Show HN: PPO agent reduces elevator wait times by 84% vs. classical dispatching (github.com)

A recent project by Jonas Brahmst has demonstrated that a Proximal Policy Optimization (PPO) reinforcement learning agent can significantly outperform a classical Destination Dispatching algorithm in managing elevator systems. In a custom-built Gymnasium simulation modeling a 20-floor building with four elevators, the PPO agent reduced average passenger wait times by an impressive 84.5%, decreasing from 600.61 steps to just 93.06. This achievement highlights the potential of reinforcement learning to optimize operations in environments typically governed by deterministic algorithms, which may struggle to adapt to dynamic traffic patterns. The comparison was meticulously designed, with both agents evaluated under identical conditions across 1,000 independent episodes. The PPO agent not only delivered faster service but also achieved a 7% increase in the number of completed passenger journeys per episode, proving that it learned to prioritize passenger needs effectively. This study illustrates the feasibility of leveraging reinforcement learning to improve operational efficiency in real-world applications like elevator dispatching, paving the way for future advancements in smart building technologies. Additionally, ongoing enhancements to the simulation, including the introduction of realistic elevator dynamics and capacity constraints, promise even deeper insights into the benefits of AI-driven dispatch strategies.