Hoping to minimize human decision making in ship navigation, researchers at the University of Southern California’s (USC) Viterbi School of Engineering are developing an automated system that relies on artificial intelligence (AI) and data analytics to avoid collisions.

Because a significant number of marine accidents and casualties are due in part to human error, researchers Xiongqing "Vincent" Liu and Edwin Williams are developing an automated system to prevent collisions. Liu, a mechanical engineering student at the Viterbi School of Engineering, developed the AI part of the system, while Williams, an aerospace engineering Ph.D. student, developed the data analytics component.

Liu used a machine learning method called reinforcement learning to teach the system how to avoid collisions with ships and other obstacles through a series of boating simulations. Initially, the system knew nothing, which forced it to explore the simulated environment independently, according to Liu. If the system collided with obstacles, it was penalized. However, if the system managed to avoid the obstacles, it was rewarded. Following thousands of simulations, the system learned how to avoid particular outcomes and collisions much the same way that humans learn.

However, alone, the AI system is not error-proof, according to Liu, as there are knowledge gaps because the system relies on scenarios Liu fed the system. William’s data analytics model reportedly fills some of those voids, using 20 years of historical boating data on past ships’ decisions and outcomes to make predictions about what other ships might do. In other words, the data tells the system which path has the lowest probability of experiencing a collision, according to Williams, who explained that the amount and quality of available data will be the factors driving the system’s accuracy. For instance, with details such as what captain was driving the oncoming ship, the system will likely make more accurate predictions.

"You can imagine there's an infinite number of trajectories that the vessel could take. But each one of those infinite trajectories has a certain probability of being taken," Williams said. "What my system does is looks at the entire probability of what those trajectories are and then determines the minimum likelihood of where the other vessel is going to be at any given time."

Alone, the AI had a 100% success rate during test simulations, but Williams acknowledges that the AI is limited by the scope of the data it was fed. According to the researchers, combining AI with navigational analytics offers additional security.

The team plans on further developing the system.