高计算代价动态优化问题的代理模型辅助粒子群优化算法

Surrogate-assisted particle swarm optimization for expensive dynamic optimization problems

针对同时具备动态优化与高计算代价两种特征的高代价动态优化问题,提出一种代理模型辅助的动态粒子群优化算法。为加快种群对环境变化的响应速度,给出一种基于多方向预测的种群初始化方法,用来产生多样性好且目标值优秀的初始种群;为降低代理模型的构建代价且保持其预测精度,设计一种融合目标值预测机制的代理模型更新策略。通过处理多个典型的高计算代价动态优化问题,实验结果表明,相比已有算法,所提算法可以较快地跟踪随环境变化的问题最优解。

For dynamic optimization problems and expensive optimization problems,a novel surrogate-assisted particle swarm optimization(SDPSO)is proposed for expensive dynamic optimization problems.To speed up the population′s response to environmental changes,a population initialization method based on multi-directional prediction is given to generate initial populations with good diversity and excellent fitness.To reduce the construction cost of surrogate model and maintain its prediction accuracy,an update strategy of surrogate model incorporating the fitness value prediction mechanism is designed.Experimental results on several typical expensive dynamic optimization problems indicate that the proposed algorithm can track the optimal solutions changing with the environment fast.