This study proposes the use of generative adversarial networks (GANs) to solve two crucial problems in the unmanned ship navigation: insufficient training data for neural networks and convergence of optimal actions under discrete conditions. To achieve smart collision avoidance of unmanned ships in various sea environments, first, this study proposes a collision avoidance decision model based on a deep reinforcement learning method. Then, it utilizes GANs to generate enough realistic image training sets to train the decision model. According to generative network learning, the conditional probability distribution of ship maneuvers is learnt (action units). Subsequently, the decision system can select a reasonable action to avoid the obstacles due to the discrete responses of the generated model to different actions and achieve the effect of intelligent collision avoidance. The experimental results showed that the generated target ship image set can be used as the training set of decision neural networks. Further, a theoretical reference to optimize the optimal convergence of discrete actions is provided.
Developing Generative Adversarial Nets to Extend Training Sets and Optimize Discrete Actions
2019
Article (Journal)
Electronic Resource
Unknown
maritime education and training (met) , generative adversarial network (gan) , discrete actions , met system in japan , lifeboat , monte carlo tree search (mcts) , learning methods , unmanned ship navigation , Canals and inland navigation. Waterways , TC601-791 , Transportation and communications , HE1-9990
Metadata by DOAJ is licensed under CC BY-SA 1.0
Transportation Research Record | 2020
|