On Adaptive Multi-objective Optimization Methods For Elevator Group Control Systems

This dissertation mainly researches on adaptive multi-objective optimization methods for elevator group control systems.The problem of multi-objective elevator group optimizing has received extensive attention due to its high practical significance. An elevator group control system serves as a typical multi-objective system aiming at increasing the service for passengers and reducing the cost such as power consumption. So in order to find out the effective solution for this problem, multi-objective optimization methods for elevator group control systems are investigated systematically. After that, the architecture of self-adaptive multi-objective optimization is designed, in which the function of each module, in turn, is introduced. Later, this dissertation expatiates upon the principle of self-adaptive multi-objective optimization in detail.The most important step in optimizing a multi-objective system is optimizing parameters of the evaluation function which is also the core issue studied in this dissertation. The issue of optimizing parameters is then formulated by the framework of Markov Decision Process and elements are defined according to the specified field. Through the overview, analysis and comparison of self-adaptive optimization methods, two efficient algorithms based on reinforcement learning are brought forward and discussed carefully, i.e. SARSA(λ) and police gradient algorithms are employed which are then respectively integrated into tile coding function approximation. Furthermore, on the basis of the current theory research on reinforcement learning, convergence properties of the correlative algorithms are analyzed and proved by means of stochastic process, theory of matrix and theory of fixed points.The proposed methods for optimizing parameters of the evaluation function, however, have inevitably encountered the problem of the curse of dimensionality, which results in slow convergence and long training time. A model of reinforcement learning based on hidden biasing information is then established. As a result, improved methods are advanced accordingly. The experiments demonstrate that the improved methods can accelerate learning and have faster speed to converge.And the virtual environment structure for elevator group control is designed which incorporates with self-adaptive multi-objective optimizing unit. Moreover, the interfaces between different functional components are well defined. The simulation experiments are done in the virtual environment for elevator group control. Finally, with two different traffic flows used for simulating and training of algorithms, we validate the two different improved methods for optimizing parameters by experiments on elevator group scheduling task. Although the studied methods for use in the self-adaptive optimization for multi-objective elevator group control systems are different in respect to their learning abilities, they both have the advantage of adapting to dynamics of the environment such as different traffic patterns. The results also show better general performance of the studied methods in contrast to some other existing methods.