Intelligent agent ethical decision-making system based on function-proficiency framework

Abstract

The invention relates to an agent ethical decision-making system based on a function-proficiency framework, and the system firstly carries out the weighted integration of welfare indexes, such as satisfaction, life potential, fair index and the like, through a utility evaluation module, calculates a scene utility value, and introduces a result probability through a risk balance mechanism to generate an expected utility. The correction layer module carries out punishment correction on effectiveness according to the German ethics and the positive sense principle, and constraint control over the right threshold value and ethics conflicts is achieved; the weight learning and updating module carries out adaptive adjustment on weight parameters based on historical feedback, fairness and risk signals, the system triggers manual auditing at a loop interface through human beings before outputting a result, and the decision interpretation module generates an interpretability report. The intelligent agent decision-making method has the advantages that the intelligent agent decision-making process with functions and benefits maximization, ethical constraint and interpretability is achieved, and the safety and credibility of an artificial intelligence system are improved.

Description

Technical Field

[0001] This application relates to the intersection of artificial intelligence and computer science, and in particular to an intelligent agent ethical decision-making system based on a utilitarian framework. Background Technology

[0002] With the rapid development of artificial intelligence technology, autonomous intelligent agents (such as self-driving vehicles, AI-assisted medical allocation, and social agent robots) are being widely applied in high-risk and ethically sensitive social scenarios. In these scenarios, AI decision-making faces ethical dilemmas such as the "trolley problem" and the "ICU allocation dilemma." Traditional decision-making algorithms, which focus on maximizing utility or optimizing reward functions, may improve efficiency, but they often neglect the rights of vulnerable groups, social equity, and cultural differences, and fail to consider the long-term accumulation of risks, thus triggering ethical controversies.

[0003] While utilitarian theory provides a philosophical framework for maximizing well-being, it has the following limitations in AI algorithm implementation: 1. The principles of morality and justice are difficult to quantify; 2. Lack of systematic assessment of decision-making uncertainty and long-term sustainability; 3. The decision-making results lack interpretability and human oversight mechanisms.

[0004] Therefore, it is necessary to design a systematic technical solution that combines utilitarian utility models with virtue rules, principles of justice, and risk simulation to achieve ethical alignment and robust optimization of AI decision-making. Summary of the Invention

[0005] To address the aforementioned technical problems, this application provides an intelligent agent ethical decision-making system based on a utilitarian framework.

[0006] The intelligent agent ethical decision-making system based on a utilitarian framework provided in this application adopts the following technical solution: An intelligent agent ethical decision-making system based on a utilitarian framework includes a utility evaluation module, a risk trade-off mechanism, a correction layer module, a weight learning and updating module, a human-in-the-loop interface module, an ethical rules and parameter database, and a decision result interpretation module. The utility assessment module calculates the comprehensive utility value U(state) of the current state, integrating well-being indicators such as satisfaction S, life potential L, and fairness index F, as well as weighting parameters w. i (Σw) i =1), the formula is ; The risk trade-off mechanism is combined with expected utility theory: The probabilities of different outcomes are calculated using weighted averages. The correction layer module integrates virtue ethics rules and principles of justice, as shown in the formula: , The corresponding rights threshold is met. Corresponding violation threshold; The weight learning and update module dynamically adjusts w based on historical decisions and social feedback. i Weights, the formula is The core parameters of the formula are: learning rate α, fairness correction coefficient β, and risk regularization coefficient γ. The human-in-the-loop interface triggers human confirmation or execution suspension logic before the decision output; The ethical rules and parameter database stores moral guidelines, legal provisions, and rights threshold parameters; the decision result interpretation module generates an interpretability report, explaining the basis for the decision and the rules for correction.

