Elevator scheduling seems simple — pick up passengers, deliver them to their floor, minimize waiting time. But the combinatorial explosion is severe. Six elevators, 40 floors, stochastic passenger arrivals: the number of possible dispatch decisions at any moment dwarfs chess. Heuristic controllers (nearest car, zone-based dispatch) handle common traffic patterns but break down during peak flows, lunch rushes, and mixed traffic.
A reinforcement learning approach (arXiv:2507.00011) introduces three innovations that outperform traditional dispatchers: a novel action space encoding that reduces the combinatorial explosion, infra-steps that model continuous passenger arrivals between dispatch decisions, and a tailored reward signal that balances waiting time against energy consumption.
The counterintuitive finding: the RL agent learns to keep elevators idle at strategic floors rather than continuously servicing calls. During low-traffic periods, the agent parks cars at floors that minimize expected future response time — which are not the lobby, not the most-used floors, but positions that optimize coverage given the building's traffic pattern. The optimal standby configuration is learned, not designed.
This challenges the “busy is productive” assumption. Human-designed heuristics tend to keep elevators moving — returning to the lobby after a delivery, responding immediately to the nearest call. The RL agent discovers that strategic idleness reduces aggregate waiting time because well-positioned idle elevators can respond to the next call faster than a busy elevator finishing its current trip.
The broader pattern: in systems with stochastic demand and multiple servers, the optimal policy often includes deliberate inactivity — positioning resources to minimize response time rather than maximizing utilization. Idle isn't wasted; it's invested.