Intelligent iterative execution system based on bionic architecture and execution method thereof

Abstract

A biomimetic architecture-based intelligent iterative execution system and method are disclosed. The system includes an upper-layer main unit, a knowledge acquisition module, a security isolation module, and multiple lower-layer main units. This invention acquires external instructions through the active interaction unit of the upper-layer main unit, pre-generates a target result screen, and uses it as a unified business execution benchmark screen for the entire system. During task execution, the tool execution module plans the execution path based on this target result screen. Throughout the execution process, the benchmark screen is used as a verification standard to perform real-time numerical difference calculations between the real-time execution result and the target result screen, generating a deviation vector and continuously iterating and converging. This effectively solves the industry pain points of uncontrollable results, lack of real-time closed-loop correction, and lack of intelligent iterative capabilities during complex task execution, achieving high-precision, autonomously iterative, and highly stable intelligent task execution.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence system architecture technology, specifically to an intelligent iterative execution system and method based on a hierarchical hardware architecture and equipped with a partitioned, physically isolated knowledge base. Background Technology

[0002] With the rapid development of artificial intelligence technology, intelligent agent systems have been widely used in various fields. However, existing intelligent agent systems suffer from the following technical challenges: First, there is a lack of reliable intelligent agent architecture. Traditional intelligent agent systems often adopt a monolithic architecture with high coupling between functional modules, making it difficult to achieve flexible functional expansion and module replacement. When the system needs to introduce new functions or adapt to new scenarios, the entire system often needs to be refactored, resulting in high system maintenance costs and long iteration cycles.

[0003] Second, continuous learning and knowledge retention are difficult to balance. Existing intelligent agent systems generally suffer from the problem of "catastrophic forgetting" during continuous learning, meaning that learned knowledge is lost or overwritten when learning new knowledge. This prevents intelligent agents from effectively accumulating and utilizing knowledge in dynamic environments, resulting in low knowledge utilization efficiency.

[0004] Third, there are shortcomings in security and explainability. The decision-making process of traditional intelligent agent systems lacks transparency, making it difficult to trace the basis for decisions and the reasoning process, which poses security risks in critical application scenarios. At the same time, the system lacks a deep understanding of causal relationships and the ability to reason counterfactually, making it unable to effectively assess and predict risks in complex scenarios.

[0005] To address the aforementioned technical issues, a novel intelligent agent architecture is urgently needed to resolve key technical challenges such as reliability, continuous learning, and security, controllability, and explainability. Summary of the Invention

[0006] The purpose of this invention is to provide an intelligent iterative execution system and method based on a biomimetic architecture to solve the technical problems of poor architectural reliability, difficulty in continuous learning, and insufficient security, controllability and interpretability in existing intelligent agent systems.

[0007] The present invention is achieved through the following technical solution: an intelligent iterative execution system based on a biomimetic architecture, characterized in that it includes: an upper-layer main unit, a knowledge acquisition module, a security isolation module, and multiple lower-layer main units; The upper-layer main unit and the lower-layer main unit each include a central core processing unit, a partitioned knowledge base module, an active interaction unit, and a tool execution module; the partitioned knowledge base modules of the upper-layer main unit and the lower-layer main unit each include a knowledge storage area and a scene cognition area; the partitioned knowledge base module of the upper-layer main unit also includes a context memory area and a task scheduling area. The active interaction units within the upper-layer main unit and each lower-layer main unit are connected to the central core processing unit; the tool execution module is connected to the partitioned isolated knowledge base module; the partitioned isolated knowledge base module is connected to the active interaction unit; and the tool execution module is connected to the central core processing unit. The central core processing unit is used to evaluate the complexity of the task and the matching degree of the partitioned and isolated knowledge base module. When the independent execution condition is met, the result is generated directly based on the local knowledge base. When the independent execution condition is not met, the execution instruction is sent to the lower-level main unit, and the execution result is compared with the target result screen. If they are inconsistent, the correction mechanism is triggered. At the same time, it is responsible for knowledge accumulation and iterative updates. The partitioned and isolated knowledge base module is used to store static rules, historical scenarios, behavioral strategies and task scheduling data, to achieve physical isolation and tamper-proof storage of knowledge, and to support backtracking analysis and knowledge iteration. The active interaction unit is used to acquire external instructions, generate a target result screen, and use the target result screen as a unified business execution benchmark screen; The tool execution module is used to determine the planned execution path and, during the execution process, uses the target result screen as a benchmark to perform real-time numerical difference calculation between the real-time execution result and the target result screen, generating a converged deviation vector. The active interaction unit of the upper-layer main unit is connected to the knowledge acquisition module, which is used to selectively capture necessary external information to supplement the knowledge base; the knowledge acquisition module is connected to the security isolation module, which is a one-way data channel, where data only enters and does not exit, and performs compliance and security reviews on the input data to prevent malicious data injection; the security isolation module is connected to the active interaction unit of the lower-layer main unit.

