Reinforcement Controller Design and Implementation for an Inversed-Pendulum System Using Parallel Genetic Algorithms

• Main Authors: Yi-Fong Chang, 張奕峰
• Other Authors: 張耀仁
• Format: Others
• Language: zh-TW
• Published: 1998
• Online Access: http://ndltd.ncl.edu.tw/handle/48612581516529051275

碩士 === 長庚大學 === 機械工程研究所 === 86 === The problem of controller design for a plant with unknown dynamics is considered in this thesis. With the assumption that the available information is restricted to final evaluations of control results, the strategy of reinforcement learning was chosen to improve the controller performance. In this paper, we explore the search capability of Genetic Algorithms to the design of both a fuzzy logic controller and a neural nerwork controller (NNC), for the balance of an inverted pendulum. The inverted pendulum is chosen due to its capability to withstand system failure in the course of iterated trials. Many successful applications of genetic algorithms can be found in the literature, however, the detailed implementation choices and design factors, such as encoding, genetic operators, evolutionary parameters etc., are still open and unclear. These issues typically have great influence on final results. In this thesis, Genetic Algorithms are implemented in a parallel framework called the island model. Different evolutionary parameters are applied on the population in each island, while allowing migration to take place between islands. We implemented six islands on a network of Transputer modules to evolve for control parameters. The proposed architecture of parallel Genetic Algorithms (PGAs) is demonstrated to offer consistent and effective search capability. Furthermore, experimental results using the controllers found from PGAs are provided to justify the learning reults.