基于KTSR自适应判别的双层多目标优化算法研究

A Bilevel Multi-Objective Optimization Algorithm Based on Adaptive KTSR Judgment

摘要: 双层多目标优化问题(BLMOPs)中，上下层目标冲突程度具有高度不确定性，传统单一优化策略往往无法有效平衡求解质量与计算效率。尽管现有多形式优化(MFO)算法在求解弱冲突问题时具备较高的搜索效率，但面对强冲突与高欺骗性环境时，常规任务重构策略极易失效并导致计算资源浪费。为此，本文提出一种基于知识迁移成功率(KTSR)自适应判别的双层多目标优化算法(cGAN-BLMFO)。首先，设计以KTSR为核心的在线评估与自适应机制，实时量化辅助任务的知识传递有效性；其次，构建“双引擎”自适应求解框架：在弱冲突阶段，利用常规多形式辅助任务实现Pareto前沿的快速逼近；当两个辅助任务的KTSR同时且连续三次低于预设阈值时，算法自适应切换至条件生成对抗网络(cGAN)重构模式，借助网络判别器软化下层约束，引导种群突破欺骗性壁垒。基准测试集上的实验结果表明，该算法不仅有效克服了强冲突环境下的失效瓶颈，同时在弱冲突问题上保持了较高计算效率，提升了算法综合性能。

Abstract: In bilevel multi-objective optimization problems (BLMOPs), the highly uncertain conflict degree between upper- and lower-level objectives makes it difficult for traditional single optimization strategies to balance solution quality and computational efficiency. Although existing multiform optimization algorithms are efficient in solving weak-conflict problems, conventional task reformulation strategies are highly prone to failure in strong-conflict and highly deceptive environments, leading to severe computational waste. To address this, we propose a bilevel multi-objective optimization algorithm based on adaptive Knowledge Transfer Success Rate (KTSR) judgment. An online KTSR evaluation mechanism is designed to dynamically quantify the effectiveness of knowledge transfer from auxiliary tasks. Accordingly, an adaptive solving framework is constructed: during the weak-conflict stage, conventional multiform auxiliary tasks are utilized to rapidly approximate the Pareto front. Once the KTSRs of two auxiliary tasks simultaneously fall below a preset threshold for three consecutive iterations, the algorithm adaptively switches to a conditional Generative Adversarial Network (cGAN) reformulation mode. This mode leverages the network’s discriminator to soften stringent lower-level constraints, guiding the population to break through deceptive barriers. Benchmark experiments demonstrate that the proposed algorithm effectively overcomes the failure bottleneck in strong-conflict environments while maintaining high efficiency in weak-conflict scenarios, significantly enhancing the comprehensive optimization performance.