Ship Collision Avoidance Using Constrained Deep Reinforcement Learning

In recent years, the rapid development of mobile technology and application platforms has provided better services for life and work. Artificial intelligence and mobile technology have made traffic ever more convenient. As an artificial intelligence method that intersects with multiple disciplines and fields, reinforcement learning has been proved to be highly effective in the automatic driving of vehicles. However, there are still many difficulties in ship collision avoidance, because it involves continuous actions and complicated regulations. We find that by constraining the states, actions and regulation of reinforcement learning, we can well apply reinforcement learning to ship collision avoidance with vast states and actions at the same time. Hence, we propose Constrained-DQN(Deep Q Network), which is used to limit the state and action set, and separate reward value via different regulations. Experiments show that Constrained-DQN is more stable and adaptive in handling continuous space than traditional path planning algorithms.