Reinforcement learning is to learn how to act optimally in an unknown environment. It requires only a scalar reinforcement signal as performance feedback from the environment. Q-learning is one of the famous algorithms for the reinforcement learning. This paper presents a new method that is able to treat the continuous states and actions in the Q-learning. That is because a Q-function is smoothly approximated by using regularization theory.
In this paper, a genetic algorithm using a Pareto partitioning method to multiobjective optimization problems is proposed. The purpose of the proposed method is to generate a set of non-dominated solutions that is properly distributed in the neighborhood of the trade-off surface. The genetic search using the proposed method uniformly controls the convergence of non-dominated solutions in the objective space. Simulation results show that the GA using the Pareto partitioning method has good performances better than the traditional GA approaches for several 2-objective function optimization problems and 2-objective flow-shop scheduling problems.
We discuss a property of the limiting form derived by letting the weighting matrix with output vector tend to infinity in a standard linear regulator problem. In this paper, it is shown that there exists an explicit relationship between the limiting form and decoupling control in continuos linear systems. The new designing procedure for deriving the decoupling control law is presented, using this limiting property. We try to apply the proposed procedure to control the Shinansha.
The generalized Hamiltonian systems are generalization of well-known Hamiltonian systems, which include various passive electric circuit systems as well as mechanical ones. This paper introduces the canonical transformation for the generalized Hamiltonian systems, which preserves the Hamiltonian structure of the original system and is expected to provide new insights and useful tools for analysis and synthesis of such systems. First, the class of such transformations and some of their properties are clarified. Second, we show how to stabilize the generalized Hamiltonian systems by using the transformation. This method works even when we can not stabilize them by conventional unity feedback without canonical transformation. Furthermore, it is shown that the proposed stabilization method includes the well-known one which exploits the virtual potential energy.
This paper concerns a nonlinear controller design for linear systems. We give a concrete method to design a nonlinear controller which smoothly switches linear compensators depending on the state of the plant. Then we apply the method to a plant with an input saturation and evaluate its effectiveness by simulations.
This paper proposes a control strategy based on an array representation of input-output data of a plant. The control strategy requires no traditional mathematical model such as a transfer function or a state-space equation. The plant dynamics is represented as a set of basis arrays whose elements are plant input-output data. Then a set of canonical basis arrays is derived from the set of basis arrays by a sweep-out calculation. The control input for dead-beat tracking is directly computed using the set of canonical basis arrays.
To realize brain surgery simulation system or automatic diagnostic system of brain diseases which make use of cross-sectional images taken by MRI, it is necessary to extract and recognize principal tissues in the image automatically. There have been many researches about extraction of tissues, but previous methods have weak points : that is, the process of the extraction is slow and its result is rough. In order to overcome these drawbacks, we propose a new method which enables us to extract principal tissues automatically and fast by using knowledge. This method is based on binarization and classification of regions. At the classification process, we make use of an information on shapes of ventricles and relative locations between ventricles, which are given beforehand as knowledge. Our method uses T1 and T2 weighted images taken by Spin Echo method as original images and we extract white matter, gray matter, ventricles and celebrospinal fluid other than ventricles in these images.