|
In many real-world production lines, there is a production system, which is mainly composed of a production station, such a system is called conveyor-serviced production station (CSPS).This paper is concerned with the optimal control problem of CSPS system, and the objective is to obtain the optimal long-run expected cost of the system by choosing a suitable look-ahead control strategy. Theoretically, the optimization problem can be treated by exact solution techniques. However, a major difficulty, so-called“the curse of modeling”, arises in such solution. Potential-based online policy iteration algorithm can avoid the curse of modeling, but it must reserve the performance values table, and looking up table leads to lacking of information generalization. At the same time, the action set of CSPS system is continuous, when it is discretized, discrete particle size will affect the optimization performance of system. So, this paper applies the cerebellar model articulation controller neural network and the online support vector machine to optimizate CSPS system on-line.First, the CMAC neural network is used to approximate continuous action Q-values of the online Q-learning. Then, it is used to approximate the Q-values or the potentials of OPI algorithm to construct OPI-Q algorithm and OPI-Qg algorithm. The simulation results show that, the algorithms which are based upon CMAC have faster learning and convergence, the average cost of system is closer to the theoretical optimization value, have a good optimization results.The online SVM is also used to approximate Q-values of the online Q-learning, and two online SVM-based Q-learning are given. One is OSVM-Q, online SVM is set for each exploration state. The other is OSVM-Q-1, only one online SVM is set for all state-action of CSPS system. Simulation results show that the optimization performance of system is improved by these algorithms.
|