機器應用的強化式學習研究

Abstract

本篇論文提出將強化式學習(Reinforcement Learning)結合深度神經網路套用在機器手臂物件夾取任務的學習方法，透過在環境中進行嘗試並取得相對應的獎勵，Agent 可以學到判斷物件位置並選擇所正確的動作。我們提出了將影像及額外資訊作為輸入的訓練方法，並藉由額外資訊提升訓練的穩定度；同時，透過結合課程式學習(Curriculum Learning)設計由簡單到困難的訓練環境，以及藉由模仿式學習(Imitation Learning)有效的利用領域知識來加速訓練。在兩個不同的實驗環境中都顯示所提出的方法可以得到非常好的效果。

In this thesis, a learning method that combines reinforcement learningwith deep neural network is applied to the robotic grasping problem.Through interacting with environment and receive corresponding rewards,agent is able to learn to locate target object location and to generate actions that can grasp the target object. We propose a training method that takes image and extra low dimensional information as input. With extra low dimensional information, the training process is more stable. Meanwhile,with the concept of curriculum learning, we propose a guideline to design the environmental settings from easy to complex. Furthermore, we provide an approach to exploit domain knowledge through imitation learning in order to accelerate the training process. We applied our method to 2 tasks:reacher and cube grasping. In reacher, the convergence problem was solved and got an average score of 10 with our method. In cube grasping, the success rate reached 98% in relatively short training time.