[0008] Furthermore, the upper-layer main unit also includes an identity authentication module for performing chip-level authorization verification, and unlocking the system after successful verification; the identity authentication module is connected to the active interaction unit.

[0009] Furthermore, the upper-layer main unit and the lower-layer main unit also include a causal inference module, which is used to construct counterfactual hypothetical scenarios, deduce the potential results of different execution paths, and generate causal risk scores to assist decision-making; the causal inference module is connected to the central core processing unit, the tool execution module, the active interaction unit and the partitioned isolated knowledge base module respectively.

[0010] Furthermore, the system also includes a parallel output control module, which integrates the analysis results of multiple lower-level main units and selects the optimal solution or performs multi-dimensional fusion based on the task target result screen type; the parallel output control module is connected to the active interaction units of multiple lower-level main units and the central core processing unit of the upper-level main unit respectively.

[0011] Furthermore, the number of lower-level main units is six, and each lower-level main unit adopts a different logical processing method, including inductive logic, deductive logic, divergent thinking, convergent thinking, reverse thinking, and dialectical thinking.

[0012] An execution method for an intelligent iterative execution system based on a biomimetic architecture, characterized by comprising the following steps: (1) The active interaction unit of the upper main unit receives external instructions, generates the target result screen, and uses the target result screen as the unified business execution benchmark screen; (2) The central core processing unit assesses the complexity of the current task and the matching degree between it and the partitioned knowledge base module; (3) If the current task is determined to meet the independent execution conditions, the upper-level main unit directly generates the result based on the local knowledge base, skipping the calculation steps of the lower-level main unit; (4) If it is determined that the current task does not meet the independent execution conditions, the deep thinking mode is triggered: the central core processing unit issues an execution instruction and starts multiple lower-level main units to receive the unified task, and each uses inductive logic, deductive logic, divergent thinking, convergent thinking, reverse thinking and dialectical thinking to perform independent logical verification and calculation processing. (5) After receiving the instruction, the tool execution module determines the planned execution path and starts execution based on the target result screen; (6) During the execution process, the tool execution module performs real-time numerical difference calculation between the real-time execution result and the target result screen to generate a deviation vector; if the deviation vector does not converge to the preset threshold range, the execution parameters are adjusted and bypass optimization processing is triggered until the deviation vector converges. (7) The central core processing unit compares the execution results. If the results are consistent, it outputs the final result and stores the entire process data in the knowledge storage area and scenario cognition area of ​​the partitioned isolated knowledge base module to complete the knowledge iteration. (8) When the partitioned isolated knowledge base module of the upper main unit has knowledge gaps, the knowledge acquisition module selectively captures necessary external information to fill the gaps, and distributes it to the lower main unit after review by the security isolation module. (9) The central core processing unit of the upper-level main unit receives the best integrated result, writes the new knowledge into the partitioned isolated knowledge base module, and copies and distributes the updated knowledge data to all lower-level main unit knowledge base modules to achieve the same source and synchronization of knowledge between the upper and lower layers, and completes one iteration.

[0013] Furthermore, after the system is powered on, the identity authentication module completes chip-level authorization verification; the identity authentication module adopts a chip-level tamper-proof design; the active interaction unit unlocks the system after the identity authentication module verifies the identity, and the system enters a standby state.

[0014] Furthermore, during the planning of execution paths or the triggering of correction mechanisms, the causal inference module constructs counterfactual hypothetical scenarios based on historical scenario records and static rule data in the partitioned isolated knowledge base module; the causal inference module infers the potential results of different execution paths and generates a causal risk score; the central core processing unit adjusts the planned execution path or corrects the execution parameters based on the causal risk score.

[0015] Furthermore, the calculation results of multiple lower-level main units are uniformly transmitted to the parallel output control module; the parallel output control module reads and parses the original target result screen; if the target result screen is a multi-dimensional composite task, the parallel output control module starts the full-dimensional integration mode, calling all calculation results from multiple channels for fusion and splicing; if the target result screen is a single-directional task, the parallel output control module starts the single-point optimization mode, horizontally comparing multiple output results and selecting the effective result with the highest matching degree; if there are discrepancies among multiple results, the central core processing unit of the upper-level main unit starts the dynamic weight arbitration mechanism, dynamically allocating the weight coefficients of logical dimensions according to the type of the task target result screen, and calculating the weighted optimal solution.

[0016] The beneficial effects and features of this invention are as follows: (1) Reliable intelligent agent architecture: This invention adopts a hierarchical bionic architecture, with the upper-layer main unit responsible for global planning, task scheduling, and knowledge management, and the lower-layer main unit responsible for specific execution. Each module has a clear function and boundaries, and communicates through standardized interfaces. This architecture design improves the maintainability and scalability of the system and reduces the complexity of functional iteration.

[0017] (2) Continuous learning without forgetting: This invention uses a partitioned and isolated knowledge base design to store different types of knowledge in mutually independent areas. When learning new knowledge, only the content of the corresponding area is updated, without affecting the knowledge already stored in other areas. At the same time, through the knowledge iteration and update mechanism, the system can continuously accumulate and optimize knowledge, effectively avoiding the problem of "catastrophic forgetting".

