Route selection for a three-dimensional elevator using deep reinforcement learning

Abstract

While the car of the conventional elevator system moves only vertically in one dimension (up and down), the car of the three-dimensional elevator system travels in three perpendicular dimensions. The elevator moves through a vertical shaft to a certain floor and then the elevator serves multiple passengers distributed among different rooms at that floor. The controller decides which route should be taken to serve the passengers. This article proposes the use of deep reinforcement learning to select a route for the three-dimensional elevator. Deep reinforcement learning method learns from experiencing a large number of scenarios generated using Monte Carlo simulation offline. Once trained, deep reinforcement learning can select the route online. Numerical experimentations are used to show the superiority of deep reinforcement learning in finding an optimum or near optimum-route instantaneously. Although deep reinforcement learning is closer to finding the optimum route than other methods, finding an optimum route is not always guaranteed. Deep reinforcement learning has some limitations that include the long training time and the difficulties in training the neural networks.

Practical application: Multidimensional elevators have been of expanding interest to the elevator industry as well as to traffic analysis engineers. This article demonstrates that deep reinforcement learning surpasses other methods in finding an optimum or near-optimum route for the three-dimensional elevator, and it also overcomes the challenges of the non-intelligent methods. This article can help enterprises that develop multidimensional elevators in overcoming the challenges of the controller in addition to boosting the feasibility of multidimensional elevators.