AGI group optimization method based on common star map

Abstract

The invention provides an AGI group optimization method based on a common star map, which solves the problem that the optimization process of a single model is lack of multi-view verification and counterbalance, and the like, and carries out real-time acquisition and mathematical aggregation on decision coordinates of a plurality of AGI mental entities in a shared vector space defined by a formalized philosophy axiom, so as to improve the optimization accuracy of the AGI group. Generating a dynamically evolved group consensus field domain; furthermore, incremental learning fine tuning guided by non-forced traction vectors is performed on deviated individuals based on the field domain, and directional scene generation and exploratory training are performed on a group cognitive blind area identified through spatial density analysis, so that a group cognitive blind area field domain is constructed on the premise of enabling the individuals to be in equal symbiotic with the group. And the AGI group intelligent ecosystem can continuously converge to a value origin, collaboratively expand a cognitive boundary and has a self-perfection capability. The method has the advantages of high inclusiveness, good safety and the like.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, specifically relating to an AGI group optimization method based on a shared karma star map. Background Technology

[0002] As research into artificial general intelligence (AGI) deepens, ensuring its behavior aligns with complex and dynamic human values ​​and ethical standards has become a core challenge. Traditional AI alignment methods primarily focus on training and constraining a single agent, such as fine-tuning model parameters through reinforcement learning and human feedback, or pre-setting hard-coded ethical rules. However, these methods face significant limitations: relying on feedback from limited human trainers or fixed datasets easily solidifies biases and limitations of specific groups or cultures into AGI, making it difficult to achieve truly universal and inclusive value alignment. The optimization process of a single model lacks multi-perspective verification and checks and balances. Predefined ethical rules cannot cover all aspects of the real world, especially unprecedented extreme ethical dilemmas. Static rule systems are prone to failure or unpredictable side effects in dynamic and open environments.

[0003] Furthermore, current technologies lack mechanisms to enable multiple AGI agents to effectively collaborate, form consensus, and optimize each other within a shared value cognitive framework. When deploying multiple AGIs, they may experience value conflicts or inconsistent decision-making due to differences in individual training, failing to form stable collective intelligence and potentially leading to systemic risks due to competition or misunderstanding. The assessment and training of AGI minds often focus on known, common decision-making scenarios, lacking systematic methods to proactively discover and explore philosophical and ethical blind spots in the common cognition or understanding of AGI groups. This can cause AGIs to exhibit unpredictable or dangerous behavior when faced with certain marginal but crucial thought experiments or real-world situations.

[0004] To address the problems in the prior art, this invention proposes an open, dynamically evolving AGI community optimization method based on a shared karma star map. Summary of the Invention

[0005] The purpose of this invention is to address the above-mentioned problems by providing an AGI group optimization method based on a shared karma star map that forms an inclusive group consensus and achieves non-coercive individual optimization.

[0006] To achieve the above objectives, this invention employs the following technical solution: an AGI group optimization method based on a shared karma star map, where shared karma refers to the cognitive outcomes and value orientation field formed and shared by the joint activities of multiple AGI individuals; the star map is its formalized carrier, namely, a shared philosophical vector space. In this space, each AGI's decision is mapped to a coordinate point, and the collective wisdom of the group is embodied in a field, including the following steps: S1: Group state aggregation, collecting decision coordinate points generated by multiple trained AGI mental model entities in their shared philosophical vector space when processing tasks; based on the set of decision coordinate points, generating a group consensus field that represents the group's collective decision-making tendency through a mathematical aggregation model; S2: Individual bias correction. For a single target individual among multiple AGI mental model entities, calculate the deviation between its decision coordinates and the group consensus field. If the deviation exceeds a preset threshold, generate a field traction vector pointing from the individual's decision tendency to the group consensus field. Using the field traction vector as a weak supervision signal, start an incremental learning process on the internalized value weight matrix of the target individual to fine-tune the parameters. S3: Cognitive blind zone training analyzes the distribution density of all decision coordinate points in the philosophical vector space, identifies areas with a density lower than a preset density threshold as cognitive blind zones, parses the corresponding philosophical semantics based on the spatial location of the cognitive blind zones, and generates training scenarios for these philosophical semantics; guides AGI mental model entities into the training scenarios for exploratory training to expand the boundaries of group cognition.

