發展粒子引導式演化策略演算法以處理實數參數之全域最佳化問題

Abstract

遺傳演算法為一種模仿自然界演化現象所發展出的一套隨機搜尋的方法，在這些方法中，又以演化策略與粒子最佳化演算法為兩個最熱門的研究主題。這兩種方法皆為處理實數參數最佳化的問題，但其搜尋的行為卻迥然不同。藉由觀察演化策略與粒子演算法的搜尋行為，我們發現兩者皆有其優點與缺點。在演化策略中，起因於隨機擾動，突變運算子缺乏一個明顯的機制來引導搜尋行為至預期會有較佳解的區域，除此之外，個體在搜尋空間中的移動行為缺乏群體相互協調的機制，然而，其強大的篩選機制可將較好的個體保留至下一代；在粒子演算法中，其搜尋機制使用群體合作的概念，每一個個體在搜尋空間的移動行為將直接朝向預期會有較佳解的方向前進。本研究的目的試圖將演化策略與粒子演算法做一個概念層級上的結合，我們引進粒子演算法中群體合作的機制，以此為概念來對於演化策略中偏向於隨機搜尋的突變運算子進行突變方向上的引導。本文將提出一個新的突變機制，稱之為引導式突變，結合引導式突變於傳統演化策略的搜尋架構下，進而發產出粒子引導式演化策略的最佳化方法。效能評估將建立在一組精心設計的測試平台下，由實驗結果得知所提的方法可達到良好的搜尋效能。

Evolutionary algorithms are stochastic search methods that mimic the metaphor of natural biological evolution. Among of these algorithms, evolution strategy (ES) and particle swarm optimization (PSO) are two of the most popular research topics. Both of ES and PSO are deal with real-parameter optimization problems but have different search behaviors. By observing the search behaviors of ES and PSO, we find both of them have strengths and weaknesses. In ES, the mutation operator lacks an explicit mechanism to guide the search into promising direction due to random variance. Moreover, there is no coordination in the movement of individuals within the search space. However, the powerful selection procedure allows solutions with superior characteristics to pass these from generation to generation. In the PSO, the search mechanism used the swarm cooperation concept. Each particle will move toward the direction which is expected to be good. The objective of this article is tried to combine ES and PSO at the concept level. We introduce the concept of swarm cooperation of PSO into the mutation operator of ES for reducing the disturbance of mutation in the mutation direction. We proposed a new mutation operator called guided mutation. Combining the guided mutation into the traditional ES framework and develop a new optimization algorithm called particle swarm guided evolution strategy. Numerical experiments are conducted on a set of carefully designed benchmark functions and demonstrate good performance achieved by the proposed methodology.