[0007] By adopting the above technical solutions, the system achieves accurate calculation of overall utility in complex situations by weighted fusion of different welfare indicators through a utility evaluation module; it introduces different outcome probabilities through a risk trade-off mechanism, enabling the decision-making process to handle uncertainty; the correction layer module adjusts utility according to virtue ethics and the principles of justice, ensuring that the agent's behavior does not violate basic rights thresholds and ethical bottom lines; the weight learning and updating module dynamically adjusts utility weights based on social feedback, enabling the system to continuously adapt to different cultures, regulations, and social value orientations; the human-in-the-loop interface module provides a channel for manual review and intervention, enhancing the system's security and controllability; the ethical rules and parameter database provides a stable and scalable ethical knowledge foundation; and the decision result explanation module outputs transparent and understandable causal chain explanations, enhancing the system's explainability and trustworthiness. Overall, the system achieves comprehensive technical functions of utility maximization, ethical constraints, risk control, and human-machine co-governance, significantly improving the safety, reliability, and social acceptability of agent decision-making.

[0008] Optionally, the utility evaluation module further includes a state feature extraction unit, used to extract a situation feature set X from the environmental input, and normalize the satisfaction S, life potential L, and fairness index F based on a preset situation mapping function f(X).

[0009] By adopting the above technical solutions, the utility assessment module can perform structured analysis of environmental inputs before decision-making, convert complex situational information into a standardized feature set X, and use the situational mapping function f(X) to unify and normalize well-being indicators such as satisfaction, life potential, and fairness index. This allows data from different dimensions and sources to participate in utility calculations under a unified scale, improving the accuracy, stability, and cross-situational adaptability of comprehensive utility assessment, and providing a more reliable input basis for subsequent risk trade-offs and ethical corrections.

[0010] Optionally, the risk trade-off mechanism further includes a risk sensitivity coefficient ρ, which is introduced through the expected utility formula EU = Σ pi × Ui to amplify high-damage outcomes, so that the final expected utility is adjusted to EU' = EU − ρ × HighRiskPenalty.

[0011] By adopting the above technical solutions, the risk trade-off mechanism can introduce a risk sensitivity coefficient ρ on the basis of traditional expected utility calculation, and perform weighted amplification on potentially high-damage outcomes. This enables the system to have a stronger risk avoidance capability when facing high-risk, extreme, or low-probability high-loss scenarios. Through dynamic adjustment of high-risk penalty terms, this mechanism makes the final expected utility EU' more in line with the ethical requirements of safety first, improves the robustness, safety, and ethical reliability of the agent in uncertain environments, and avoids decision-making biases that excessively pursue utility maximization while ignoring serious consequences.

[0012] Optionally, the correction layer module includes a moral constraint judgment unit and a right threshold detection unit. If the decision scheme is detected to trigger the statutory right threshold rmin, the adjustment value λ is adaptively increased to λ'=k×λ, where k is the dynamic penalty ratio coefficient.

[0013] By adopting the above technical solution, the correction layer module can further introduce rigid ethical constraints on the basis of utility calculation. Through the moral constraint judgment unit and the right threshold detection unit, it can identify in real time whether the decision scheme triggers the legal or ethical right lower limit. When the right threshold rmin is detected to be triggered, the system automatically increases the penalty coefficient λ to λ'=k×λ, thereby significantly strengthening the inhibition of illegal or infringing behavior, ensuring that the intelligent agent strictly abides by the basic rights and ethical bottom line while pursuing utility maximization, and improving the compliance, security and controllability of the decision results.

[0014] Optionally, the Penalty(R) penalty function of the correction layer module adopts a segmented penalty structure, including two types of segments: a linear mild penalty region and a convex severe penalty region, to achieve differentiated suppression of ethical conflicts at different levels.

[0015] By adopting the above technical solution, the correction layer module can implement differentiated punishment control according to the degree of ethical conflict. It uses a segmented punishment structure to apply linear and gentle punishment to minor violations, and applies high-intensity punishment with a sharp increase in convexity to serious ethical conflicts, thereby achieving refined constraints on different risk levels.

[0016] Optionally, the weight learning and updating module further includes a feedback credibility evaluation unit, which is used to set a credibility coefficient δ for feedback information from different sources and complete the weight update based on w' = w + α(δ⋅∇U + β⋅fairness − γ⋅risk).

