Optimizing elevator systems with deep Q-Networks

Abstract

This project aims to investigate the collective dynamics of elevator systems under varying passenger arrival rates and travel patterns. A key observation is the elevator system’s natural tendency to bunch, particularly when operating near or beyond its handling capacity. This emergent synchronization is driven by uneven passenger distribution across floors and is further intensified by the mechanics of passenger pickup. As the system transitions into synchronization, its dynamics resemble a second-order phase transition, similar to that observed in the Kuramoto model. This bunching behaviour leads to increased passenger waiting times. To address this challenge, we investigate the effectiveness of a Deep Q-Network (DQN) based elevator control strategy. While the DQN performs comparably to traditional scheduling algorithms under uniform interfloor traffic, it exhibits a distinct advantage in scenarios with directional or patterned demand. In these cases, the DQN successfully disrupts elevator synchronization, thereby reducing passenger wait times. These findings highlight the value of viewing elevator systems through the lens of complex systems and synchronization theory. Moreover, they demonstrate the potential of reinforcement learning to enhance the efficiency and responsiveness of elevator control, especially in non uniform traffic conditions. Overall, the study contributes to a deeper understanding of elevator dynamics and offers insights for the development of intelligent elevator control systems.