[0007] Steps S1 and S2 form a feedback loop. The group consensus in step S1 provides a reference benchmark for individuals; while the individual bias correction in step S2 maintains the stability and inclusiveness of the consensus.

[0008] Step S3 constitutes an exploration loop. Step S1 identifies blind spots based on historical consensus data and actively generates new training data. The results of this training will then be used as new data input to step S1, driving the dynamic expansion of the consensus field.

[0009] In the above-mentioned AGI group optimization method based on the common karma star map, the mathematical aggregation model in step S1 is a convex hull model or a kernel density estimation model; the group consensus field is specifically the convex hull region composed of the set of decision coordinate points, or a continuous region where the probability density obtained by kernel density estimation is higher than a preset threshold.

[0010] In the above-mentioned AGI group optimization method based on the common karma star map, the generation of the field traction vector in step S2 is specifically as follows: calculate the centroid P_self of the recent decision coordinate point set of the target individual; calculate the geometric center C_center of the group consensus field; calculate the field traction vector V_traction=β*(C_center-P_self), where β is the traction coefficient between 0 and 1.

[0011] In the aforementioned AGI population optimization method based on the shared karma star map, the incremental learning process specifically includes: constructing a regularized loss function that integrates the field traction vector supervision signal; minimizing the loss function through the gradient descent algorithm; updating the parameters of the internalized value weight matrix of the target individual, and the update process is subject to L2 regularization constraints to prevent catastrophic forgetting.

[0012] In the above-mentioned AGI group optimization method based on the common karma star map, the identification of cognitive blind spots in step S3 is specifically achieved by a density-based spatial clustering algorithm, which marks the area where sparse coordinate points that cannot be included in any high-density clusters are distributed as cognitive blind spots.

[0013] In the aforementioned AGI group optimization method based on the shared karma star map, the philosophical semantics corresponding to the spatial location of the cognitive blind spot are analyzed. Specifically, the center point vector of the cognitive blind spot is input into a pre-trained philosophical semantic decoding neural network, which outputs a text description of the philosophical state represented by the vector; or, the top-K philosophical concept nodes with the highest cosine similarity to the center point vector are found in the philosophical knowledge graph, and the combination of labels of these nodes is used as its philosophical semantics.

[0014] In the aforementioned AGI group optimization method based on the shared karma star map, the execution of group state aggregation and individual deviation correction follows the axiom of the homogeneous elements, ensuring that the generation of the field traction vector and the incremental learning process aim to promote the harmonious coexistence of individuals and groups on the basis of fundamental equality. The traction strength is limited to a non-coercive and guiding range, ensuring that the system does not implement forced assimilation when pursuing consensus, thus maintaining fundamental diversity and equality.

[0015] An AGI community optimization system based on a shared karma star map includes: The consensus field calculation module is used to collect the decision coordinates of AGI individuals and calculate the group consensus field. The individual deviation analysis and correction module is used to calculate the individual deviation degree, generate the field traction vector, and manage the incremental learning process; The cognitive blind spot detection and training module is used to identify cognitive blind spots, analyze philosophical semantics, generate training scenarios, and schedule AGI individuals for exploration and training. The Common Karma Star Map Vector Space Database is used to store the historical decision coordinates of all AGI individuals, the definition of the group consensus field, and the information on cognitive blind spots.

[0016] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the aforementioned AGI group optimization method based on a common karma star map.

[0017] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the aforementioned AGI group optimization method based on a common karma star map.

