Application of Q-learning to elevator dispatcidng

Abstract

This paper addresses the problem of designing an elevator dispatcher to minimize the average passenger waiting time in an office building served by a group of elevator cars. The paper shows how an adaptive dynamic programming algorithm, Q-learning, can be used for both analysis and design. The analysis gives a better understanding of the dispatching problem and the importance of various state information to making dispatching decisions. The design gives a dispatcher which outperforms a commercially representative dispatcher. The results clearly demonstrate the potential capabilities and usefulness of Q-learning for solving discrete event system control problems.