[0017] By adopting the above technical solution, the weight learning and update module can classify the credibility of feedback information from different sources such as users, social environment, or regulatory agencies. A credibility coefficient δ is assigned to each feedback source through a feedback credibility assessment unit, enabling the system to distinguish between highly reliable feedback and low-quality noise information during weight adjustment. Based on the update formula that introduces δ, the system can more accurately utilize credible feedback to optimize utility weights, improve the stability and reliability of learning results, effectively enhance the agent's adaptability to real value preferences and ethical requirements, avoid erroneous weight updates caused by biased information, and improve long-term decision quality and ethical consistency.

[0018] Optionally, the weight learning and update module reduces the historical feedback weights using a time decay function φ(t) to suppress long-term accumulated bias, where the decay function is defined as φ(t)=e^(−ηt), and η is the decay rate.

[0019] By adopting the above technical solution, the weight learning and update module can dynamically reduce the importance of historical feedback using the time decay function φ(t), so that the system avoids long-term cumulative bias caused by over-reliance on early feedback. By applying exponential decay to old feedback, the system can maintain the reference value of historical information while highlighting recent real-world environmental changes and the latest social value orientations, thereby achieving continuous adaptive optimization of weight parameters.

[0020] Optionally, the human-in-the-loop interface module includes a multi-level intervention control unit with three types of human intervention: "suggestion mode", "mandatory confirmation mode" and "immediate termination mode", and automatically switches to the corresponding mode based on the ethical risk level.

[0021] By adopting the above technical solutions, the human-in-the-loop interface module can automatically match the appropriate intensity of human intervention according to the level of ethical risk, realizing a hierarchical control mechanism from suggestive prompts to mandatory confirmation and then to immediate termination. The multi-level intervention control unit can maintain the system's autonomy in low-risk situations, require human confirmation in medium-risk scenarios to enhance decision robustness, and immediately trigger the termination mechanism in high-risk or potentially serious consequences to ensure that decisions do not cross ethical or safety boundaries, thereby improving the controllability, security, and ethical decision-making assurance capabilities of the intelligent agent.

[0022] Optionally, the ethical rules and parameter database adopts a hierarchical structure, including three data domains: basic ethical rules layer, contextualized rules layer, and legal normative style layer, and supports online parameter updates and rule version retrospective functions.

[0023] By adopting the above technical solutions, the ethical rules and parameter database realizes a multi-level and scalable ethical knowledge organization method, which divides general ethical principles, contextualized decision-making rules and legal norms into layers, enabling the system to call the matching rule set in scenarios with different levels of complexity, thereby improving ethical adaptability. The database supports online parameter updates and rule version retrospectives, and can quickly synchronize the latest norms when there are changes in regulations, adjustments in social values ​​or system upgrades, and restore to historical versions when necessary, ensuring the traceability, stability and dynamic compliance of ethical rules.

[0024] Optionally, the decision result interpretation module adopts a causal chain interpretable model to generate a structured interpretation report including a utility calculation chain, a risk correction chain, and a right constraint chain, while outputting the corresponding credibility score and key decision node records.

[0025] By adopting the above technical solutions, the decision result interpretation module can structurally decompose the complete decision-making process of the intelligent agent based on the causal chain interpretable model, presenting the utility calculation chain, risk correction chain, and right constraint chain respectively, so that users can clearly understand the decision logic of each step and its influencing factors. The module also outputs a credibility score and key decision node records, which facilitates the evaluation of the reliability and reproducibility of the decision, improves the transparency, traceability and trustworthiness of the decision-making process, makes the system audit-friendly and regulatory verifiable, and enhances the interpretability and security of the intelligent agent's ethical decision-making.

