The control of a group of elevators is a difficult stochastic control problem, because of the random and unpredictable passenger arrivals. Here we propose a new method for constructing an adaptive controller for stochastic systems, by using a combination of neural networks and an on-line reinforcement learning based on stochastic approximation. This method combines an efficient supervised learning with a general reinforcement adaptation. The supervised learning is used to prepare the controller with domain-specific knowledge, and for initializing it to emulate an existing controller. The reinforcement learning is used for adaptation, with a special attention paid to allow online operation. The new method is used to develop an adaptive, optimal elevator group controller. Results of simulation tests indicate the feasibility of using the proposed method for industrial applications.
In this paper, we propose an expanding construction of neural networks to improve associative ability when we use the projection rule to memorize prototype vectors in the networks. In order to embed the prototype vectors in the high order networks, we add an arbitrary fixed pattern to each prototype vector at the memorizing process and add it to each key vector at the recalling process. The associative ability is concerned with the domain of attraction of each equilibrium corresponding to each prototype vector. We evaluate the domain of attraction of the networks and prove that the domain of attraction of the expanded networks is larger than that of the non-expanded networks. The evaluation of the domain of attraction depends only on the order of added pattern. The simulation results show the quantitative relations between the order of the networks and the domain of attraction.
A parallel processing scheme is described for dynamic control computation of a robot-arm on any number of parallel processors. The control law for dynamic control of the robot-arm is generally represented by the sum of products. Such a control law usually consists of a huge number of operations. However, some of these operations are redundant and may be reduced by factorization. This paper proposes a practical scheduling algorithm which assigns the operations of the control law to parallel processors considering how many of them could be reduced by factorization. Based on the proposed algorithm, a scheduling system is also developed by using an object oriented approach (g++). Then, a wonderful effect of the proposed algorithm is demonstrated through examples.
This paper proposes a design scheme of variable structure (VS) type model reference adaptive control systems (MRACS) which is robust to a plant with time-variable parameters. In this design scheme, the influence of time-variable parameters is regarded as a disturbance to a nominal model of the plant and is eliminated by using a VS method. Furthermore, parameters of a controller are also adjusted by using the VS method.
A design method of PID controller for a two disk type mixed sensitivity problem is proposed. Permissible sets of PID gains which satisfy constraints on the sensitivity and complementary sensitivity functions are obtained at each frequency and a solution can be obtained as the intersection of these parameter sets and an optimal solution can be obtained by this method. The datum necessary for the design is the frequency response of the plant and no mathematical model is necessary. The algorithm is simple and easy for programming. A few numerical examples are given.