Q-Learning traffic-distributing approach to managing multiple-destination traffic

Abstract

In disastrous situations (e.g., Tsunami or Flash flood), people must be evacuated to safety (e.g., shelter or high ground) as fast as possible, which depends directly on the number of available evacuation routes and the traffic flow along those routes. This study proposes a Q-Learning traffic-distributing approach to managing multiple-destination traffic. The approach processes current traffic on every available route and provides possible routes for people to take to evacuate to safety. This approach uses a reinforcement-learning algorithm to dynamically find and provide the best possible routes for people evacuating from different locations and times. The algorithm was developed from existing machine learning algorithms especially for processing traffic data. Evaluation and comparison of the developed algorithm, basic A*, and Yen's k-shortest path algorithms were conducted in terms of average travel time for various traffic and route situations. The developed algorithm provided the shortest average travel time for three tested situations. The evaluation test results may directly benefit developers and planners of evacuation programs in their effort to devise the best possible evacuation plans.