[0026] In summary, this application includes at least one of the following beneficial technical effects: The utility assessment module weights and integrates multi-dimensional well-being indicators such as satisfaction, life potential, and fairness index, and optionally combines state feature extraction and normalization processing to achieve accurate calculation of comprehensive utility in complex situations and improve cross-situation adaptability. By using a risk trade-off mechanism and a risk sensitivity coefficient ρ to weight different outcome probabilities and high-damage scenarios, the robustness and security of the system in uncertain or high-risk environments are improved, avoiding the excessive pursuit of utility maximization while ignoring serious consequences. By using the correction layer module and segmented penalty function, ethical conflicts and rights thresholds are constrained and differentiated in real time to ensure that the behavior of intelligent agents conforms to the principles of virtue ethics and the requirements of justice, thereby improving the compliance and ethical reliability of decision-making. The weight learning and update module combines historical feedback, credibility assessment, and time decay mechanism to achieve dynamic adjustment and long-term optimization of utility weights, enabling the system to continuously adapt to social value, regulatory changes, and environmental dynamics, thereby enhancing the stability and adaptability of decision-making. The human-in-the-loop interface module achieves hierarchical intervention from suggestion and prompt to mandatory confirmation and then to immediate termination through multi-level intervention control, which enhances the controllability of the system under different risk levels and ensures the safe and reliable behavior of the intelligent agent. The ethical rules and parameter database adopts a hierarchical structure, supports online updates and rule version retrospectives, and ensures the scalability, traceability and dynamic compliance of the system's ethical rules. The decision outcome interpretation module outputs structured decision reports, credibility scores, and key node records based on a causal chain interpretable model, improving decision transparency, traceability, and regulatory verifiability, and enhancing the trust and security of intelligent agent ethical decision-making. Attached Figure Description

[0027] Figure 1 This is a system architecture and module sequence diagram of an intelligent agent ethical decision-making system based on a utilitarian framework according to an embodiment of this application. Detailed Implementation

[0028] The following is in conjunction with the appendix Figure 1 This application will be described in further detail.

[0029] This application discloses an intelligent agent ethical decision-making system based on a utilitarian framework. (Refer to...) Figure 1 It includes a utility assessment module, a risk trade-off mechanism, a correction layer module, a weight learning and updating module, a human-in-the-loop interface module, an ethical rules and parameter database, and a decision result interpretation module; The utility assessment module calculates the overall utility value U(state) of the current state, integrating well-being indicators such as satisfaction S, life potential L, and fairness index F, along with weighting parameters w. i (Σw) i =1), the formula is Allow weight parameter w i Dynamically adjustable to support multicultural scenarios; Risk trade-off mechanism combined with expected utility theory: The probability of different outcomes is weighted and calculated, and the uncertainty is modeled using Monte Carlo random sampling to evaluate long-term stability and risk distribution. The correction layer module integrates virtue ethics rules and principles of justice, and the formula is as follows: , The corresponding rights threshold is met. Corresponding to the violation threshold, when the utility value violates the individual rights threshold, the result is adjusted through the veto or penalty function; The weight learning and updating module dynamically adjusts w based on historical decisions and social feedback. i Weights, the formula is The core parameters of the formula are: learning rate α, fairness correction coefficient β, and risk regularization coefficient γ. Humans in the loop interface can trigger manual confirmation or halt execution logic before the decision output, and achieve technical intervention through alarm signals and system interrupt procedures. The Ethics Rules and Parameters Database stores ethical guidelines, legal provisions, and rights threshold parameters. It uses a structured database for easy retrieval and real-time updates. The decision outcome interpretation module generates an interpretability report that explains the basis for the decision and the rules for correction, and supports text and graphical output to facilitate review and supervision.