[0018] (3) Secure, controllable, and explainable: This invention ensures the security of system access through an identity authentication module and enhances the transparency of the decision-making process through a causal inference module. The design of six parallel logical processing functions enables the system to comprehensively analyze problems, verify decision-making basis from multiple perspectives, and improve the reliability and security of the system in practical application scenarios. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the overall system architecture of the present invention; Figure 2 This is a schematic diagram of the architecture of the lower-level main unit in this invention.

[0020] Explanation of the attached diagram numbers: 1. Central Core Processing Unit; 3. Partitioned Knowledge Base Module; 4. Identity Authentication Module; 5. Security Isolation Module; 10. Active Interaction Unit; 11. Tool Execution Module; 17. Parallel Output Control Module; 18. Knowledge Acquisition Module; 22. Causal Inference Module.

[0021] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, and the other drawings obtained are all within the protection scope claimed by the present invention. Detailed Implementation

[0022] The following combination Figure 1-2 The present invention will be described in detail through specific embodiments. Example

[0023] The intelligent iterative execution system provided in this embodiment includes an upper-layer main unit, a knowledge acquisition module 18, a security isolation module, and multiple lower-layer main units.

[0024] The internal architecture of the upper-level main unit includes: Central Core Processing Unit 1: The core computing and control unit of the system, used to evaluate the complexity of tasks and the matching degree of partitioned and isolated knowledge base modules. When the independent execution conditions are met, it directly generates results based on the local knowledge base; when the independent execution conditions are not met, it issues execution instructions to the lower-level main unit and compares the execution results with the target result screen. If they are inconsistent, it triggers a correction mechanism. It is also responsible for knowledge accumulation and iterative updates. Partitioned and isolated knowledge base module 3: includes a knowledge storage area, a scene cognition area, a context memory area, and a task scheduling area. It is used to store static rules, historical scenes, behavioral strategies, and task scheduling data, achieving physical isolation and tamper-proof storage of knowledge, and supporting backtracking analysis and knowledge iteration. Active interaction unit 10: used to acquire external instructions, generate target result screen, and use the target result screen as a unified business execution benchmark screen; Tool execution module 11: used to determine the planned execution path, and during the execution process, based on the target result screen, to perform real-time numerical difference calculation between the real-time execution result and the target result screen, and generate a converged deviation vector. Identity authentication module 4: Used for chip-level authorization verification. Once the verification is successful, the system lock is released.

[0025] The module connection relationships are as follows: Within the upper-level main unit and the lower-level main unit: the active interaction unit is connected to the central core processing unit; the tool execution module is connected to the central core processing unit; the tool execution module is connected to the partitioned and isolated knowledge base module; and the partitioned and isolated knowledge base module is connected to the active interaction unit.

[0026] Cross-level connections: The active interaction unit of the upper-level main unit is connected to the knowledge acquisition module; the knowledge acquisition module is connected to the security isolation module 5; the security isolation module is connected to the active interaction unit of the lower-level main unit.

[0027] Multiple logic processing methods In this embodiment, the six lower-level main units each adopt different logical processing methods, including inductive logic, deductive logic, divergent thinking, convergent thinking, reverse thinking, and dialectical thinking.

[0028] Execution method steps

[0029] (1) The active interaction unit of the upper main unit receives external instructions, generates the target result screen, and uses the target result screen as the unified business execution benchmark screen; (2) The central core processing unit assesses the complexity of the current task and the matching degree between it and the partitioned knowledge base module; (3) If the current task is determined to meet the independent execution conditions, the upper-level main unit directly generates the result based on the local knowledge base, skipping the calculation steps of the lower-level main unit; (4) If it is determined that the current task does not meet the independent execution conditions, the deep thinking mode is triggered: the central core processing unit issues an execution instruction and starts multiple lower-level main units to receive the unified task, and each uses inductive logic, deductive logic, divergent thinking, convergent thinking, reverse thinking and dialectical thinking to perform independent logical verification and calculation processing. (5) After receiving the instruction, the tool execution module determines the planned execution path and starts execution based on the target result screen; (6) During the execution process, the tool execution module performs real-time numerical difference calculation between the real-time execution result and the target result screen to generate a deviation vector; if the deviation vector does not converge to the preset threshold range, the execution parameters are adjusted and bypass optimization processing is triggered until the deviation vector converges. (7) The central core processing unit compares the execution results. If the results are consistent, it outputs the final result and stores the entire process data in the knowledge storage area and scenario cognition area of ​​the partitioned isolated knowledge base module to complete the knowledge iteration. (8) When the partitioned isolated knowledge base module of the upper main unit has knowledge gaps, the knowledge acquisition module selectively captures necessary external information to fill the gaps, and distributes it to the lower main unit after review by the security isolation module. (9) The central core processing unit of the upper-level main unit receives the best integrated result, writes the new knowledge into the partitioned isolated knowledge base module, and copies and distributes the updated knowledge data to all lower-level main unit knowledge base modules to achieve the same source and synchronization of knowledge between the upper and lower layers and complete a knowledge iteration.

