建構一個學習分類元系統的改良架構

Constructing an Improved Framework of Learning Classifier System

林豐澤(Feng-Tse Lin)

《智慧科技與應用統計學報》 5卷1期 (2007/06) Pp. 15-27

https://doi.org/10.29807/JTITAS.200706.0002

摘要

分類元系統是建構在基因演算法之上的一種機器學習系統，它依照簡單”condition-action”的規則對環境產生動作，當環境有所反應，系統會修改內部的規則，這些規則通稱為分類元。適應度高的分類元可以生存下去，適應度低的分類元會逐漸被淘汰。因此，分類元系統是以規則為單元，透過訊息傳遞方式，以大量並行處理及接受環境回饋來達成學習的一種機器學習機制。Holland早期提出的分類元系統曾被應用到股票管理、產生決策、以及存貨管理等問題上，然而它的學習機制及系統架構有些缺點，造成實用上的困難。於本文中，我們針對Holland分類元系統的缺點提出改良的系統架構，此架構具有增強式學習能力，能夠適應動態環境的變化來增強或減弱分類元的適應性，甚至可以產生高適應性的新分類元，達成機器學習的目的。本論文先就改良復分類元系統的功能、架構、以及作業方式提出說明，最後再討論加入模糊規則子系統的執行效能。

關鍵字

分類元系統 ； 基因演算法 ； 機器學習 ； 增強式學習 ； 模糊規則

並列摘要

Learning Classifier System (LCS) is a machine learning system based on the scheme of genetic algorithms. LCS receives outside information through detectors according to some simplified condition-action rules to generate actions against the environment. Any reaction come from the environment causes LCS to modify its internal rules. These rules are then called classifiers. The classifiers with higher fitness will survive and the classifiers with lower fitness will gradually die. Hence, the learning mechanism in LCS is based on classifiers in which massively parallel, message passing, and environmental feedback are the way of learning for classifiers. In the earlier days, Holland's classifier was successfully applied to stock managements, decision-making, and inventory control. However, LCS has difficulties used in practical applications due to some deficiencies in its learning mechanism and architecture. In this study, we propose an improved framework of Holland's classifier with reinforcement learning capability. This improved model adapts the environmental changes to intensify or diminish the strength of classifiers and even produces high strength classifiers for machine learning purpose. In this paper, we first present the function, the framework, and the procedure of the proposed improved framework of Holland's classifier, and then introduce the fuzzy rule subsystem and also discuss the system's learning performance.

並列關鍵字

Learning classifier system ； genetic algorithms ； machine learning ； reinforcement learning ； fuzzy rule