A knowledge-assisted reinforcement learning optimization for road network design problems under uncertainty

Abstract

A knowledge-assisted reinforcement learning evolution optimization (KARLEO) is presented for a road network under uncertain demand and capacity. In order to hedge against stochastic link capacity and travel demand, a knowledge-assisted reinforcement learning optimization is proposed. Different from earlier studies, a stochastic link traffic model is first presented to capture time-varying cost incurred by traffic flow when link capacity is uncertain. In order to manage interaction between stochastic environment and traffic dynamics, a knowledge-assisted learnable environment for time-varying traffic flow and stochastic travel demand is proposed. To this end, a knowledge-assisted cooperative coevolution optimization is presented to solve a reinforcement learning-guided network design problem. In order to demonstrate computational performance of proposed model and approach, numerical experiments are performed at a real-world city under various kinds of traffic conditions. To investigate scalability of proposed approach, computational comparisons are made with a stochastic bi-level programming model at large-scale traffic grids under varying level of stochasticity. As it reported, the proposed approach achieved considerable improvement over conventional approaches by effectively reducing total cost at fairly low computational expense in all cases.