification, a critical path algorithm based on RL-CPM is introduced. Historical data is used to evaluate and rank the importance, urgency, or difficulty of solving subtasks, and the critical path weights are dynamically adjusted according to real-time feedback. Step 2.5, Human-Machine Verification Loop: Manual intervention is performed as needed to adjust and correct the execution of the task.

[0015] Furthermore, the implementation method of step 2.4 of the present invention includes: The implementation method of step 2.4 includes: (1) Dynamic task classification and path pre-calculation Input: Set of subtasks { Task dependency matrix Historical execution data

[0016] Tiered strategy: =α*Importance( )+β*Urgency( )+y*Difficulty( ) in: Subtasks The overall priority score is used for task scheduling and sorting; Importance Indicates based on historical execution data The importance of the task being assessed; Urgency ) indicates a task The degree of urgency; Difficulty ) represents based on the task dependency matrix The difficulty of task execution is assessed; α, β, and y are adjustable weights, α+β+y=1, and the initial weights are set to (0.4, 0.3, 0.3). Time window calculation:

[0017]

[0018]

[0019] in: Indicates task The earliest start time; Indicates task The earliest completion time; Indicates task The set of all prerequisite tasks; Indicates task The latest start time; Indicates task Total float time; Initial critical path:

[0020] (2) Dynamic critical path optimization State space:

[0021]

[0022] in: Indicates the real-time change in expert opinions; Indicates the tool execution latency rate; This indicates the current resource utilization rate; This represents the gradient of changes in the quality of the task plan; Indicates a deviation in task execution; The entropy value, representing expert consensus, reflects the degree of dispersion of group opinions; Action space:

[0023] Weight coefficient update rules:

[0024]

[0025] (4) Multi-objective Pareto optimization Objective function:

[0026]

[0027]

[0028]

[0029] in: This represents the time objective function, which is the sum of the time of all tasks on the critical path. This represents the resource objective function, which is the weighted sum of resource usage across all tasks. Represent the quality objective function; Indicates task The estimated execution time; CriticalPath represents the set of tasks on the current dynamic critical path; Indicates task Resource demand; Indicates task Resource utilization rate during execution; Indicates task The execution quality score.

[0030] The beneficial effects of this invention are: This invention achieves breakthrough improvements in multiple dimensions compared to existing technologies. A dynamic critical path optimization mechanism compresses task adjustment response time, reduces response delays to sudden priority changes, and lowers resource waste. The multi-objective decision-making module improves Pareto solution adoption rate and solution quality score while reducing multi-objective conflict rate through an improved NSGA-II algorithm and real-time expert intervention loop. The intelligent routing engine improves the success rate of abnormal switching, supports dynamic registration and millisecond-level load awareness. A hierarchical knowledge base architecture combined with hybrid retrieval technology improves semantic matching accuracy. In typical scenarios such as strategic simulation and engineering demonstration, the system supports improved expert consensus and accelerated OODA decision-making loop, overcoming technical bottlenecks such as poor dynamic adaptability, inaccurate multi-objective trade-offs, and fragile toolchains in traditional HWME systems. It provides a new generation solution for real-time collaborative decision-making in complex scenarios, possessing high robustness, strong adaptability, and flexible human-machine collaboration. Attached Figure Description

[0031] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings: Figure 1 This is an agent architecture based on a large language model, as described in this embodiment of the invention. Figure 2 This is the system architecture of an embodiment of the present invention; Figure 3 This is the dynamic task decomposition process in an embodiment of the present invention. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0033] Example 1 The AI ​​discussion system for dynamic critical path multi-objective optimization according to this invention consists of five layers: basic resource layer, technology layer, tool layer, business layer, and application layer.

[0034] 1. The basic resource layer consists of network infrastructure, server arrays, storage systems, distributed computing clusters and related hardware devices, providing the underlying physical architecture and basic technical support for the entire technology platform or application system.

