DNA演化模糊系統應用於移動機器人控制

Abstract

遺傳演算法是以隨機多點同時搜尋的方式，而非傳統單點依序搜尋方式，因此可以避免侷限在某一區域的最佳解上，而得到整體區域的最佳解。然而，傳統遺傳演算法在演化模糊邏輯系統內部的架構與參數當中，其性能及族群大小有著密切關係等限制。為了改善此限制，於是以DNA-Based演化演算法來取代傳統的遺傳演算法，在實現上與傳統遺傳演算法一樣需要精確的模擬器與大量演化時間。 本論文提出以整合DNA-Based演化模糊邏輯系統及Q-Learning之適應性學習方法，以使移動機器人能適應實際複雜的環境。藉由Q-Learning來輔助DNA-Based演化模糊邏輯系統的方式，利用Q-Learning在實際環境互動當中產生最佳行為，且不需要建構環境動態模型，以避免因精確模擬器不易建立而造成模擬器和實際環境的誤差。由模擬結果證實，Q-Learning提供DNA-Based演化的過程當中適應函數所產生的方式，在不需要任何環境的座標資訊，有效提昇與實際環際互動當中的精確度。 最後，將本論文在模擬器上所演化完成的參數應用於sputnik移動機器人，以驗證實際環境互動的可行性與效能。Genetic algorithm(GA) is a way which many random points are being searched at the same time, and which is not a traditional single point being searched in sequence. Therefore the GA can avoid limiting the optimal solution of someone area. However, the performance of the traditional GAs is closely related with the population size for the evolvement of the parameter values of the fuzzy logic system . In order to improve this limit of the GA, a DNA-Based evolution algorithm can replace the traditional GAs, regardless of the DNA-Based evolution algorithm or the traditional GAs, the implementation requires a precise simulator and a significant amount of time. This thesis presents an adaptive learning approach of integrating DNA-Based evolution fuzzy logic system and Q-learning to enable a mobile robot to adapt a real and complex environment. Q-learning is used to assist the method of the DNA-Based which evolves the fuzzy logic system. Taking advantage of the optimal behavior of Q-learning in real environment and do not need to establish the environmental dynamic model. Thus, the optimal behavior of Q-learning in real environment can avoid the error between the simulator and real environment because the precise simulator is difficult to design. According to the simulation results, Q-learning do not need any environmental coordinate information and effectively promote the accuracy in real environment. Finally, one experiment for sputnik mobile robot using complete parameter in simulator has performed to demonstrate the feasibility and the performance in real environment.