以詢問式學習法改良粒子群演算法

Abstract

動機：粒子群最佳化(particle swarm optimization, PSO)演算法是目前人工智慧研究方面極受重視的子領域之ㄧ。其求解快速有效率近年來在國際間得到認同與肯定，也有許多學者提出相關缺失，本論文探討PSO演算法的缺點並提出解決的方法，以提升其效能。 作法：本研究結合所發展的詢問式學習法則，試圖透過含糊地帶的加強學習，擴展尋優廣度，藉此提高演算法整體的求解精準度。傳統粒子群演算法容易使粒子陷入局部最佳解陷阱，使所有粒子走向錯誤的方向。我們的方法則可以走出此陷阱，提高搜尋到真正解答的可能性。論文中以容易理解的二維函數配合二維的實驗結果來做說明。 成果：本論文為首度將詢問式學習觀念應用到粒子群演算法，實驗結果顯示，我們的方法在整體表現上都優於傳統的PSO。透過粒子群主動進行詢問式學習，以目前尚含糊不清的解空間區域做為學習的指引，本研究所提出的機制確實改善粒子群演算法陷入局部最佳解缺點，並在整體求解精準度獲得改善。

Motivation: PSO (particle swarm optimization) is one of the most important research topics on artificial intelligence. PSO still remain some disadvantages. This paper tries to discuss the disadvantages of PSO and to find a solution for improving its performance. Method: We apply the query-based learning method proposed in our previous papers to PSO. It leads the particles to extend their search area. Thus, not only the precision of solution but also the time consumed is improved. We visualize the mechanism through a two-dimension PSO and verify the mechanism by several functions. Conventional PSO usually leads the particles go into the wrong direction of evolution. To resolve this drawback, when particles tend to converge, we spread some particles into ambiguous solution space. Furthermore, PSO has been well improved. Achievement: This thesis, in our knowledge, is the first study that applies the QBL concept in Particle Swarm Optimization. The experiment results show the proposed approach is able to prevent the system from falling into local optimal and improve the performance of PSO.