[0035] 2. The technical layer integrates the core technical components and development framework supporting the operation of the comprehensive seminar system, including intelligent agent systems, graph databases, the Spring Boot framework, and core elements such as the PPO algorithm in reinforcement learning and the NSGA-II algorithm in multi-objective optimization. The intelligent agent module constructs a functional system through specialized intelligent agents for task parsing, coordination and scheduling, verification and evaluation, and policy optimization. Specifically: The task decomposition adopts a dual-modal driving mechanism of chained reasoning (CoT) and PPO reinforcement learning. It dynamically reconstructs the task network by adjusting the critical path coefficients and parallel thresholds in real time through a six-dimensional state space (task latency, resource utilization, quality gradient, expert entropy, etc.). The optimized decision fusion improved NSGA-II algorithm constructs a three-dimensional objective function (time-resource-quality), generates a Pareto optimal solution set based on the constraint violation screening mechanism, and embeds an expert correction verification loop.

[0036] 3. The tool layer establishes a unified tool management platform to enable cross-system resource registration, scheduling, and collaborative invocation. Core tools for expert decision support include data modeling and analysis engines, intelligent decision-making algorithm libraries, knowledge graph construction frameworks, and multi-dimensional simulation and deduction systems. The tool layer utilizes an intelligent routing engine to achieve on-demand scheduling of simulation and data analysis tools.

[0037] 4. The business layer carries core business functions such as expert file management and full-cycle control of seminars and conferences. Through the domain rule engine and process orchestration center, it coordinates and configures business logic to achieve the standardized operation of the entire chain of seminar tasks and decision-making processes.

[0038] 5. The application layer user system is divided into two core permission groups: domain expert users and system operation and maintenance administrators, which correspond to the core functional requirements of participating in discussions and system operation and maintenance, respectively.

[0039] In a preferred embodiment of the present invention, the five-layer architecture of the present invention is implemented through the following functional modules: The intelligent agent module of this invention can be specifically implemented as follows: a large language model serves as the core controller, which, together with a perception input module, a collaborative interaction module, a cognitive decision engine, a knowledge storage unit, and a task execution mechanism, constitutes a composite artificial intelligence agent. The agent's workflow includes instruction reception, understanding based on the large language model (LLM), knowledge base matching, task planning, action execution, and strategy optimization.

[0040] The LLM (Limited Language Management) serves as the core controller of the system, leveraging natural language understanding and generation capabilities to parse tasks and coordinate inter-module scheduling. Specifically, it decomposes complex problems into structured operational processes and coordinates the sequential operation of various functional units. The knowledge storage unit employs a distributed data warehouse architecture, integrating domain knowledge graphs, behavioral rule bases, and experience case databases to provide multi-dimensional knowledge support for system operation. The cognitive decision engine integrates deep reinforcement learning algorithms and Bayesian inference models, using a multi-source heterogeneous data processing layer to extract features and semantically associate environmental state information collected by the perception input module, forming a decision tree with temporal dependencies. The task execution mechanism includes an API interface adaptation layer and a physical operation converter, compiling decision instructions into executable code sequences and achieving precise mapping between digital space operation instructions and physical execution terminals through dynamic parameter tuning. All functional modules exchange data via an event-driven bus, forming a closed-loop feedback intelligent decision-making and execution system.

[0041] Example 2 The AI-based method for dynamic critical path multi-objective optimization according to embodiments of the present invention includes the following steps: Step 1: Construct an AI discussion system for dynamic critical path multi-objective optimization and perform knowledge base matching: The core architecture of the comprehensive discussion system's knowledge base focuses on the collaborative design of three major technical components: a multimodal knowledge classification system, a distributed storage engine, and a semantic enhancement retrieval algorithm.

[0042] 1.1 Multimodal Knowledge Classification System The multimodal knowledge classification system adopts a hierarchical storage architecture, dividing the memory system into long-term memory and short-term memory. The former uses vector databases and semantic indexing technology to achieve long-term storage and semantic retrieval of structured knowledge such as event consensus sets, algorithm rule bases, and experience case bases; the latter is based on an in-memory database to build a real-time state tracking module, integrating a time-series marking mechanism to manage dynamic data such as task execution context, meeting process snapshots, and user interaction logs.