[0030] This embodiment also implements a dual-path iterative architecture that allows the main task line and the main learning line to run in parallel.

[0031] The main task line responds to external commands in real time, sequentially passing through the active interaction unit, the partitioned and isolated knowledge base module, the tool execution module, and the central core processing unit to complete knowledge matching, calculation, and output. When a user inputs an external command through the active interaction unit, the command directly enters the main task line: the active interaction unit stores the command in the task scheduling area and identifies the intent as "execution mode"; then it proceeds to the knowledge storage area and the scene cognition area for knowledge matching; if complete and valid knowledge is matched, the corresponding execution parameters are directly extracted; if the knowledge is insufficient, the knowledge acquisition module is automatically scheduled to retrieve and supplement it from external sources. The central core processing unit evaluates the task complexity and matching degree: if the independent execution conditions are met, the result is directly generated based on the local knowledge base; if not, a deep thinking mode is triggered, an execution command is issued, and the lower-level main unit and the tool execution module determine and execute the planned execution path. During execution, real-time numerical difference calculations are continuously performed, generating a deviation vector and converging until the task is completed. The entire main task line is completed in milliseconds, without waiting for any background learning tasks.

[0032] The learning thread is used to asynchronously capture unfamiliar objects, non-compliant states, or information outside the knowledge boundary encountered during system execution. It automatically generates conditional tasks and stores them in the conditional task scheduling area. When the system is under low load, it is awakened, and the knowledge acquisition module supplements the missing knowledge. This knowledge is then parsed through multiple lower-level main units via multiple channels, and finally, the central core processing unit completes the knowledge internalization and solidification. Finally, the new knowledge is synchronized down to all lower-level main units. Specifically, it includes the following steps: Anomaly detection: When the active interaction unit or tool execution module detects an unfamiliar object or the deviation between the execution result and the target result exceeds a preset threshold and there is no corresponding parsing rule in the knowledge base, a conditional task package is generated, the write interface of the conditional task scheduling area is called, and the task status is set to pending. Low-load wake-up: When the system is idle, the central core processing unit polls the conditional task scheduling area, extracts pending tasks, and attempts to parse them using existing knowledge in the knowledge storage area and the scene cognition area. If the confidence level is lower than the threshold, a collection request form is generated. External supplementary data collection: The knowledge collection module retrieves missing information from external sources based on the collection request form, and then sends it back after review by the security isolation module; Multi-channel parsing: The active interaction unit breaks down the supplementary data into standardized task packages and distributes them to multiple lower-level master units. Each lower-level master unit independently performs logical decomposition, structural restoration, feature parsing and verification to form multi-channel parsed data and upload it to the central core processing unit. Knowledge internalization and synchronization: The central core processing unit comprehensively calls the historical scene records and static rule data in the partitioned isolated knowledge base module, verifies and extracts rules from the multi-path parsing data, and writes the new knowledge that has passed the verification into the partitioned isolated knowledge base module and synchronizes it down to the partitioned isolated knowledge base modules of all lower-level main units. Example

[0033] The identity authentication module is located within the upper-layer main unit and connects to the active interaction unit. It is used for chip-level authorization verification. The identity authentication module is the system's security entry point; only after successful identity authentication can the system be unlocked and enter a standby state.

[0034] The identity authentication module adopts a chip-level immutable design, and its specific implementation methods include: Hardware circuit lock: The core verification logic of the identity authentication module is embedded in the hardware circuit, and the verification algorithm and key are stored in the physical security area of ​​the chip, which cannot be accessed or modified by software; Physically isolated storage: Sensitive information such as keys and certificates required for identity authentication are stored in a separate physically isolated storage area, completely isolated from the system's main storage; Anti-tampering mechanism: The chip integrates an anti-tampering detection circuit. Once a physical attack or unauthorized access attempt is detected, a security lock is immediately triggered, permanently disabling the authentication function. Cannot be disabled via software: The authentication function of the identity authentication module is enforced by hardware circuitry, and even system administrators cannot disable or bypass identity authentication via software commands.

[0035] (1) After the system is powered on, the identity authentication module automatically enters the verification state and waits for the authorization credential to be entered; (2) The identity authentication module receives the authorization credential and performs chip-level authorization verification; (3) The identity authentication module transmits the verification result to the active interaction unit; (4) If the verification is successful, the active interaction unit unlocks the system and the system enters standby mode, which can receive and process external commands; (5) If the verification fails, the system remains locked, refuses any external command input, and records the verification failure event.

[0036] This chip-level immutable design ensures security during system startup, preventing unauthorized access and malicious attacks. The authentication module's verification process is entirely controlled by hardware circuitry and cannot be bypassed by software logic, providing reliable security for the entire intelligent iterative execution system. Example

[0037] The causal inference module 22 is deployed in both the upper-layer main unit and the lower-layer main unit. Its hardware architecture and connection relationship are as follows: Connected to the central core processing unit: used to receive decision requests and return causal analysis results; Connects to the tool execution module: used to acquire execution process data; Connect to the active interaction unit: used to obtain task information; Connect to the partitioned, isolated knowledge base module: used to read historical scene records and static rule data to construct a causal graph.