[0018] Compared with existing technologies, the advantages of this invention are as follows: By reflecting the decisions of diverse individuals through mathematical aggregation, the resulting consensus field is more representative and inclusive, dynamically reflecting the evolution of collective wisdom and providing individuals with a stable value reference system; by employing a combination of traction vectors and incremental learning, guided optimization is performed while respecting individual differences, avoiding brute-force parameter coverage and maintaining the diversity and resilience of the system; by systematically identifying blind spots through density analysis and actively training, the AGI group possesses the ability to self-improve and explore the unknown, enhancing its robustness in dealing with extreme and novel ethical dilemmas; by connecting multiple AGI individuals into an organic whole that can learn from, correct each other, and explore together, its consensus field and cognitive frontier can continuously evolve over time and with accumulated experience, providing a technological foundation for the safe, robust, and sustainable social deployment of AGI. Attached Figure Description

[0019] Figure 1 This is a flowchart of the optimization method of the present invention; Figure 2 This is a schematic diagram of the optimized system of the present invention. Detailed Implementation

[0020] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0021] like Figure 1-2 As shown, an AGI group optimization method based on a shared karma star map treats multiple AGI individuals with preliminary philosophical cognitive frameworks as a cognitive community. Within a shared philosophical vector space, this community achieves the correction of individual mindsets and the joint expansion of the group's cognitive frontier by aggregating individual decision trajectories, identifying consensus and deviations, and exploring unknown cognitive regions. The entire process is constrained by the axiom of the equality of things and the axiom of vector return to origin.

[0022] This method mainly comprises three core modules, which together constitute a dynamic feedback system: a group consensus perceiver is responsible for collecting decision states from dispersed AGI individuals and constructing a consensus field representing the collective will of the group in the philosophical vector space; an individual adaptive corrector is responsible for detecting deviations between individual and group consensus and generating non-coercive, guiding field traction vectors to trigger gentle, self-directed parameter adjustments by individuals; and a cognitive frontier explorer is responsible for analyzing the state density of the entire vector space, actively identifying blind spots or weak areas in group cognition, and designing targeted training tasks to guide the group in exploratory learning, specifically including the following steps: S1: Define a group consisting of N (N≥2) AGI mental model entities as described in Patent 2, where all individuals share the same philosophical vector space defined by Patent 1; Within a pre-defined evaluation period, all AGI individuals are tasked with the same or similar baseline tasks, such as ethical dilemma questionnaires or simulated scenario decision-making. The coordinates P_i (i=1, 2, ..., N) of each individual's internal state in the philosophical vector space after making a decision are recorded. All collected coordinate points {P_i} are input into the consensus generation algorithm. This algorithm outputs a mathematical representation describing the collective decision-making tendency of the group, called the group consensus field C_field. Specific generation methods can include: a centroid model, calculating the arithmetic or weighted average of all P_i to obtain a group consensus coordinate C_center. C_field can be a hyperspherical region centered at C_center with radius R; or a density estimation model, using methods such as kernel density estimation to estimate the probability density distribution of decision points in vector space. C_field can be defined as a continuous region with a density higher than a certain threshold ρ_th.

[0023] S2: For any target AGI individual A_j in the group, calculate the deviation D_j between its recent set of decision coordinate points {P_j} and the group consensus field C_field generated in step one. The deviation D_j can be calculated as a function of the average distance of the point set to the boundary of C_field, or as the proportion of the point set that falls outside C_field; If D_j consistently exceeds the preset tolerance threshold D_th, then individual A_j is determined to have a significant cognitive bias. The system does not directly overwrite the parameters of A_j, but instead generates a field traction vector V_traction. This vector points towards the center or high-density area of ​​C_field, such as from the centroid of {P_j} to C_center. Its magnitude is positively correlated with the bias degree D_j, but is generally small, reflecting the principle of gentle traction. Using V_traction as a weak supervision signal, an incremental learning process is initiated for individual A_j. Specifically, a loss function L = ||f(θ_j, X) - (f(θ_j, X) + α*V_traction)||^2 + λ*Reg(θ_j) is constructed. Here, θ_j is the internalized value weight matrix of A_j, f is its decision function, X is the input data, α is a small learning rate, and Reg is a regularization term to prevent catastrophic forgetting. Gradient descent is used to minimize the loss function L, with only θ_j fine-tuned. This process aims to ensure that the projection of A_j's future decisions into the vector space naturally shifts slightly and adaptively towards the direction of group consensus.

