分類元系統構成中之長期記憶形成與類推式認知學習研究

Abstract

通用型分類元系統雖然在理論上及許多的研究報告裡有著非常出色的表現，但是從實際的執行運轉中，我們卻發現通用型分類元系統會產生失憶的現象。因為，傳統的分類元系統是假設環境的變化是微量且平順的。所以，當環境改變得太激烈時，分類元系統過去所學的經驗無法派上用場，也沒有辦法將目前從環境中獲取到的知識應用在未來類似的狀況。因此，在本篇論文中，為了消除失憶的症狀，我們嘗試整合通用型分類元系統與長期記憶等元件，提出一個擁有類推式認知學習能力的分類元系統模型。本篇論文的研究目標包括： 1.解釋分類元系統失憶的現象從何而來。 2.證明加入長期記憶及其它元件後，分類元系統不僅消除健忘的症狀，還因此額外擁有類推式認知學習的能力。 3.評估擁有長期記憶的分類元系統可以改善多少的學習績效。 在本篇論文中，我們將實作影像辨識與分類的問題來驗證此模型。從實驗中得知，擁有長期記憶的分類元系統不但成功地解決健忘的問題，而且執行效能表現優異，所需要的額外系統資源也不多。此外，藉由改變系統的一些關鍵參數，我們也觀察到一些有趣的特性，及這些特性所衍生出來的各種變化。

Although general-purpose classifier system has an impressive level of performance both in theories and its empirical behavior, we found it lack the ability to improve its performance gracefully and incrementally in the long run with a limited population size, and it produced forgetfulness when coming to executing. The standard classifier system is implemented in a hypothetically smooth environment that has slight change. Thus, the classifier system is still brittle to the extent that they are unable to benefit from the obtained problem solving skills when changes are drastic or dramatic. By dynamically modifying a small set of rules in the population, the problem of rapidly forgetting inactive information over a short period of time is unavoidable and intolerable for real world system. As a result, in this thesis, we attempt to integrate long-term memory and other components into general-purpose classifier system to present a learning production system model that has ability to learn through analogical recognition. The purpose of this paper is: 1. To explain the reason that classifier system produces forgetfulness. 2. To prove that classifier system can not only solve the problem of forgetfulness but also additionally possess the ability to learn through analogical recognition. 3. To evaluate the learning efficiency of classifier system having long-term memory. In this thesis, we implement adaptive image learning and classification to verify this model. According to the result of the experiment, the classifier system with long-term memory succeeds in eliminating forgetfulness and have excellent performances with not much added systematic resources. Moreover, we observe a variety of changes deriving from several interesting characteristics of the system by changing some significant parameter.