Part 2 of 2 Parts (Please read Part 1 first)
     Researchers looking for a better approach to optimizing the layout of nuclear fuel rods decided to explore deep reinforcement learning (DRL). This is an AI technique that has made it possible for AI systems to beat human beings at very complex games such as chess and Go. DRL combines deep neural networks with reinforcement learning. Deep neural networks are excellent at picking out patterns in huge amounts of data. Reinforcement learning ties learning to reward signals like winning a game or reaching a high score in a game.
     In the study reported in this post, the researchers trained their AI system to position the fuel rods under a set of constraints. The system earns more points with each favorable choice. Each of the constraints or rules chosen by the researchers is based on decades of expert knowledge that is supported by the laws of physics. The AI system might score points by moving the low uranium fuel rods to the edges of the fuel assembly to slow down the fission reactions there. Spreading out the gadolinium rods that suppress fission will gain points by maintaining consistent burn levels. It has been found that limiting the number of gadolinium rods to between sixteen and eighteen is a good choice.
      Majdi Radaideh is a postdoc in Shirvan’s lab and the lead author of the study. He said, “After you wire in rules, the neural networks start to take very good actions. They’re not wasting time on random processes. It was fun to watch them learn to play the game like a human would.”
      Although reinforcement learning has allowed AI systems have learned how to beat humans at increasingly complex games, the capabilities of such AI systems to do productive work in the real world has been largely unexplored. This study shows that reinforcement learning has potentially powerful commercial applications.
      Joshua Joseph is a research scientist at the MIT Quest for Intelligence and a coauthor of the report detailing the study. He said, “This study is an exciting example of transferring an AI technique for playing board games and video games to helping us solve practical problems in the world.”
      Exelon is currently testing a beta version of the AI system in a virtual environment that simulates the fuel assembly in a boiling water reactor. The researchers are also simulating two hundred fuel rods in a pressurized boiling water reactor. Pressurized boiling water reactors are the most common type of commercial nuclear power reactors in operation today. Exelon is based in Chicago, Illinois. It owns and operates twenty-one nuclear reactors across the U.S. A company spokesperson says that it could be ready to implement the new AI system in the next year or two.