A Sample-Efficiency Comparison Between Evolutionary Algorithms and Deep Reinforcement Learning for Path Planning in an Environmental Patrolling Mission

摘要

For water environmental monitoring tasks, the use of Autonomous Surface Vehicles has been a very common option to substitute human interaction and increase the efficiency and speed in the water quality measuring process. This task requires an optimization of the trajectories of the vehicle following a non-homogeneous interest coverage criterion which is a hard optimization problem. This issue is aggravated whenever the resolution of the water resource to be monitored scales up. Since two of the preferred approaches for path planning of autonomous vehicles in the literature are Evolutionary Algorithms and Reinforcement Learning, in this paper, We compare the performance of both techniques in a simulator of Ypacarai Lake in Asunción (Paraguay). The results show that the evolutionary approach is 50% more efficient for the lowest resolution but scales badly. Regarding the learning stability and sparsity of the trajectory optimality, the Double Deep Q-Learning algorithm has better convergence, but it appears to be less robust than the evolutionary approach. Finally, in a generalization analysis, the Deep Learning approach proves to be 35% more effective in reacting to scenario changes.