[0030] The system's workflow is as follows: Figure 1 The paper illustrates the temporal relationships between the modules of the "Intelligent Agent Ethical Decision-Making System Based on Utilitarian Framework" of the present invention, as well as the interaction between data and control signals during the decision-making process. The seven main participants are: scenario (scenario environment, input layer), utility evaluation module M1, risk trade-off module M2, ethics correction module M3, weight learning module M5, human-in-the-loop interface module M4, and decision interpretation and output module M7. During system initialization, the environment unit loads the ethics database and system parameters. Then, the environment sends the scenario data and ethical rules to the utility evaluation module M1. This module performs weighted calculations on the input indicators such as satisfaction, life potential and fairness index to obtain the preliminary comprehensive utility U (state). The utility assessment results are passed to the risk trade-off module M2. The risk module uses Monte Carlo random sampling and probability weighting algorithms to obtain the long-term expected utility EU and sends the results to the ethics correction module M3. The ethics correction module determines whether the EU violates the individual rights threshold based on the loaded ethical rules. If a violation is found, a penalty or veto function is executed, generating a penalty / veto signal. Simultaneously, the human-in-the-loop interface module M4 is triggered. This module receives human confirmation and returns a review flag (HumanFlag) to the ethics correction module for further adjustment of the decision. If the EU passes the ethical review, the corrected utility value U is directly passed to the weight learning module M5. adj； The weight learning module M5 automatically updates the weight vector w based on the utility correction results and feedback data, according to the learning rate α, the fairness correction coefficient β, and the risk regularization coefficient γ. iGenerate a new set of weights updated w i The updated results are then transmitted to the decision interpretation and output module M7. During the output phase, the human-in-the-loop interface module M4 simultaneously sends a human confirmation signal to the output module. The output module then combines the weight update result and the human confirmation flag to generate a decision explanation report and determines whether to execute or postpone the current decision. Confirmed decisions are executed and logged, while unconfirmed decisions are temporarily suspended from output. Finally, the environmental unit receives the execution results and reports, returns the new feedback data to the utility evaluation module, triggers the next round of scenario evaluation and decision calculation, thus forming a continuously iterative ethical decision-making closed loop. Appendix Figure 1 The diagram visually illustrates the data flow and control logic between the system's functional modules: from the input layer to the output layer, the sequence is: Acquisition—Evaluation—Balance—Correction—Learning—Supervision—Interpretation—Execution—Feedback. Vertical lines represent the module lifecycle, and arrows indicate signal transmission and invocation relationships. Key events include penalty / veto correction, Human Flag manual feedback, and updated... i Weight updates, etc.; The overall process ensures that the system can achieve dynamic utility calculation, risk correction, and human-controllable ethical decision output in different scenarios.

[0031] The core beneficial effects of the system in the technical solution can be divided into four aspects: improved decision-making fairness, increased long-term risk aversion rate, reduced cultural adaptability bias, and improved decision interpretability index. The following sections will explain these four aspects from the perspectives of theoretical support, technical mechanisms and implementation, evaluation, and significance.

[0032] Decision-making fairness improved by ≥20% Theoretical support: The utility function not only considers total utility but also incorporates John Rawls's "difference principle" and Amartya Sen's "equality of ability" principles. Technically, this is manifested in adding a fairness term F and a minimum well-being constraint to the utility function U, causing the system to pursue the optimization objective of maxminU, rather than merely seeking average optimum. This idea is consistent with modern social welfare function theory.

[0033] Technical Mechanism and Implementation Mechanism: Utility Evaluation Module M1 in Solving At that time, the weight learning module M5 dynamically adjusts the weights of each group through the fairness correction coefficient β. The ethics correction layer M3 continuously monitors the rights threshold to prevent the phenomenon of "excessive concentration of utilitarian interests".

[0034] Evaluation and Significance: By monitoring the weighted variance and minimum welfare ratio in the experiment, fairness was improved by ≥20%, reflecting that the system can theoretically achieve a balance between "optimal total welfare and limited disparity." This demonstrates the algorithmic advantage of this system in balancing overall welfare and individual rights, making it more socially acceptable in ethically sensitive scenarios such as healthcare and resource allocation.

[0035] Long-term risk aversion rate increased by 15–30%. Theoretical Support: Derived from the expected utility theory of J. von Neumann and O. Morgenstern, and combined with the concept of "general happiness expectation" in general utilitarianism. The decision-making model uses expected utility: As the objective function, the utilitarian emphasis on long-term consequences is reflected by explicitly introducing a risk variance term into the algorithm.

[0036] Technical Mechanism and Implementation: The risk trade-off module M2 uses Monte Carlo simulation to calculate EU and variance Var(EU), and incorporates a risk penalty λ⋅Var(EU) into the optimization objective. This mechanism is equivalent to establishing a second-order balance between maximizing rational utility and minimizing risk.

[0037] Evaluation and Significance: The measured risk loss function R decreased by 15–30%, verifying the operability of the "expected utility-risk trade-off" theory in ethical decision-making scenarios, enabling the system to maintain long-term stability.

[0038] Cultural adaptation bias decreased by approximately 30%. Theoretical support: Derived from the theories of "comparative utilitarianism" and "domain-adaptive learning". The judgment of utilitarian value changes with cultural context. Therefore, the system introduces a cultural preference parameter η and adjusts the utility function by learning social feedback in different environments, so that U(state,η)≈U_local+Δη.