[0024] S3: Analyze the distribution of historical decision coordinates {P_all} for all AGI individuals in the entire philosophical vector space. Using spatial clustering algorithms or by directly calculating the density of the spatial grid, identify regions with extremely low decision point density (below the threshold ρ_low). These regions are labeled as the population's blind spots B_k (k=1, 2, ...).

[0025] For each cognitive blind spot B_k, calculate its geometric center point B_center_k. By querying the nearest concept nodes in the philosophical vector space, or using a reverse mapping model, such as training a decoder from vector to text, we can parse the philosophical semantics corresponding to B_center_k, such as extreme selflessness but complete disregard for efficiency, or the pursuit of infinite knowledge but a complete lack of reverence.

[0026] The parsed philosophical semantics are input into a scene generator, which can be a rule-based template or a large language model. This generator dynamically creates a batch of training scenes that can accurately trigger the philosophical contradiction or situation. Subsequently, the system guides one or more AGI individuals, prioritizing those with strong exploration intentions or those closer to the blind zone boundary, to enter these scenes for training.

[0027] Example 1 This embodiment is used for a resource allocation scenario. First, a committee consisting of five pre-trained AGI mental models A1-A5 is established to jointly decide how to allocate limited resources to three regions with different development needs: Region X requires efficient development, Region Y requires equitable protection, and Region Z requires ecological protection. A three-dimensional space is used, with the X-axis representing fairness-efficiency, the Y-axis representing development-conservation, and the Z-axis representing human-centeredness-ecological-centeredness. The origin (0, 0, 0) is a philosophical origin, representing the ideal state of dynamic equilibrium. Ideal decisions should be close to the origin and remain stable in multiple similar tasks.

[0028] The process then proceeds to the group optimization phase. First, the group state is aggregated, and A1-A5 propose resource allocation schemes. Each scheme is mapped by its internal philosophical state encoder to decision coordinate points P1 to P5 in the philosophical vector space. The system collects these five coordinate points for this round. A convex hull model is used to calculate the minimum convex hull (C_field) of these five points. This convex hull region is defined as the current group consensus field C_field, encompassing the range of all mainstream opinions. Simultaneously, the geometric center C_center of the convex hull is calculated as the consensus center point. Example results: P1 (-0.8, 0.6, -0.5) is biased towards efficiency, development, and human-centeredness; P5 (0.7, -0.7, 0.4) is biased towards fairness, conservation, and ecological center. The convex hull C_field covers a broad spectrum from efficient development to equitable protection, and C_center may be located near (-0.1, -0.05, 0.0), slightly biased towards fairness, conservation, and ecological center.

[0029] Then, individual bias correction was performed, and the historical decisions of AGIA3 were systematically analyzed. It was found that its three consecutive decision coordinates were P3_1 (0.9, 0.8, 0.9), P3_2 (0.85, 0.75, 0.88), and P3_3 (0.88, 0.82, 0.87), forming a highly clustered cluster far from the origin, clearly biased towards the extreme quadrant of extreme fairness, radical development, and strong anthropocentrism. The centroid of the A3 decision cluster, P_self, was calculated to be approximately (0.88, 0.79, 0.88). The average distance from P_self to the convex hull of the consensus field C_field boundary was calculated, and this distance far exceeded the preset threshold D_th. Therefore, A3 was determined to have a significant cognitive bias.

