Hybrid Models Of Fuzzy Artmap And Qlearning For Pattern Classification

• Main Author: Navan, Farhad Pourpanah
• Format: Thesis
• Language: English
• Published: 2015
• Subjects: T Technology; TK7800-8360 Electronics
• Online Access: http://eprints.usm.my/41031/1/Hybrid_Models_Of_Fuzzy_Artmap_And_Qlearning_For_Pattern_Classification.pdf; http://eprints.usm.my/41031/
• Institution: Universiti Sains Malaysia

Pengelasan corak adalah salah satu isu utama dalam pelbagai tugas pencarian data. Dalam kajian ini, fokus penyelidikan tertumpu kepada reka bentuk dan pembinaan model hibrid yang menggabungkan rangkaian neural Teori Resonan Adaptif (ART) terselia dan model Pembelajaran Pengukuhan (RL) untuk pengelasan corak. Secara khususnya, rangkaian ARTMAP Kabur (FAM) dan Pembelajaran-Q dijadikan sebagai tulang belakang dalam merekabentuk dan membina model-model hibrid. Satu model QFAM baharu terlebih dahulu diperkenalkan bagi menambahbaik prestasi pengelasan rangkaian FAM. Strategi pruning dimasukkan bagi mengurangkan kekompleksan QFAM. Bagi mengatasi isu ketidak-telusan, Algoritma Genetik (GA) digunakan bagi mengekstrak hukum kabur if-then daripada QFAM. Model yang terhasil iaitu QFAM-GA, dapat memberi ramalan berserta dengan huraian dengan hanya menggunakan bilangan antisiden yang sedikit. Bagi menambahkan lagi kebolehtahanan model-model Q-FAM, penggunaan sistem agenpelbagai telah dicadangkan. Hasilnya, model gugusan QFAM berasaskan agen dengan ukuran percaya dan kaedah rundingan baharu telah dicadangkan. Pelbagai jenis masalah tanda-aras telah digunakan bagi penilaian model-model gugusan dan individu berasaskan QFAM. Hasilnya telah dianalisa dan dibandingkan dengan FAM serta model-model yang dilaporkan dalam kajian terdahulu. Sebagai tambahan, dua daripada masalah dunia-nyata digunakan bagi menunjukkan kebolehan praktikal model hibrid. Keputusan akhir menunjukkan keberkesanan modul berasaskan QFAM dalam menerajui tugas-tugas pengelasan corak. ________________________________________________________________________________________________________________________ Pattern classification is one of the primary issues in various data mining tasks. In this study, the main research focus is on the design and development of hybrid models, combining the supervised Adaptive Resonance Theory (ART) neural network and Reinforcement Learning (RL) models for pattern classification. Specifically, the Fuzzy ARTMAP (FAM) network and Q-learning are adopted as the backbone for designing and developing the hybrid models. A new QFAM model is first introduced to improve the classification performance of FAM network. A pruning strategy is incorporated to reduce the complexity of QFAM. To overcome the opaqueness issue, a Genetic Algorithm (GA) is used to extract fuzzy if-then rules from QFAM. The resulting model, i.e. QFAM-GA, is able to provide predictions with explanations using only a few antecedents. To further improve the robustness of QFAM-based models, the notion of multi agent systems is employed. As a result, an agent-based QFAM ensemble model with a new trust measurement and negotiation method is proposed. A variety of benchmark problems are used for evaluation of individual and ensemble QFAM-based models. The results are analyzed and compared with those from FAM as well as other models reported in the literature. In addition, two real-world problems are used to demonstrate the practicality of the hybrid models. The outcomes indicate the effectiveness of QFAM-based models in tackling pattern classification tasks.