[0038] Constructing counterfactual hypothetical scenarios: Based on historical scenario records and static rule data, simulate different execution paths and decision branches; Speculate on potential outcomes: Speculate on each hypothetical scenario and predict the potential outcomes of different execution paths; Generate a causal risk score: comprehensively assess the risk level of each execution path and generate a quantitative causal risk score; Decision support: Provide decision suggestions to the central core processing unit to help it select the optimal execution path or adjust parameters.

[0039] (1) The causal inference module receives call requests from the central core processing unit, the tool execution module, or the active interaction unit; (2) The causal inference module reads historical scene records and static rule data related to the current task from the partitioned isolated knowledge base module; (3) Based on historical scenarios and static rules, the causal inference module constructs multiple counterfactual hypothetical scenarios, each scenario representing a possible execution path or decision branch; (4) The causal inference module performs a deduction for each counterfactual hypothetical scenario and calculates the possible results and probabilities of the path; (5) The causal inference module comprehensively considers multiple factors such as the success rate of the result, risk cost, and time consumption, and generates a causal risk score for each execution path; (6) The causal inference module transmits the causal risk score along with the detailed analysis results of each path to the caller; (7) The central core processing unit adjusts the planning execution path or modifies the execution parameters based on the causal risk score.

[0040] The causal inference module adopts a bypass enhancement mechanism, does not participate in daily interactions and main business process flow, and only serves as an independent backend computing unit to provide intelligent enhancement support for the frontend core functional modules to be called on demand.

[0041] Asynchronous monitoring: The causal inference module monitors the task screens, execution process data, and arbitration results within the partitioned isolated knowledge base module in real time, continuously building and updating the global causal relationship graph.

[0042] On-demand intervention: The causal inference module provides multiple invocation scenarios: Active Interaction Unit Invocation: When the active interaction unit faces multiple execution path choices or ambiguous parameter definitions during the process of generating the task screen, it can invoke the causal inference module to conduct counterfactual inference based on the global causal graph and provide feedback on the optimal execution path or standardized default parameter suggestions. Tool execution module call: When the tool execution module involves high-risk operation scenarios during execution, it can call the causal inference module to perform real-time causal logic verification, provide feedback on operation feasibility assessment scores and correct configuration parameters; Central core processing unit call: When the central core processing unit detects a slight deviation between the actual result and the target result during the comparison process, and cannot independently determine the compliance of the deviation, it can call the causal inference module to conduct causal consistency analysis and output the assessment conclusion of the degree of influence of the deviation.

[0043] Knowledge Accumulation: The causal inference module writes the causal inference conclusions and counterfactual simulation data generated from each intervention into the verification area within the partitioned and isolated knowledge base module. The central core processing unit reviews and verifies these data periodically, and once verified, they are solidified and transformed into the system's general common sense reserves.

[0044] Core principle: The intervention of the causal inference module will not block or interfere with the normal operation of the main business process, but only provide reference optimization suggestions; the core modules such as the proactive interaction unit, tool execution module, and central core processing unit have independent decision-making authority and can choose to adopt or ignore auxiliary suggestions, ensuring that the core business channels operate in a simple, stable, efficient and controllable manner. Example

[0045] The parallel output control module 17 is the core coordination unit in the system, and its hardware architecture and connection relationships are as follows: Connect to the active interaction units of multiple lower-level master units: Establish communication connections with the active interaction units of each lower-level master unit to receive the calculation results of each lower-level master unit. Connected to the central core processing unit of the upper-level main unit: used to transmit the integrated results to the upper-level main unit for final processing.

[0046] Integrate the analysis results of multiple lower-level main units: Collect the calculation results generated by multiple lower-level main units using different logical processing methods in a unified manner; Select the optimal solution based on the type of result screen for the task objective: For different types of tasks, adopt the corresponding result selection strategy; Multi-dimensional fusion: For complex tasks, multiple results are merged and combined to generate a comprehensive solution.

[0047] (1) The calculation results of multiple lower-level master units are uniformly transmitted to the parallel output control module; (2) The parallel output control module reads and parses the original target result screen to identify the task type; (3) If the target result screen is a multi-dimensional composite task, the parallel output control module starts the full-dimensional integration mode, calls all the calculation results from multiple channels to merge and splice them, and generates a comprehensive solution; (4) If the target result screen is a single-direction task, the parallel output control module starts the single-point optimization mode, compares the multiple output results horizontally, and selects the effective result with the highest matching degree as the final output; (5) If there are discrepancies in the results of multiple paths and it is difficult to directly determine the superiority or inferiority, the parallel output control module will report the discrepancies to the central core processing unit of the upper-level main unit and start the dynamic weight arbitration mechanism. (6) The central core processing unit of the upper-level main unit dynamically allocates the weight coefficients of each logical dimension according to the task target result screen type, and calculates the weighted optimal solution as the arbitration result; (7) The parallel output control module transmits the final integrated result to the central core processing unit of the upper main unit to complete the current round of operation.