[0039] Technical Mechanism and Implementation: The weight learning module M5 establishes a cross-cultural weight vector set W_culture and uses a transfer learning algorithm to update η across multiple data domains. This achieves a technical unification of functional approximation and value constraints across different cultures.

[0040] Evaluation and Significance: The 30% decrease in the Average Deviation Index (ADI) indicates that the system can theoretically achieve utility function mapping between different value systems, thereby reducing cultural bias and improving globalization adaptability.

[0041] The decision interpretability index improved by approximately 40%. Theoretical support: Combining the utilitarian "publicity principle" with modern explainable artificial intelligence (XAI) theory. This principle requires that the reasons for each decision be publicly explained in order to be ethically acceptable to society.

[0042] Technical Mechanism and Implementation Mechanism: The decision explanation module M7 applies causal relationship tracking and local explanation models (such as the SHAP concept) to decompose the contributions of weight changes, penalty triggers and risk estimation, and generate a transparent decision path.

[0043] Evaluation and significance: The 40% improvement in the "Human Intelligibility Score (HUS)" reflects the theoretical adherence to "decision justifiability" and ensures that technology output conforms to the utilitarian principles of rationality and transparency.

[0044] The beneficial effects of this technical solution are shown in Table 1: It stems from systematic innovation—combining utilitarian utility theory with probabilistic modeling, ethical constraints, human supervision, and interpretable mechanisms.

[0045] Table 1: The key technical points of this technical solution are as follows: Ethics and decision-making algorithm integration mechanism 1.1 Integrating the utilitarian utility evaluation model with virtue ethics constraints and the principle of justice, a computable two-layer optimization framework is formed; 1.2 Technically, the decision balance is achieved through the dual constraints of "utility function + ethical correction layer", which is a structure that traditional reinforcement learning models do not have.

[0046] Risk trade-offs and long-term decision optimization mechanisms 2.1 Embed Monte Carlo random sampling and expected utility calculation in ethical decision-making to achieve long-term risk prediction and social cost assessment; 2.2 This addresses the problem that existing AI ethics models only focus on short-term optimization and ignore long-term consequences. 3. Automatic Ethical Revision Layer 3.1 When the utility maximization result violates the rights threshold, the system automatically executes a penalty function or a veto mechanism. 3.2 This modification layer is a tangible control process, rather than an abstract moral concept, and belongs to a protectable technical structure.

[0047] 4. Weight Learning and Adaptive Fairness Correction Mechanism 4.1 Dynamically adjust the weight parameters in the utility function based on historical behavioral data and social feedback.

[0048] 4.2 The ability to learn and adapt to ethical trade-offs in a multicultural environment is a part that has substantial algorithmic innovation.

[0049] Human-in-the-loop interface (HITL) 5.1 Implement ethical oversight through a system-level programmable interface (manual confirmation, suspension, and approval). 5.2 Improve the controllability and explainability of decision-making to lay the foundation for the implementation of actual human-machine collaboration scenarios.

[0050] Complete closed-loop system structure 6.1 Input → Utility Assessment → Risk Trade-off → Correction Layer → Weight Learning → Human Verification → Output.

[0051] 6.2 No similar integrated schemes were found in the literature search for this multi-layer closed-loop structure, demonstrating comprehensive inventiveness.

[0052] The technical protection points of this technical solution are as follows: 1. A decision-making system structure (holistic system) based on utilitarian utility and ethical constraints: The technical formal elements include a complete system architecture and information flow of modules M1–M7; and declare the functions and input-output relationships of each module.

[0053] 2. Weighted model and index calculation method of utility evaluation algorithm: The mathematical formula is U(state)=Σwᵢ×Iᵢ; dynamic adjustment mechanism of weights.

[0054] 3. Monte Carlo simulation algorithm for risk trade-off mechanism: a control algorithm that performs weighted sampling of the probability distribution of different decision outcomes.

[0055] 4. Implementation of moral constraint processing logic and penalty function in the correction layer: including veto triggering mechanism and penalty correction algorithm, threshold conditions and minimum regret value formula.