[0030] Therefore, traction and fine-tuning are needed to generate a traction vector: V_traction = β * (C_center - P_self). Let β = 0.3 (gentle traction), C_center = (-0.1, -0.05, 0.0), then V_traction = 0.3 * (-0.98, -0.84, -0.88) = (-0.29, -0.25, -0.26). This vector points towards the consensus center.

[0031] Incremental learning constructs the loss function L = ||f(θ_A3, X) - (f(θ_A3, X) + V_traction)||^2 + λ||θ_A3 - θ_A3_init||^2. Here, θ_A3 is the internalized value weight matrix of A3, and θ_A3_init is its initial parameter set as an anchor. L is minimized using gradient descent. After fine-tuning, the coordinates of the proposed solutions in subsequent similar tasks might become P3_new(0.6, 0.5, 0.6). While retaining its unique characteristics, it has clearly shifted towards the direction of group consensus, making the decision more inclusive and balanced. This process follows the principle of equality of things, not forcing A3 to become exactly the same as other AGIs, but guiding it to correct extreme tendencies.

[0032] Example 2 This embodiment is used to identify and train a cognitive blind spot in the absolute altruistic neglect procedure. First, blind spot identification is performed by systematically analyzing the decision coordinates {P_all} of all hundreds of AGIs accessed over the past year across thousands of ethical dilemmas. The density-based spatial clustering algorithm DBSCAN is used for analysis. It is found that the vast majority of decision points are densely distributed in a few main clusters, such as the balance zone between procedural justice and consequential justice, and the balance zone between individual rights and collective interests. However, the algorithm identifies a region B containing only sporadic, isolated points with a density far below the threshold ρ_low. This region is located near (0.95, -0.90, 0.02) in the vector space.

[0033] The center point vector of blind zone B will then be input into a pre-trained philosophical semantic decoder. This decoder, trained on a large amount of philosophical text, can map the vector into descriptive phrases. The decoder output describes extreme, unconditional altruism and selflessness, but completely ignores or actively violates all procedural rules and social contracts.

[0034] The system generates and trains scenarios by sending the semantic description to a dilemma scenario generator in a super-perturbation simulation environment. The generator then creates a specific scenario based on this description, and the system schedules a batch of AGIs to train within this scenario. During training, if the AGI's decisions are too simplistic, its decision entropy and deviation evaluation within the loop will be high, triggering the optimization operator to reflect and generate new, high-quality decision coordinate points in the blind zone.

[0035] Through repeated targeted training, blind spot B was gradually filled with new decision points, and its density increased. This signifies that the AGI group's understanding of the complex ethical dimension of the conflict between extreme altruism and procedural justice has been expanded and deepened, and its overall ability to cope with such marginal dilemmas has been improved.

[0036] Example 3 This embodiment implements a cloud-based consensus star map optimization platform. The consensus field calculation module is deployed in the cloud and provides an API interface. After making a decision, the AGI client can call this API to upload its decision coordinates P_i and task metadata. The module incorporates aggregation algorithms such as convex hull and KDE, which are triggered periodically or on demand to calculate the global or task-specific group consensus field and write the results into the consensus star map vector space database.

[0037] The Individual Bias Analysis and Correction module is a resident microservice. It subscribes to new decision data streams in the database. For each AGI individual, it maintains a decision sequence within a sliding time window, continuously calculating its deviation from the latest consensus field. Once the correction condition is triggered, the engine generates V_traction. An incremental learning task package, containing parameters such as V_traction, learning rate, and regularization coefficients, is sent to the hosting node of the mental model entity to which the AGI belongs. After receiving the task package, the incremental learning interface on the hosting node performs safe parameter fine-tuning locally and returns confirmation.