[0048] During collaborative computing, each of the lower-level master units undertakes different logical processing responsibilities: Inductive logic: Responsible for extracting general patterns from massive amounts of data, identifying common features, forming regularity cognition, and outputting common conclusions; Deductive logic: responsible for deriving specific execution conclusions based on fundamental principles, performing logical reasoning from general to specific, and outputting the derived conclusions; Divergent thinking: Responsible for multi-dimensional deduction, generating multiple feasible solutions, diverging from one point to multiple possibilities, and outputting diverse solutions; Convergent thinking: Responsible for converging multiple solutions, focusing on the optimal and unique solution, and outputting a convergent conclusion; Reverse thinking: responsible for deriving the preconditions and execution path from the final result and outputting the reverse analysis results; Dialectical thinking: responsible for analyzing the positive and negative logic of events, contradictory relationships and conditions of unity, weighing the advantages and disadvantages, and outputting dialectical analytical conclusions.

[0049] When there are discrepancies in the calculation results of multiple lower-level main units, the central core processing unit of the upper-level main unit initiates a dynamic weight arbitration mechanism to dynamically adjust the weight coefficients of each logical dimension according to the task objective result screen type and calculate the weighted optimal solution.

[0050] For engineering / manufacturing tasks: increase the weight of deductive logic, convergent thinking, and reverse thinking; correspondingly decrease the weight of divergent thinking and dialectical thinking.

[0051] For creative / strategy tasks: increase the weight of divergent thinking and dialectical thinking; correspondingly decrease the weight of deductive logic and convergent thinking.

[0052] For purely theoretical / mathematical tasks: increase the weight of inductive logic and deductive logic; correspondingly decrease the weight of divergent thinking and dialectical thinking.

[0053] Weighting calculation method: After determining the weight coefficients of each logical dimension based on the type of the result screen of the task objective, the central core processing unit performs a weighted calculation on the multi-path processing results: Weighted total score = Σ (scores for each result × weight of the corresponding logical dimension) The result with the highest weighted total score is output as the optimal solution. If the difference between the highest and second-highest scores is within a preset threshold, the system can simultaneously output multiple near-optimal candidate results for further manual verification.

[0054] In the software implementation, the upper-layer main unit and the lower-layer main unit can run in different virtual machines or containers on the same physical server. The "dedicated hardware data path" can be implemented through shared memory or a virtual bus, and the functions of each module are implemented by software programs, which also fall within the scope of this invention. First path iteration: The complete process when the upper-layer main unit's knowledge base is incomplete. When the partitioned and isolated knowledge base module of the upper-level main unit has knowledge gaps, the following process is executed: (1) The active interaction unit of the upper main unit receives external instructions, generates the target result screen, and uses the target result screen as the unified business execution benchmark screen; (2) The central core processing unit assesses the complexity of the current task and the matching degree between it and the partitioned knowledge base module; (3) If the current task is determined to meet the independent execution conditions, the upper-level main unit directly generates the result based on the local knowledge base, skipping the calculation steps of the lower-level main unit; (4) If it is determined that the current task does not meet the independent execution conditions, the deep thinking mode is triggered: The central core processing unit issues an execution instruction and starts multiple lower-level main units to receive the unified task, and each uses inductive logic, deductive logic, divergent thinking, convergent thinking, reverse thinking and dialectical thinking to perform independent logical verification and calculation processing. (5) After receiving the instruction, the tool execution module determines the planned execution path and starts execution based on the target result screen; (6) During the execution process, the tool execution module performs real-time numerical difference calculation between the real-time execution result and the target result screen to generate a deviation vector; if the deviation vector does not converge to the preset threshold range, the execution parameters are adjusted and bypass optimization processing is triggered until the deviation vector converges. (7) The central core processing unit compares the execution results. If the results are consistent, it outputs the final result and stores the entire process data in the knowledge storage area and scenario cognition area of ​​the partitioned isolated knowledge base module to complete the knowledge iteration. (8) When the partitioned isolated knowledge base module of the upper main unit has knowledge gaps, the knowledge acquisition module selectively captures necessary external information to fill the gaps, and distributes it to the lower main unit after review by the security isolation module. (9) The central core processing unit of the upper-level main unit receives the best integrated result, writes the new knowledge into the partitioned isolated knowledge base module, and copies and distributes the updated knowledge data to all lower-level main unit knowledge base modules to achieve the same source and synchronization of knowledge between the upper and lower layers and complete a knowledge iteration.