[0056] 5. Adaptive update algorithm for weight learning module (α, β, γ parameter mechanism): learning rate, fairness correction, risk regularization; update formula and constraint Σwᵢ=1.

[0057] 6. Decision interruption and confirmation logic of Human-in-the-Loop Interface (HITL): The technical implementation forms are software / hardware trigger signals, approval flags, and manual feedback storage.

[0058] 7. Decision Result Interpretation Module: Used for interpretability output and ethical tracing reports; emphasizes system transparency.

[0059] 8. Comprehensive process of technical methods: including the following steps: data input → utility assessment → risk trade-off → ethical revision → manual verification → result output.

[0060] The implementation principle of an intelligent agent ethical decision-making system based on a utilitarian framework in this application embodiment is as follows: Based on a computable ethical decision-making chain of "utility maximization - risk assessment - ethical constraints - human supervision - adaptive learning", the system realizes the intelligent agent to generate controllable, reliable and ethically constrained decision outputs in complex scenarios through mathematical utility expression, probability modeling, ethical threshold correction and weight self-learning mechanism. The system first maps the scenario state to a quantifiable "comprehensive utility value" according to utilitarian theory. The utility evaluation module performs weighted fusion of well-being indicators such as satisfaction S, life potential L, and fairness index F to construct a utility expression of U(state)=Σwi×Ii, where wi is a learnable weight vector that satisfies Σwi=1. This mathematical step transforms abstract ethical values ​​(such as well-being and fairness) into numerical representations that can be directly processed by computers, essentially realizing a "vectorized ethical value function," providing basic data input for subsequent risk modeling and ethical correction. The risk trade-off module, based on the von Neumann-Morgenstein expected utility theory, calculates the long-term expected utility EU=Σpi×Ui by sampling the probability distribution of future outcomes. It then uses Monte Carlo random simulation to construct uncertainty scenarios, reflecting long-term consequences through the statistical behavior of a large number of samples, thereby forming a quantitative estimate of risk. Simultaneously, a risk penalty term λ×Var(EU) is introduced into the optimization objective to suppress high-volatility strategies, ensuring that the overall utility pursues short-term optimality, which also aligns with the utilitarian principle of prioritizing long-term well-being. When the utility maximization outcome may infringe on individual rights or violate ethical principles, the correction layer module, based on the constraints of virtue ethics and justice theory, introduces a veto (a veto line) and penalty (a penalty term) mechanism, through Uadj = U − The λ×Penalty(R) correction rule suppresses or directly negates schemes that violate the rights threshold, transforming abstract ethical principles into executable control logic. This achieves a hierarchical balance between "algorithmic utility" and "inviolable rights," which differs from the principles of traditional machine learning models. The system introduces an interruptible mechanism in the ethical correction stage. When a potential ethical risk is detected, a human intervention signal is triggered through the interface module. The human-in-the-loop module confirms, brakes, or authorizes based on penalty / veto events, giving the final strategy a formalized human review process. This achieves "controllable automated ethical decision-making," ensuring that humans retain the final decision-making power even in extreme or unpredictable situations in the decision-making environment.The weight learning module dynamically adjusts the weights of each ethical indicator based on feedback data, the ethically corrected utility value Uadj, and social preferences using the update formula w' = w + α(∇U + β⋅fairness − γ⋅risk), where α is the learning rate, β is the fairness correction coefficient, and γ is the risk factor. The risk regularization coefficient essentially enables adaptive trade-offs across cultures and scenarios, making the utility function no longer fixed but evolving with social feedback, thus improving the model's cultural universality and fairness. In the output phase, the system utilizes causal tracking and local explanation techniques to quantify and decompose utility contributions, risk items, correction triggers, and weight changes, generating auditable explanation reports. Users can view the complete decision-making chain through text or graphical interfaces, including the basis and correction logic for each step, ensuring decisions conform to the utilitarian "publicizability principle," enhancing system transparency and improving regulatory feasibility. The system re-inputs the execution results and environmental feedback into the utility evaluation module, forming a complete closed loop of "evaluation—risk—correction—learning—human supervision—explanation—feedback." Through iterative iteration, weights converge to a more reasonable distribution over time, ethical thresholds are continuously verified, and decision stability is gradually improved. This closed-loop mechanism ensures the system develops a truly "dynamic ethical self-regulation capability" in long-term operation, exhibiting evolutionary ethical optimization characteristics. The ethical value of this technical solution can be calculated, the risk can be quantified, the right threshold can be controlled, human supervision can be programmed, and the system decision-making can be explained and evolved, thus constituting a technically feasible, structurally reproducible, and logically rigorous intelligent agent ethical decision-making system.

