Scholarlink

<p>Constrained optimization in computational engineering treats feasibility as physical admissibility, encoding the value judgments of system designers as neutral facts. Many engineering failures on record trace to constraints that were never formalized because the discipline lacked tools to handle them. This paper develops multi-domain feasibility, in which a solution must satisfy three epistemologically distinct constraint classes at once: physical (discovered through natural law), statistical (inferred through data and uncertainty quantification), and normative (negotiated through social, ethical, and legal processes). The classes differ in formulation, temporal dynamics, and the revision procedures available when no solution satisfies all of them. Two concepts organize the analysis. The moral drift problem is that normative constraints evolve with legal frameworks and social consensus while optimization systems treat their constraints as fixed after training. The ethical singularity is the threshold of system autonomy and operational speed beyond which retrospective normative governance becomes structurally impossible; the paper formalizes it with a governance-bandwidth metric from control theory and derives a rank-deficiency condition for regulatory architectures. These ideas motivate the constitutional optimizer, a meta-governance architecture that treats the precedence ordering among constraint classes, the authority to revise it, and the procedure for handling infeasible intersections as first-class design problems, drawing on lexicographic decision theory, Rawlsian political philosophy, and public law. The paper maps the framework against the EU AI Act, the NIST AI Risk Management Framework, and ISO/IEC 42001, identifying which elements they instantiate and which they leave unaddressed. For a computational-engineering readership the paper states the framework as a canonical lexicographic program with explicit feasibility tests, gives the constitutional optimizer and its infeasibility diagnosis as algorithms, positions it against robust optimization, chance-constrained programming, safe reinforcement learning, constrained Markov decision processes, control barrier functions, and multi-objective optimization, proves a control-theoretic governability limit on a worked model, and examines the non-convexity and worst-case hardness of the combined feasible set. The reframing moves normative questions earlier in the engineering process, widens participation in system design, and reveals a kind of failure that current benchmarks miss.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Embedding ethical and safety constraints into data-driven optimization for computational engineering systems

  • 期刊论文

  • 发表日期：

    2026年7月7日

查看全文
  • Babu George

摘要

Constrained optimization in computational engineering treats feasibility as physical admissibility, encoding the value judgments of system designers as neutral facts. Many engineering failures on record trace to constraints that were never formalized because the discipline lacked tools to handle them. This paper develops multi-domain feasibility, in which a solution must satisfy three epistemologically distinct constraint classes at once: physical (discovered through natural law), statistical (inferred through data and uncertainty quantification), and normative (negotiated through social, ethical, and legal processes). The classes differ in formulation, temporal dynamics, and the revision procedures available when no solution satisfies all of them. Two concepts organize the analysis. The moral drift problem is that normative constraints evolve with legal frameworks and social consensus while optimization systems treat their constraints as fixed after training. The ethical singularity is the threshold of system autonomy and operational speed beyond which retrospective normative governance becomes structurally impossible; the paper formalizes it with a governance-bandwidth metric from control theory and derives a rank-deficiency condition for regulatory architectures. These ideas motivate the constitutional optimizer, a meta-governance architecture that treats the precedence ordering among constraint classes, the authority to revise it, and the procedure for handling infeasible intersections as first-class design problems, drawing on lexicographic decision theory, Rawlsian political philosophy, and public law. The paper maps the framework against the EU AI Act, the NIST AI Risk Management Framework, and ISO/IEC 42001, identifying which elements they instantiate and which they leave unaddressed. For a computational-engineering readership the paper states the framework as a canonical lexicographic program with explicit feasibility tests, gives the constitutional optimizer and its infeasibility diagnosis as algorithms, positions it against robust optimization, chance-constrained programming, safe reinforcement learning, constrained Markov decision processes, control barrier functions, and multi-objective optimization, proves a control-theoretic governability limit on a worked model, and examines the non-convexity and worst-case hardness of the combined feasible set. The reframing moves normative questions earlier in the engineering process, widens participation in system design, and reveals a kind of failure that current benchmarks miss.