加強式學習於避碰運動控制之應用

Abstract

本論文將加強式學習應用在避碰運動控制的問題上。加強式學習一般可分 為直接法和間接法兩大類。本論文所要探討的內容著重在比較這二種學習 方式的異同，此外，更要評定它們應用在機器人的避碰運動控制問題上的 成果及效率。雖然經由模擬我們發現利用直接法所設計出的控制器在一般 化的能力上要比間接法來得好，但是這二種學習方法都有各自的限制及缺 點，因此在目前很難評斷那一種學習方法較佳。然而在結構不甚明確的控 制問題上，加強式學習仍然是一種有效的解決方法。 This thesis applies reinforcement learning to the collision- free motion control problem. There are two general reinforcement learning methods, the direct and indirect method. We present a direct method, the stochastic learning scheme, and an indirect method, the model-based learning scheme. The focus is on justify the effectiveness and efficiency of these schemes on the robot collision-free motion control problem. The results of simulations show that the stochastic learning scheme outperforms the model-based learning scheme, but both learning methods have their own limitations and disadvantages. It is suggested that reinforcement learning control is an effective alternative in dealing with less-structured control problems.