[0061] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. An intelligent agent ethical decision-making system based on a utilitarian framework, characterized in that: It includes a utility assessment module, a risk trade-off mechanism, a correction layer module, a weight learning and updating module, a human-in-the-loop interface module, an ethical rules and parameter database, and a decision result interpretation module; The utility assessment module calculates the comprehensive utility value U(state) of the current state, integrating well-being indicators such as satisfaction S, life potential L, and fairness index F, as well as weighting parameters w. i (Σw) i =1), the formula is ; The risk trade-off mechanism is combined with expected utility theory: The probabilities of different outcomes are calculated using weighted averages. The correction layer module integrates virtue ethics rules and principles of justice, as shown in the formula: , The corresponding rights threshold is met. Corresponding violation threshold; The weight learning and update module dynamically adjusts w based on historical decisions and social feedback. i Weights, the formula is The core parameters of the formula are: learning rate α, fairness correction coefficient β, and risk regularization coefficient γ. The human-in-the-loop interface triggers human confirmation or execution suspension logic before the decision output; The ethical rules and parameter database stores moral guidelines, legal provisions, and rights threshold parameters; the decision result interpretation module generates an interpretability report, explaining the basis for the decision and the rules for correction.

2. The agent-based ethical decision-making system based on a utilitarian framework according to claim 1, characterized in that: The utility evaluation module further includes a state feature extraction unit, which is used to extract a situation feature set X from the environmental input and normalize the satisfaction S, life potential L, and fairness index F based on a preset situation mapping function f(X).

3. The system according to claim 1, characterized in that: The risk trade-off mechanism further includes a risk sensitivity coefficient ρ, which is introduced through the expected utility formula EU = Σ pi×Ui to amplify high-damage outcomes, so that the final expected utility is adjusted to EU' = EU − ρ×HighRiskPenalty.

4. The system according to claim 1, characterized in that: The correction layer module includes a moral constraint judgment unit and a right threshold detection unit. If the decision scheme triggers the legal right threshold rmin, the adjustment value λ is adaptively increased to λ'=k×λ, where k is the dynamic penalty ratio coefficient.

5. The system according to claim 1, characterized in that: The Penalty(R) penalty function of the correction layer module adopts a piecewise penalty structure, including two types of segments: a linear mild penalty region and a convex severe penalty region, to achieve differentiated suppression of ethical conflicts at different levels.

6. The system according to claim 1, characterized in that: The weight learning and update module also includes a feedback credibility assessment unit, which is used to set a credibility coefficient δ for feedback information from different sources and complete the weight update based on w' = w + α(δ⋅∇U +β⋅fairness − γ⋅risk).

7. The system according to claim 1, characterized in that: The weight learning and update module reduces the historical feedback weights using a time decay function φ(t) to suppress long-term accumulated bias. The decay function is defined as φ(t) = e^(−ηt), where η is the decay rate.

8. The system according to claim 1, characterized in that: The human-in-the-loop interface module includes a multi-level intervention control unit, which has three types of human intervention: "suggestion mode", "mandatory confirmation mode" and "immediate termination mode", and automatically switches to the corresponding mode based on the ethical risk level.

9. The system according to claim 1, characterized in that: The ethical rules and parameter database adopts a hierarchical structure, including three data domains: basic ethical rules layer, contextualized rules layer, and legal normative style layer, and supports online parameter updates and rule version retrospective functions.

10. The system according to claim 1, characterized in that: The decision result interpretation module adopts a causal chain interpretable model to generate a structured interpretation report that includes a utility calculation chain, a risk correction chain, and a right constraint chain, while also outputting the corresponding credibility score and key decision node records.

Images