Application of Reinforcement Learning in Decision Systems: Lift Control Case Study

Abstract

This study explores the application of reinforcement learning (RL) algorithms to optimize lift control strategies. By developing a versatile lift simulator enriched with real-world traffic data from an intelligent building system, we systematically compare RL-based strategies against well-established heuristic solutions. The research evaluates their performance using predefined metrics to improve our understanding of RL’s effectiveness in solving complex decision problems, such as the lift control algorithm. The results of the experiments show that all trained agents developed strategies that outperform the heuristic algorithms in every metric. Furthermore, the study conducts a comprehensive exploration of three Experience Replay mechanisms, aiming to enhance the performance of the chosen RL algorithm, Deep Q-Learning.