[0043] 1.2 Distributed Storage Engine To improve retrieval efficiency, knowledge base design typically involves extracting and mapping long-term memory information such as domain knowledge and common sense, storing it in a structured, ordered, and hierarchical manner. Short-term memory, on the other hand, is stored in the model's internal state, with persistence depending on business requirements. Based on the knowledge data type, relational databases, graph databases, vector databases, and caches are selected for storage.

[0044] 1.3 Semantic Enhancement Knowledge Retrieval This system constructs a multimodal semantically enhanced knowledge retrieval system, adopting a three-layer processing architecture: 1) Indexing mechanism: By building an index structure for core data fields, the efficiency of knowledge retrieval is improved; 2) Semantic matching: Based on natural language processing and semantic understanding algorithms, the semantic association mapping between query statements and knowledge base content is completed; 3) Caching mechanism: A memory-level cache pool is used to store frequently accessed knowledge nodes, thereby optimizing query performance by reducing the backend storage I / O load.

[0045] Step 2, Task Breakdown As the core coordination hub of the integrated discussion system, the intelligent agent relies on a large language model to transform and parse unstructured requirements into atomic tasks, and improves task execution accuracy through sequential planning. However, pure model-driven task decomposition faces challenges such as rigid processes, redundant execution paths, and dynamic disturbances, such as a real-time change rate of expert decision parameters exceeding 30%. This system innovatively adopts a two-stage task optimization mechanism: the first stage generates the task topology structure through the Chain of Thought (CoT) technology and sets up an expert intervention verification mechanism to ensure logical completeness; the second stage establishes a dynamic critical path evolution model based on the deep coupling of the Proximal Policy Optimization (PPO) algorithm and the Critical Path Method (CPM). Through the construction of a time-resource constraint matrix and an entropy evaluation model (including five dimensions of parameters such as task density and resource utilization ratio), the critical path is reconstructed online, reducing the dynamic adjustment response time of the task execution sequence to within 1.2 seconds.

[0046] 2.1. Dynamic Event Perception Stage of Discussion Task: Based on the preset discussion topic, the problem task is defined, and the task boundary is intelligently identified by relying on the pre-built knowledge base system. Core constraints and semantic keywords are extracted simultaneously.

[0047] 2.2 Dynamic Task Decomposition Phase: Based on the core elements and constraints of the task, and combined with a historical knowledge base, the task decomposition phase utilizes chain-reasoning technology to implement multi-level decomposition, generating an atomized subtask topology with executable granularity. A dynamic entropy feedback mechanism captures expert opinion iteration parameters in real time (e.g., requirement change frequency > 3 times / hour), and optimizes the decomposition path weight coefficient α∈[0.2, 0.8] accordingly. The process supports dynamic parameter tuning to optimize the decomposition mode and subtask priority. An example of an adaptive subtask structure formed through dynamic decomposition in a system overview design workshop scenario is shown below.

[0048] "System Design": { "Dynamic Parameters": { "Alpha": {"Range": [0.0, 1.0], "Current": 0.7}, "Beta": {"Range": [1, 5], "Current": 3} }, "Discussion Stages": [ { Name: "Preparation Stage", "Subtasks": [ { "Number": "T1.1", "Task": "", "DynamicParams": { "AdjustHistory": ["2024-03-01: Alpha+0.1"], "CurrentWeight": 0.8 }, "Outputs": ["System Scope"] }, { "Number": "T1.2", "DynamicParams": { "AdjustHistory": [], "CurrentWeight": 0.6 } }, / / ... Add the same structure to other subtasks ] }, / / ... The structure of other stages remains unchanged. ], "Workflow Relations": { "Sequential Flow": [ { "Path": "T1.1→T1.2→T1.3→T1.4→T1.5", "TriggerCondition": "ResourceUtilization>70%" }, { "Path": "T2.1→T2.2→T2.3", "TriggerCondition": "ChangeFrequency<3" } ], "Parallel Flow": [ { "Tasks": ["T1.3", "T1.4"], "MaxParallelism": 5 } ], "Feedback Loop": [ { "Path": "T3.2→T2.2", "UpdateRule": "ExpertDivergence>0.4" } ] } } 2.3 Review Mechanism: Establish a multimodal verification process and set up an expert verification loop in the task decomposition stage. When the subtask weight adjustment amount Δ>0.3 or the constraint compliance index<85%, the manual confirmation step is triggered. The task scheduling engine can only be activated after the topology structure is iteratively optimized through dynamic parameter tuning and expert double signature mechanism.