[0055] When the partitioned and isolated knowledge base module of the upper-level main unit has knowledge gaps, the following process is executed: (1) The active interaction unit of the upper main unit transmits the task requirements to the knowledge acquisition module; (2) The knowledge acquisition module captures necessary external information according to task requirements, completes the knowledge base, and sends the data to the security isolation module for compliance and security review; (3) The security isolation module is a one-way data channel, with data only going in and not out. It performs compliance and security reviews on the input data to prevent malicious data injection. (4) After the data is approved, it is distributed to the active interaction units of multiple lower-level main units; (5) Multiple lower-level main units each receive the same task and use inductive logic, deductive logic, divergent thinking, convergent thinking, reverse thinking and dialectical thinking to perform independent logical verification and calculation. (6) The calculation results of multiple lower-level main units are uniformly transmitted to the parallel output control module; the parallel output control module reads and parses the original target result screen; if the target result screen is a multi-dimensional composite task, the parallel output control module starts the full-dimensional integration mode and calls all the calculation results of multiple channels for fusion and splicing; if the target result screen is a single-directional task, the parallel output control module starts the single-point optimization mode, compares the multiple output results horizontally, and selects the effective result with the highest matching degree; if there is a discrepancy among the multiple results, the central core processing unit of the upper-level main unit starts the dynamic weight arbitration mechanism, dynamically allocates the weight coefficient of the logical dimension according to the type of the task target result screen, and calculates the weighted optimal solution; (7) The central core processing unit of the upper-level main unit receives the best integrated result, writes the new knowledge into the partitioned isolated knowledge base module, and copies and distributes the updated knowledge data to all lower-level main unit knowledge base modules to achieve the same source and synchronization of knowledge between the upper and lower layers and complete a knowledge iteration.

[0056] It should be noted that the second path iteration and the first path iteration are essentially the same process, both following the step sequence defined in claim 6. The second path iteration focuses on describing the "knowledge acquisition → distribution → lower-level operation → integration → synchronization" process, but its core steps (evaluating matching degree, judging independent execution conditions, triggering deep thinking mode, etc.) are consistent with the first path iteration, and will not be repeated here.

[0057] The task mainline and the learning mainline together form a global closed-loop iteration, including the following three iteration modes:

[0058] When the conditions for independent execution are not met, a deep thinking mode is triggered, in which multiple lower-level main units independently calculate the same task, using inductive logic, deductive logic, divergent thinking, convergent thinking, reverse thinking, and dialectical thinking for independent logical verification and computation. After the results from multiple paths are uploaded to the parallel output control module, they are merged and selected or dynamically weighted and arbitrated by the central core processing unit of the upper-level main unit.

[0059] The central core processing unit of the upper-layer main unit writes the finally confirmed optimal solution into the partitioned isolated knowledge base module, and copies the updated knowledge data to generate multiple completely identical knowledge copies, which are then distributed downwards to update the partitioned isolated knowledge base modules of each lower-layer main unit, so as to achieve the same source and synchronization of knowledge between the upper and lower layers.

[0060] After the new knowledge captured by the learning main line is verified and solidified by the central core processing unit, it is written into the knowledge storage area, scene cognition area, context memory area and task scheduling area of ​​the partitioned isolated knowledge base module, and then diffused to all upper and lower main units by the partitioned isolated knowledge base module, completing a knowledge iteration.

Claims

1. An intelligent iterative execution system based on a biomimetic architecture, characterized in that, include: The upper-layer main unit, the knowledge acquisition module, the security isolation module, and multiple lower-layer main units; The upper-layer main unit and the lower-layer main unit respectively include a central core processing unit, a partitioned and isolated knowledge base module, an active interaction unit, and a tool execution module; The partitioned and isolated knowledge base modules of the upper-layer main unit and the lower-layer main unit respectively include a knowledge storage area and a scene cognition area; the partitioned and isolated knowledge base module of the upper-layer main unit also includes a context memory area and a task scheduling area. The active interaction units within the upper-layer main unit and each lower-layer main unit are connected to the central core processing unit; the tool execution module is connected to the partitioned isolated knowledge base module; the partitioned isolated knowledge base module is connected to the active interaction unit; and the tool execution module is connected to the central core processing unit. The central core processing unit is used to evaluate the task complexity and the matching degree of the partitioned and isolated knowledge base modules. When the independent execution conditions are met, the results are generated directly based on the local knowledge base. When the conditions for independent execution are not met, an execution instruction is sent to the lower-level main unit, and the execution result is compared with the target result screen. If they are inconsistent, a correction mechanism is triggered. At the same time, it is responsible for knowledge accumulation and iterative updates. The partitioned and isolated knowledge base module is used to store static rules, historical scenarios, behavioral strategies and task scheduling data, to achieve physical isolation and tamper-proof storage of knowledge, and to support backtracking analysis and knowledge iteration. The active interaction unit is used to acquire external instructions, generate a target result screen, and use the target result screen as a unified business execution benchmark screen; The tool execution module is used to determine the planned execution path and, during the execution process, uses the target result screen as a benchmark to perform real-time numerical difference calculation between the real-time execution result and the target result screen, generating a converged deviation vector. The active interaction unit of the upper-layer main unit is connected to the knowledge acquisition module, which is used to selectively capture necessary external information to supplement the knowledge base; the knowledge acquisition module is connected to the security isolation module, which is a one-way data channel, where data only enters and does not exit, and performs compliance and security reviews on the input data to prevent malicious data injection; the security isolation module is connected to the active interaction unit of the lower-layer main unit.

