基於增強式學習架構下使用合作式粒子群最佳化演算法建構模糊類神經控制系統

Abstract

在機器學習中，增強式學習有監督式學習比不上的優點，但卻有搜尋空間較大的問題。而粒子群最佳化演算法有搜尋範圍大的優點，但是以監督式學習做為機器學習方法，使得在應用層面上受到限制；且在演化過程會有「前進兩步，退後一步」的問題，導致在需要維度較大的系統中，沒有搜尋到全域最佳解的能力。因此希望能利用合作式粒子群最佳化演算法以增強式訊號為適應值，使系統應用層面更廣泛。本篇論文提出一個架構，使用增強式學習機制建構模糊類神經控制系統，在學習中使用合作式粒子群最佳化演算法來學習模糊類神經網路控制器的參數。且利用平行雙倒單擺系統和球桿系統的控制問題為例，測試本系統達成控制目標與學習的功能。由模擬結果得知，本論文所提出演算法可達到滿意的結果。

In the machine learning, the reinforcement learning has advantages that supervise learning can’t reach, but it needs a larger searching space. The particle swarm optimization (PSO) algorithm has the advantage of large searching space, but it nowadays uses supervised learning as the method for machine learning, which limits its application. In addition, in the process of calculation, a “two-step-forward, one-step-back” pattern may emerge. This will lead to a disability to search for the global best solution in a system with larger dimension. Therefore, the present study hopes to use the cooperative particle swarm optimization (CPSO) algorithm with reinforced signal as fitness to make the system application more general. In this thesis, a neuro-fuzzy control system using reinforcement learning is constructed. The system uses CPSO algorithm to learn the parameters of a neural-fuzzy controller. Moreover, the system took, for example, the control of the tandem pendulum system and the ball and beam system to test its capability to implement the target control and to learn. From the simulated results, the algorithm proposed in this study can achieve satisfactory results.