[0049] 2.4 RL-CPM Collaborative Optimization Scheduling Execution: Based on the task classification in the previous step, a critical path algorithm based on RL-CPM is introduced. By using historical data, the importance, urgency, or difficulty of solving subtasks are evaluated and ranked. At the same time, the critical path weights are dynamically adjusted according to real-time feedback (changes in expert opinions, delays in tool execution).

[0050] RL-CPM Collaborative Optimization Scheduling Execution Process (1) Dynamic task classification and path pre-calculation Input: Set of subtasks { Task dependency matrix Historical execution data

[0051] Tiered strategy: =α*Importance( )+β*Urgency( )+y*Difficulty( ) in: Subtasks The overall priority score is used for task scheduling and sorting; Importance Indicates based on historical execution data The importance of the task being assessed; Urgency ) indicates a task The degree of urgency; Difficulty ) represents based on the task dependency matrix The difficulty of task execution is assessed; α, β, and y are adjustable weights, α+β+y=1, and the initial weights are set to (0.4, 0.3, 0.3). Time window calculation:

[0052]

[0053]

[0054] in: Indicates task The earliest start time; Indicates task The earliest completion time; Indicates task The set of all prerequisite tasks; Indicates task The latest start time; Indicates task Total float time; Initial critical path:

[0055] (2) Dynamic critical path optimization State space:

[0056]

[0057] in: Indicates the real-time change in expert opinions; Indicates the tool execution latency rate; This indicates the current utilization rate of resources (CPU / memory, etc.); Indicates the gradient (trend) of changes in the quality of the task plan; Indicates a deviation in task execution; The entropy value, representing expert consensus, reflects the degree of dispersion of group opinions; Action space:

[0058] Weight coefficient update rules:

[0059]

[0060] (5) Multi-objective Pareto optimization Objective function:

[0061]

[0062]

[0063]

[0064] in: This represents the time objective function, which is the sum of the time of all tasks on the critical path. This represents the resource objective function, which is the weighted sum of resource usage across all tasks. This represents the quality objective function (taking negative values ​​to achieve minimization). Indicates task The estimated execution time; CriticalPath represents the set of tasks on the current dynamic critical path; Indicates task Resource demand; Indicates task Resource utilization rate during execution; Indicates task Execution quality score; The AI ​​seminar system for dynamic critical path multi-objective optimization employs an agent capability matrix and resource state awareness mechanism to drive task assignment and execution sequence planning, with task-agent matching degree serving as the basis for scheduling decisions. During execution monitoring, a dynamic response engine detects abnormal events such as task state transitions and environmental parameter drifts in real time, triggering online scheduling optimization algorithms to achieve second-level response for task topology reconstruction and resource reallocation strategies.

[0065] 2.5 Human-Machine Verification Loop: In this stage, manual intervention can be carried out at any time as needed to adjust and correct the execution of the task.

[0066] Step 3: Tool Integration To enable tool calls driven by the process engine, the intelligent agent needs to bind the external tool functions to the model in advance, and perform function matching, parameter generation and service docking during the interaction process.

[0067] Taking a conference management system as an example, during the preparation stage of a seminar, the agent triggers an email notification service through automated agenda arrangement, which includes core elements such as the list of participants, meeting time and topic extracted from the expert database. This structured data is standardized and encapsulated in JSON format.

