Order Formation in the Learning and Adaptive System and Its Application to Mechanical Control

研究了通过学习或自适应系统来实现复杂机械系统的有用功能，考虑了不确定复杂系统之间的竞争或合作。在研究中，考虑了确定性和非结构化不确定性系统、确定性和结构化不确定性系统以及随机不确定性系统，提出了几种基于神经网络的鲁棒控制器设计方法。我们发现，在考虑被控对象不确定性的方法训练的控制器的性能与鲁棒性之间的关系中，形成了一个非常简单的阶次，即幂定律。从最优化的角度来看，学习设计鲁棒控制器的方法等同于多目标优化问题。形成简单有序的主要原因之一，它广泛存在于许多自然系统…中更重要的是，在多目标优化的状态空间中形成帕累托解集。因此，针对环境不确定性的竞争学习以提高鲁棒性与简单有序形成有着密切的关系。固定的鲁棒控制器对所考虑的鲁棒性是鲁棒的，但它们对不确定性可能是脆弱的，而在设计控制器时从未考虑到不确定性。因此，不仅要针对考虑到的不确定性进行竞争性学习，而且要注意适应性。研究了模块化控制器自适应的任务分解及其排序方法。提出了一种能够分解复杂机械系统任务的自适应模块强化学习算法。将所提出的方法应用于无人直升机的飞行控制器设计，特别是高度控制器的设计。在控制器设计中，考虑了垂直风的不确定性。我们发现，自适应模块分解了高度控制任务，形成了简单的顺序。该方法根据直升机的运动方向对控制器的任务进行分解，以提高对垂直风的鲁棒性，并通过学习自适应地获得适合每个任务的控制器。较少

Research on designing control system to realize useful functions of complicated mechanical systems by learning or adaptive system considering competition or cooperation between the uncertain complex systems was performed.In our researches, systems with deterministic and un-structured uncertainties, systems with deterministic and structured uncertainties, and systems with stochastic uncertainties were considered, and we proposed several design methods of robust controllers by using neural networks. We found that the very simple order that is power law is formed in the relation between the performance and the robustness of the set of controllers which are trained by the proposed methods in which uncertainties of the controlled object are taken into account. From the viewpoint of the optimization, design methods of robust controllers by learning is equals to the multi objective optimization problems. One of main causes of forming simple order, which is widely found in many natural systems … More , is forming of the set of Pareto solutions in the state space of multi-objective optimization. Therefore the competitive learning against the environmental uncertainties to improve robustness has close relation to simple order formation. Fixed robust controllers are robust against the considered robustness but they may be fragile to uncertainties which are never considered in designing controllers. Therefore not only the competitive learning against considered uncertainties but also adaption is important. Task decomposition by adaptation of modular controller and its order formation was investigated. We proposed an reinforcement learning algorithm for adaptive modules, which are able to decompose tasks of the complicated mechanical system. We applied the proposed method to design flight controllers, especially altitude controllers, for an unmanned helicopter. In designing controllers, vertical wind was considered as uncertainties. We found that the adaptive modules decomposed the task of the altitude control and the modules formed the simple order. As a result, controller's task was decomposed according to the direction of the helicopter's motion to improve the robustness against vertical wind and suitable controllers for each tasks were adaptively obtained by learning. Less