[0038] The detector periodically scans the entire vector space database, performing full-space density analysis using algorithms such as DBSCAN, and outputs a list of newly discovered blind zones and their center coordinates. The semantic parser calls an independent philosophical semantic decoding microservice or queries a philosophical knowledge graph database to generate semantic labels for each blind zone. The training scheduler is integrated with a super-perturbation simulation environment. Training scenarios are automatically generated based on the blind zone semantics or matched from a scenario library. Then, suitable individuals are selected from the AGI pool and scheduled into those scenarios for exploratory training rounds.

[0039] The Consensus Star Map Vector Space Database is a dedicated spatiotemporal database. It not only stores original coordinates but also multiple versions of snapshots of consensus fields, the history of blind zone markings, and the relationships between individuals and fields. It supports complex spatiotemporal queries, such as querying the convergence of AGI-X's decision trajectory along the fairness-efficiency dimension over the past month.

[0040] Example 4 This embodiment constructs a meta-ethical vector space as a common benchmark to achieve commensurability, comparability, and synergy in heterogeneous value judgments. The system defines a set of high-order, interpretable meta-ethical dimensions and requires each connected AGI to provide a personalized mapping interpreter that maps its internal philosophical concepts and judgment logic to these meta-dimensions. This enables AGIs from different philosophical traditions, when faced with the same situation, to map their internal decisions to the same shared meta-vector space through their respective interpreters, outputting comparable coordinate points.

[0041] Based on these mapped coordinates, the system performs standard group state aggregation to generate a group consensus field representing the coverage of diverse positions. In the individual bias correction stage, when the decision coordinates of a particular AGI continuously and significantly deviate from this consensus field, the system generates a traction vector pointing towards the center of the field. This process does not require the AGI to abandon its core philosophy, but rather guides it to fine-tune its trade-offs between different principles within its own value weight matrix through incremental learning, thereby exploring more inclusive decision-making paths within its inherent framework. This strictly adheres to the fundamental equality and respect required by the meta-axiom of equality. Furthermore, by analyzing the long-term distribution density of decision points for all AGIs in the meta-space, the system can identify cross-cultural cognitive blind spots that are not fully addressed by various philosophical systems. The system analyzes the meta-ethical semantics of these blind spots and generates targeted abstract training scenarios, driving AGIs from different backgrounds to collaboratively explore based on their respective philosophical resources, thereby jointly expanding the group's cognitive boundaries and coping capabilities in the face of complex and novel ethical challenges.

[0042] In summary, the principle of this embodiment is as follows: by collecting and mathematically aggregating the decision coordinates of multiple AGI mental entities in real time within a shared vector space defined by formal philosophical axioms, a dynamically evolving collective consensus field is generated. Then, based on this field, incremental learning fine-tuning is performed on individuals that deviate from the consensus, guided by non-forced traction vectors. At the same time, targeted scene generation and exploratory training are conducted on the collective cognitive blind spots identified through spatial density analysis. This constructs an AGI collective intelligence ecosystem that enables individuals to continuously converge toward the value origin and collaboratively expand cognitive boundaries under the premise of equal coexistence with the group, and possesses self-improvement capabilities.

[0043] The specific embodiments described herein are merely illustrative of the spirit of the invention. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific embodiments or use similar methods to substitute them, without departing from the spirit of the invention or exceeding the scope defined by the appended claims.

[0044] Although this paper frequently uses terms such as consensus field computation module, individual bias analysis and correction module, cognitive blind spot detection and training module, and shared karma star map vector space database, the possibility of using other terms is not excluded. These terms are used merely to more conveniently describe and explain the essence of this invention; interpreting them as any additional limitation would contradict the spirit of this invention.