[0068] During the information preparation phase, the agent performs multi-source data fusion operations: intelligently retrieving historical cases and technical documents related to business system development from the knowledge base using natural language to SQL commands. After classification and processing, a systematic knowledge graph that experts can directly read is formed, significantly reducing the cost of manual information screening. When participants change during the meeting, the agent can dynamically optimize agenda priorities in conjunction with the evaluation model. This real-time decision-making capability relies on deep integration with toolchains such as search engines, computing engines, and calendar services. By building tool invocation middleware, the agent not only achieves efficient cross-system data flow and accurate execution of complex computational tasks, but also forms a scalable intelligent processing architecture. It provides end-to-end solutions for information retrieval, data analysis, and process optimization for multi-domain application scenarios, enabling domain experts to focus on core decision-making processes, thereby systematically improving business processing quality and knowledge collaboration efficiency.

[0069] In summary, this invention focuses on dynamic multi-objective decision-making and constructs an intelligent discussion system based on a "perception-decomposition-optimization-execution" logical chain. Its key technical implementation points include: The first aspect of this invention proposes a dynamic task decomposition mechanism. It employs a dual-channel architecture of Chain Thinking (CoT) and reinforcement learning. CoT decomposes fuzzy requirements into structured sub-task trees, including three types of dependencies: sequential flow, parallel flow, and feedback loop, ultimately generating an initial task network. Simultaneously, a PPO reinforcement learning model is constructed. Based on a six-dimensional state vector—task latency rate, CPU / memory utilization, solution quality gradient, expert participation entropy, dynamic event trigger flags, and consensus change rate—it perceives environmental changes in real time. Through a continuous action space (critical path weight coefficient α∈[0,1], parallel task threshold β∈[1,5]), the critical path is dynamically adjusted to solve the resource misallocation problem caused by traditional CPM static path planning.

[0070] The second aspect of this invention proposes a multi-objective collaborative optimization method. A three-dimensional objective function is constructed, and feasible solutions are selected by calculating constraint violation rates. The NSGA-II algorithm is improved by introducing an adaptive crossover rate and elite retention strategy to generate a Pareto optimal solution set. The multi-objective trade-offs are then visualized through an expert interactive interface.

[0071] The third aspect of this invention proposes a dynamic knowledge base management technology. A hierarchical storage architecture for long and short-term memory is designed. Long-term memory is stored using relational databases (expert information), graph databases (solution association networks), and vector databases (semantic indexes for research reports), with context states cached via Redis. An innovative expert participation entropy quantification model is introduced to dynamically adjust knowledge retrieval weights and improve the efficiency of heterogeneous data retrieval.

[0072] The fourth aspect of this invention proposes an intelligent tool routing engine. Based on standardized API interfaces, it enables dynamic registration and capability description of tools, develops a task status sensor (monitoring tool execution latency and resource utilization), and triggers tool switching decisions in conjunction with a task priority queue (sliding window mechanism).

[0073] The fifth aspect of this invention proposes a human-machine collaborative verification closed loop. A three-layer intervention mechanism is designed: the system proactively pushes path adjustment suggestions (including a visual dashboard of the adjustment basis); experts manually correct weight parameters or reject proposals; the rejection event triggers retraining of the reinforcement learning model, forming a reinforcement learning loop of "machine intelligence generation - human experience correction - model iterative evolution".

[0074] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0075] It should be understood that those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. An AI discussion system for dynamic critical path multi-objective optimization, characterized in that, The system includes: The infrastructure layer provides the underlying physical architecture and basic technical support, including: network infrastructure, server arrays, storage systems, distributed computing clusters, and hardware devices; The technical layer comprises the core technical components and development framework that support and enable the operation of the seminar system, including: an agent module, a database, the Spring Boot framework, and algorithms; the agent module is used to implement a specialized agent construction function system, including dynamic task decomposition and optimization decision-making. The tool layer is used to build a unified tool management platform and knowledge base to realize the registration, scheduling and collaborative invocation of cross-system resources; The business layer is used to implement core business functions, including expert file management and full-cycle control of seminars and conferences. The application layer is used to define core permissions for the user system, including: domain expert users and system operation and maintenance administrators, which correspond to the functional requirements of participating in discussions and system operation and maintenance, respectively.

