A Deep Reinforcement Learning Approach for Non-homogeneous Patrolling using Wi-Fi Fleet-restricted Autonomous Vehicles

Abstract

The use of intelligent autonomous vehicles to monitor natural phenomena involves the optimization of multiple policies that must comply with physical restrictions of the environment. In the patrolling problem, typically addressed in the environmental surveillance of natural scenarios, it is required to fulfill the non-homogeneous coverage of an unknown scalar map, with limitations of navigable areas and communication. This work presents a framework based on deep reinforcement learning to deal with communication restrictions for online route planning and patrolling with multiple vehicles. This algorithm, based on the Deep Q-Learning algorithm, using a customized reward function and a fleet-informed deep network, is able to optimize every vehicle policy to maintain each vehicle’s distance from another within the limits of its wireless communication protocol (WiFi). The results show better performance than other path planning heuristics, while being a model-free approach and providing an effective method to use in similar patrolling scenarios.