基于随机森林和自适应随机排序的昂贵多目标进化算法

A random forest and adaptive stochastic ranking based evolutionary algorithm

针对昂贵约束多目标离散优化问题,提出一种基于随机森林和自适应随机排序的昂贵多目标进化算法(a random forest and adaptive stochastic ranking based multi-objective evolutionary algorithm,RFASRMOEA).为了提高代理模型对离散问题的近似精度,RFASRMOEA采用随机森林作为代理模型辅助进化算法进行搜索.同时,为提升综合性能,提出一种基于平衡适应度评估策略和自适应概率操作的自适应随机排序机制.具体地,平衡适应度评估策略利用种群迭代信息结合所设计的基于目标转移的多样性评估和基于余弦的收敛性评估,充分发掘种群个体潜力.而自适应概率操作通过动态调整随机排序机制的关注点,使得算法在前期探索更多可行域而后期迅速收敛于可行域,进而平衡约束条件的满足与目标函数优化之间的冲突.在测试问题上的实验结果表明,所提出算法在处理昂贵约束多目标离散优化问题时具有较高的竞争力.

This paper proposes a random forest and adaptive stochastic ranking based multi-objective evolutionary algorithm(RFASRMOEA)for addressing expensive constrained multi-objective discrete optimization problems.To enhance the approximation accuracy of the surrogate model for discrete problems,the RFASRMOEA employs a random forest as the surrogate model to assist the evolutionary algorithm in the search process.Additionally,to improve overall performance,the algorithm introduces an adaptive stochastic ranking mechanism based on a balanced fitness evaluation strategy and adaptive probability operations.Specifically,the balanced fitness evaluation strategy uses population iteration information and incorporates diversity evaluation based on objective transfers and convergence evaluation based on cosine similarity,thus fully exploiting the potential of individuals in the population.The adaptive probability operations dynamically adjust the focus of the stochastic ranking mechanism,allowing the algorithm to explore a wider feasible domain in the early stage and then rapidly converge to the feasible domain in the later stage,thereby balancing the trade-off between satisfying the constraint conditions and optimizing the objective functions.Experimental results on test problems demonstrate that the proposed algorithm exhibits high competitiveness in dealing with expensive constrained multi-objective discrete optimization problems.