2. The AI ​​discussion system for dynamic critical path multi-objective optimization according to claim 1, characterized in that, The method for dynamic task decomposition in the intelligent agent module includes: A dual-channel architecture combining Chained Thinking (CoT) and reinforcement learning is adopted. CoT decomposes fuzzy requirements into structured subtask trees, including three types of dependencies: sequential flow, parallel flow, and feedback loop, ultimately generating an initial task network. Simultaneously, a PPO reinforcement learning model is constructed, which uses a six-dimensional state vector based on task latency rate, CPU / memory utilization, solution quality gradient, expert participation entropy, dynamic event trigger flags, and consensus change rate to perceive environmental changes in real time and dynamically adjust the critical path through a continuous action space.

3. The AI ​​discussion system for dynamic critical path multi-objective optimization according to claim 1, characterized in that, The methods for making optimization decisions in the intelligent agent module include: We construct a three-dimensional objective function of time, resources, and quality, and use constraint violation calculation to screen feasible solutions. We improve the NSGA-II algorithm by introducing an adaptive crossover rate and elite retention strategy to generate a Pareto optimal solution set and embed an expert correction verification loop.

4. The AI ​​discussion system for dynamic critical path multi-objective optimization according to claim 1, characterized in that, The knowledge base employs dynamic management technology and a hierarchical storage architecture with long short-term memory. Long-term memory uses a relational database to store expert information, a graph database for solution association networks, and a vector database for semantic indexing of research reports, all stored in a categorized manner. Short-term memory uses Redis to cache context states and introduces an expert participation entropy quantification model to dynamically adjust knowledge retrieval weights.

5. The AI ​​discussion system for dynamic critical path multi-objective optimization according to claim 1, characterized in that, The tool layer employs a tool intelligent routing engine, which uses standardized API interfaces to achieve dynamic registration and capability description of tools. A task status sensor is developed to monitor tool execution latency and resource utilization, and a task priority queue and sliding window mechanism are combined to trigger tool switching decisions.

6. An AI discussion method for dynamic critical path multi-objective optimization, employing the AI ​​discussion system for dynamic critical path multi-objective optimization as described in any one of claims 1-5, characterized in that, The method includes the following steps: Step 1: Construct an AI discussion system for dynamic critical path multi-objective optimization based on a multi-layered framework; build a knowledge base using a multimodal knowledge classification system, a distributed storage engine, and a semantically enhanced retrieval algorithm; Step 2: Input the discussion task, perceive the dynamic events of the discussion task, and extract constraints and semantic keywords; adopt a two-stage task optimization mechanism: the first stage generates the task topology structure through thinking chain technology and sets up an expert intervention verification mechanism to ensure logical completeness; the second stage combines the near-end strategy optimization PPO algorithm and the critical path method CPM to establish a dynamic critical path evolution model, realize dynamic task classification and path reconstruction. Step 3: Integrate the system, model and external tools to realize tool calls driven by the process engine, and perform function matching, parameter generation and service docking during the interaction process.