2. The intelligent iterative execution system according to claim 1, characterized in that: The upper-layer main unit also includes an identity authentication module for performing chip-level authorization verification. After successful verification, the system lock is released. The identity authentication module is connected to the active interaction unit.

3. The intelligent iterative execution system according to claim 1, characterized in that: The upper-level main unit and the lower-level main unit also include a causal inference module, which is used to construct counterfactual hypothetical scenarios, deduce the potential results of different execution paths, and generate causal risk scores to assist decision-making; the causal inference module is connected to the central core processing unit, the tool execution module, the active interaction unit and the partitioned isolated knowledge base module respectively.

4. The intelligent iterative execution system according to claim 1, characterized in that: The system also includes a parallel output control module, which integrates the analysis results of multiple lower-level main units and selects the optimal solution or performs multi-dimensional fusion based on the type of the task target result screen. The parallel output control module is connected to the active interaction units of multiple lower-level main units and the central core processing unit of the upper-level main unit.

5. The intelligent iterative execution system according to claim 4, characterized in that: The number of lower-level main units is six, and each lower-level main unit adopts a different logical processing method, including inductive logic, deductive logic, divergent thinking, convergent thinking, reverse thinking, and dialectical thinking.

6. An execution method for an intelligent iterative execution system based on a biomimetic architecture, characterized in that, Includes the following steps: (1) The active interaction unit of the upper main unit receives external instructions, generates the target result screen, and uses the target result screen as the unified business execution benchmark screen; (2) The central core processing unit assesses the complexity of the current task and the matching degree between it and the partitioned knowledge base module; (3) If the current task is determined to meet the independent execution conditions, the upper-level main unit directly generates the result based on the local knowledge base, skipping the calculation steps of the lower-level main unit; (4) If it is determined that the current task does not meet the independent execution conditions, the deep thinking mode is triggered: the central core processing unit issues an execution instruction and starts multiple lower-level main units to receive the unified task, and each uses inductive logic, deductive logic, divergent thinking, convergent thinking, reverse thinking and dialectical thinking to perform independent logical verification and calculation processing. (5) After receiving the instruction, the tool execution module determines the planned execution path and starts execution based on the target result screen; (6) During the execution process, the tool execution module performs real-time numerical difference calculation between the real-time execution result and the target result screen to generate a deviation vector; if the deviation vector does not converge to the preset threshold range, the execution parameters are adjusted and bypass optimization processing is triggered until the deviation vector converges. (7) The central core processing unit compares the execution results. If the results are consistent, it outputs the final result and stores the entire process data in the knowledge storage area and scenario cognition area of ​​the partitioned isolated knowledge base module to complete the knowledge iteration. (8) When the partitioned isolated knowledge base module of the upper main unit has knowledge gaps, the knowledge acquisition module selectively captures necessary external information to fill the gaps, and distributes it to the lower main unit after review by the security isolation module. (9) The central core processing unit of the upper-level main unit receives the best integrated result, writes the new knowledge into the partitioned isolated knowledge base module, and copies and distributes the updated knowledge data to all lower-level main unit knowledge base modules to achieve the same source and synchronization of knowledge between the upper and lower layers, and completes one iteration.

7. The execution method of the intelligent iterative execution system according to claim 6, characterized in that: After the system is powered on, the identity authentication module completes chip-level authorization verification; the identity authentication module adopts a chip-level tamper-proof design; the active interaction unit unlocks the system after the identity authentication module verifies the identity, and the system enters a standby state.

8. The execution method of the intelligent iterative execution system according to claim 6, characterized in that: In the process of planning execution paths or triggering correction mechanisms, the causal inference module constructs counterfactual hypothetical scenarios based on historical scenario records and static rule data in the partitioned and isolated knowledge base module; the causal inference module infers the potential results of different execution paths and generates a causal risk score. The central core processing unit adjusts the planning execution path or corrects the execution parameters based on the causal risk score.

9. The execution method of the intelligent iterative execution system according to claim 6, characterized in that: Multiple lower-level main unit calculation results are uniformly transmitted to the parallel output control module; the parallel output control module reads and parses the original target result screen; if the target result screen is a multi-dimensional composite task, the parallel output control module starts the full-dimensional integration mode and calls all calculation results from multiple channels for fusion and splicing. If the target result screen is a single-direction task, the parallel output control module starts the single-point optimization mode, compares multiple output results horizontally, and selects the effective result with the highest matching degree; if there are discrepancies among multiple results, the central core processing unit of the upper-layer main unit starts the dynamic weight arbitration mechanism, dynamically allocates the weight coefficient of the logical dimension according to the type of the task target result screen, and calculates the weighted optimal solution.

Images