Autonomous underwater vehicles (AUVs) path planning based on Deep Reinforcement Learning

摘要

In the path planning problem for autonomous underwater vehicles, good results can be achieved using deep reinforcement learning models, but the actual effectiveness of the algorithm is extremely dependent on the construction of the neural network, which can require tedious manual tuning. We used AUTORL to automate the search for a neural network architecture and trained a population of intelligences to search for an algorithm that max the final reward. We applied this to the DQN algorithm and trained it in a built simulation environment, demonstrating that the evolved underwater robot path planning algorithm improves both in terms of final reward and convergence speed.