7. The AI ​​discussion method for dynamic critical path multi-objective optimization according to claim 6, characterized in that, The implementation method of step 1 includes: Step 1.1, Multimodal Knowledge Classification System: A hierarchical storage architecture is adopted, dividing the memory system into long-term memory and short-term memory; long-term memory realizes the long-term storage and semantic retrieval of structured knowledge such as event consensus set, algorithm rule base, and experience case base through vector database and semantic indexing technology; short-term memory builds a real-time status tracking module based on memory database, and integrates time-series marking mechanism to manage dynamic data such as task execution context, meeting process snapshot and user interaction log; Step 1.2, Distributed Storage Engine: Extracts and maps long-term memory, storing it in an ordered and hierarchical manner in a structured way; short-term memory is stored in an in-memory database, and whether to persist it is selected according to business needs; based on the type of knowledge data, relational database, graph database, vector database, and cache are selected for storage respectively. Step 1.3, Semantic Enhancement Retrieval Algorithm, including a three-layer processing architecture: Indexing mechanism, which builds an index structure for core data fields; Semantic matching, which completes the semantic association mapping between query statements and knowledge base content based on natural language processing and semantic understanding algorithms; Caching mechanism, which uses a memory-level cache pool to store frequently accessed knowledge nodes, thereby optimizing query performance by reducing the backend storage I / O load.

8. The AI ​​discussion method for dynamic critical path multi-objective optimization according to claim 6, characterized in that, The implementation method of step 2 includes: Step 2.1, Dynamic Event Perception Stage of Discussion Task: For the discussion task, the problem task is defined based on the preset discussion topic, the task boundary is intelligently identified according to the knowledge base, and the core constraints and semantic keywords are extracted simultaneously. Step 2.2, Dynamic Task Decomposition Stage: Based on core constraints and semantic keywords, combined with historical data from the knowledge base, multi-level decomposition is implemented using chain reasoning technology to generate an atomized subtask topology structure with executable granularity; the iterative parameters of expert opinions are captured in real time through a dynamic entropy feedback mechanism, and the decomposition path weight coefficients are optimized online accordingly to achieve optimization of the decomposition mode and subtask priority, resulting in an adaptive subtask structure formed through dynamic decomposition. Step 2.3, Review Mechanism: Establish a multimodal verification process and set up expert verification loops in the dynamic task decomposition stage; Step 2.4, RL-CPM Collaborative Optimization Scheduling and Execution: Based on the obtained subtask structure, i.e. task classification, a critical path algorithm based on RL-CPM is introduced. Historical data is used to evaluate and rank the importance, urgency, or difficulty of solving subtasks, and the critical path weights are dynamically adjusted according to real-time feedback. Step 2.5, Human-Machine Verification Loop: Manual intervention is performed as needed to adjust and correct the execution of the task.

9. The AI ​​discussion method for dynamic critical path multi-objective optimization according to claim 8, characterized in that, The implementation method of step 2.4 includes: (1) Dynamic task classification and path pre-calculation Input: Set of subtasks { Task dependency matrix Historical execution data Tiered strategy: =α*Importance( )+β*Urgency( )+y*Difficulty( ) in: Subtasks The overall priority score is used for task scheduling and sorting; Importance Indicates based on historical execution data The importance of the task being assessed; Urgency ) indicates a task The degree of urgency; Difficulty ) represents based on the task dependency matrix The difficulty of task execution is assessed; α, β, and y are adjustable weights, α+β+y=1, and the initial weights are set to (0.4, 0.3, 0.3). Time window calculation: in: Indicates task The earliest start time; Indicates task The earliest completion time; Indicates task The set of all prerequisite tasks; Indicates task The latest start time; Indicates task Total float time; Initial critical path: (2) Dynamic critical path optimization State space: in: Indicates the real-time change in expert opinions; Indicates the tool execution latency rate; This indicates the current resource utilization rate; This represents the gradient of changes in the quality of the task plan; Indicates a deviation in task execution; The entropy value, representing expert consensus, reflects the degree of dispersion of group opinions; Action space: Weight coefficient update rules: (3) Multi-objective Pareto optimization Objective function: in: This represents the time objective function, which is the sum of the time of all tasks on the critical path. This represents the resource objective function, which is the weighted sum of resource usage across all tasks. Represent the quality objective function; Indicates task The estimated execution time; CriticalPath represents the set of tasks on the current dynamic critical path; Indicates task Resource demand; Indicates task Resource utilization rate during execution; Indicates task The execution quality score.

Images