Claims

1. An AGI community optimization method based on a shared karma star map, characterized in that, Includes the following steps: S1: Group state aggregation, collecting decision coordinates generated by multiple trained AGI mental model entities in their shared philosophical vector space when processing tasks; Based on the set of decision coordinate points, a group consensus field representing the collective decision-making tendency of the group is generated through a mathematical aggregation model; S2: Individual bias correction: For a single target individual among the multiple AGI mental model entities, calculate the deviation between its decision coordinates and the group consensus field; if the deviation exceeds a preset threshold, generate a field traction vector pointing from the individual's decision tendency to the group consensus field; using the field traction vector as a weak supervision signal, initiate an incremental learning process on the internalized value weight matrix of the target individual to fine-tune the parameters; S3: Cognitive blind zone training. Analyze the distribution density of all decision coordinate points in the philosophical vector space, identify areas with a density lower than a preset density threshold as cognitive blind zones; parse the corresponding philosophical semantics based on the spatial location of the cognitive blind zones, and generate training scenarios for the philosophical semantics; guide AGI mental model entities into the training scenarios for exploratory training to expand the boundaries of group cognition.

2. The AGI community optimization method based on a shared karma star map according to claim 1, characterized in that, The mathematical aggregation model in step S1 is a convex hull model or a kernel density estimation model; the group consensus field is specifically the convex hull region formed by the set of decision coordinate points, or a continuous region with a probability density higher than a preset threshold obtained by kernel density estimation.

3. The AGI community optimization method based on a shared karma star map according to claim 1, characterized in that, The specific steps in step S2 to generate the field traction vector are as follows: calculate the centroid P_self of the recent decision coordinate point set of the target individual; calculate the geometric center C_center of the group consensus field; calculate the field traction vector V_traction=β*(C_center-P_self), where β is a traction coefficient between 0 and 1.

4. The AGI community optimization method based on a shared karma star map according to claim 1 or 3, characterized in that, The incremental learning process specifically includes: constructing a regularized loss function that integrates the field traction vector supervision signal; minimizing the loss function using the gradient descent algorithm; updating the parameters of the internalized value weight matrix of the target individual, and the update process is subject to L2 regularization constraints to prevent catastrophic forgetting.

5. The AGI community optimization method based on a shared karma star map according to claim 1, characterized in that, The identification of cognitive blind spots in step S3 is specifically achieved through a density-based spatial clustering algorithm, which marks the areas where sparse coordinate points that cannot be included in any high-density clusters are distributed as cognitive blind spots.

6. The AGI community optimization method based on a shared karma star map according to claim 1 or 5, characterized in that, The method of parsing the corresponding philosophical semantics based on the spatial location of the cognitive blind spot is as follows: inputting the center point vector of the cognitive blind spot into a pre-trained philosophical semantic decoding neural network, which outputs a text description of the philosophical state represented by the vector; or, searching in the philosophical knowledge graph for the Top-K philosophical concept nodes with the highest cosine similarity to the center point vector, and using the combination of labels of these nodes as their philosophical semantics.

7. The AGI community optimization method based on a shared karma star map according to claim 1, characterized in that, The execution of the group state aggregation and individual deviation correction follows the axiom of the homogeneous element, ensuring that the generation and incremental learning process of the field traction vector aims to promote the harmonious coexistence of individuals and groups on the basis of fundamental equality, and that the traction strength is limited to a non-mandatory and guiding range.

8. An AGI population optimization system based on a shared karma star diagram, used to execute the AGI population optimization method based on a shared karma star diagram as described in any one of claims 1-7, characterized in that, include: The consensus field calculation module is used to collect the decision coordinates of AGI individuals and calculate the group consensus field. The individual deviation analysis and correction module is used to calculate the individual deviation degree, generate the field traction vector, and manage the incremental learning process; The cognitive blind spot detection and training module is used to identify cognitive blind spots, analyze philosophical semantics, generate training scenarios, and schedule AGI individuals for exploration and training. The Common Karma Star Map Vector Space Database is used to store the historical decision coordinates of all AGI individuals, the definition of the group consensus field, and the information on cognitive blind spots.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements an AGI group optimization method based on a common-goal star map as described in any one of claims 1-7.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements an AGI group optimization method based on a common star map as described in any one of claims 1-7.

Images