基于迁移学习的拐点预测策略求解动态多目标优化问题

A Knee Points Prediction Strategy Based on Transfer Learning for Dynamic Multi-Objective Optimization

动态多目标优化问题具有多个相互冲突的目标,而且这些目标也受环境的影响不断变化,为了快速准确跟踪不断变化的Pareto前沿和Pareto解集,提出一种基于迁移学习的拐点预测策略(a knee points prediction strategy based on transfer learning, TKPS)。TKPS根据记忆过去时刻种群中优秀个体,使用迁移学习算法得到映射矩阵W,然后通过映射矩阵W,把当前时刻的拐点集映射到高维希尔伯特空间,从中找到下一时刻的拐点集,引导种群收敛;同时,在拐点的邻域内选出若干个伴随个体,增加种群多样性,避免种群陷入局部最优。TKPS采用8个测试函数,并与其它3个算法结果对比分析,实验结果表明TKPS算法具有更快的响应环境变化的能力。

Dynamic multi-objective optimization problems has multiple conflicting objectives that are constantly changing due to the influence of the environment, in order to track the Pareto front and Pareto solution set rapidly and accurately, a knee points prediction strategy based on transfer learning(TKPS) is proposed. According to the memory of the excellent individuals in the population at the past moment, TKPS uses the transfer learning algorithm to obtain the mapping matrix W, and then maps the knee point set at the current moment to the high-dimensional Hilbert space through the mapping matrix W to find the inflection point set at the next moment and guide the population convergence;At the same time, several adjoint individuals are selected in the neighborhood of knee point to increase population diversity and avoid population falling into local optimal. TKPS adopts eight test functions and compares the results with the other three algorithms. The experimental results show that the TKPS algorithm has the ability to respond to environmental changes faster.