A Hybrid Structure of Fuzzy Neural Networks Generation via Q-Learning

• 作 者: Meng Joo Er
• 机 构: School of Electrical and Electronic Engineering, Nanyang Technological University, Singapore, 639798
• 出 处: International Journal of Systems and Control
• 发表日期: 2007年01期
• 关键词: automatically dynamically superiority supervised robot adjusted embedding salien

In this paper, a hybrid approach termed Biased Dynamic Self-Generated Fuzzy Q-Learning (BDSGFQL) for automatically generating Fuzzy Neural Networks (FNNs) is proposed. In the proposed method, an FNN is generated via the Q-learning and the embedded human expert knowledge. The human expert knowledge is embedded as a bias of the system according to the condence level of the knowledge. The novel BDSGFQL methodology can also automatically create, delete and adjust fuzzy rules according to the evaluations of the entire system as well as the individual fuzzy rules. The salient characteristics of the BDSGFQL approach are: 1) Capable of embedding expert knowledge according to the condence level; 2) Capable of structure self-identication and automatic parameter estimation and modication; 3) FNNs can be quickly generated without supervised learning; 4) Fuzzy rules can be created, adjusted and deleted dynamically; 5) Membership functions of an FNN can be dynamically adjusted according to the evaluation of reinforcement learning. Simulation studies of a wall-following task by a mobile robot demonstrate the superiority of the proposed method.

In this paper, a hybrid approach termed Biased Dynamic Self-Generated Fuzzy Q-Learning (BDSGFQL) for automatically generating Fuzzy Neural Networks (FNNs) is proposed. In the proposed method, an FNN is generated via the Q-learning and the embedded human expert knowledge. The human expert knowledge is embedded as a bias of the system according to the condence level of the knowledge. The novel BDSGFQL methodology can also automatically create, delete and adjust fuzzy rules according to the evaluations of the entire system as well as the The salient characteristics of the BDSGFQL approach are: 1) Capable of embedding expert knowledge according to the condence level; 2) Capable of structure self-identity and automatic parameter estimation and modication; 3) FNNs can be quickly generated without supervised 4) Fuzzy rules can be created, adjusted and deleted dynamically; 5) Membership functions of an FNN can be dynamically adjusted according to the evaluator ation of reinforcement learning. Simulation studies of a wall-following task by a mobile robot demonstrate the superiority of the proposed method.