Adaptive Critic Learning Techniques for Robotic Learning Control

适应性批评学习（ACL）技术是一种自学习技术，近年来因其在解决复杂的控制/决策问题方面的巨大潜力而受到研究者的广泛关注。然而，当该技术应用于复杂的规划问题时，存在“分块”问题，这在神经工程文献中没有得到足够的重视。本课题将探讨ACL在双足行走控制中的应用，该控制涉及多个控制动作（关节），是一个复杂的问题。两个新的先进结构，都是基于ACL，将探讨。这些结构要么是分级的批评，要么是不同层次的批评与其他适应能力的组合。层次结构被期望更适合于复杂的问题。本研究也想找出在高级ACL应用中可能存在的障碍。除了技术方面，本研究将进一步探讨ACL与人类学习过程之间的关系。

Lin 9408857 The Adaptive Critic Learning (ACL) technique is one self-learning technique, which has received much attention form researchers in the recent years because of its great potential in solving difficult control/decision-making problems. However when this technique is applied to complex planning problems, there are issues of "chunking" which have not received sufficient attention in the neuroengineering literature. This project will explore the application of ACL on biped walking control, which involves multiple control actions (joints) and is a complicated problem. Two new advanced structures, both based on ACL, will be explored. These structures have either leveled critics or a combination of critic with other adaption capability at different levels. The Hierarchical structures are expected to be more suitable for complicated problems. This study also intends to find out possible roadblocks in advanced ACL applications. In addition to the technical aspect, this study will take a closer look at the relationship between ACL and Human learning process.