A distributed learning control system for elevator groups

Human-designed elevator control policies usually perform sufficiently well, but are costly to obtain and do not easily adapt to changing traffic patterns. This paper describes an adaptive distributed elevator control system based on reinforcement learning. Whereas inspired by prior work, the design of the system is novel, developed with the intention to avoid any unrealistic assumptions that would limit its practical usefulness. Encouraging experimental results are presented with a realistic simulator of an elevator group.