A New Game in Business
How Deep Reinforcement Learning can help you to make better business decisions.
One of the great promises of digital transformation is to elicit new insights and action instructions from our data with the help of algorithms. This should then lead to better processes, products and new business models. But as is so often the case, this is easier said than done.
One promising approach comes from the field of Deep Reinforcement Learning.
Deep reinforcement learning combines deep neural networks, which excel at recognizing patterns in large amounts of data, with reinforcement learning, in which learning is tied to a reward signal, such as winning a game, as in Go, or achieving a high score, as in the game Super Mario Bros.
In March 2016, AlphaGo, a computer program based on these principles, won 4–1 against 18-time Go champion Lee Sedol. Since then, a lot has happened in the field and better and better programs have been developed and released.
The latest version of these programs is called MuZero. MuZero surpassed AlphaZero’s performance in the games of Chess and Shogi (a Japanese variant of Chess), improved its performance in Go (setting a new world record), and surpassed the state of the art in the Arcade Learning Environment Challenge. This is about how well MuZero masters a selection of 57 Atari games. It dominated them all.
It has been proven many times that programs of this type can achieve superhuman results on the most complex games. But can we use these programs to solve business problems?
63 Millionen Dollars
This was also the question posed by a joint research team from Exelon and world-renowned MIT. Exelon, based in Chicago, owns and operates 21 nuclear reactors in the United States. The question to be addressed was, “Can we use Deep Neural Networks and Deep Reinforcement Learning to help improve nuclear reactor design and save money?”
One of the best ways to save costs in this type of power generation is deep in the reactor core. When fuel rods are ideally placed, they burn less fuel and require less maintenance.
Nuclear engineers have learned to design better and better layouts through decades of development, but now artificial intelligence should give optimization a new boost.
By transforming the design process into a game, researchers could run the equivalent of 36,000 simulations to find the optimal configurations. As a result, the fuel rods’ life in an assembly could be extended by about 5 percent, which translates to savings of $3 million per year when applied to the entire reactor. For the 21 reactors Exelon operates, that’s $63 million per year.
The World Needs Translators
This example clarifies that new technologies usually require a translation or transformation to make them usable for us. In the case described, this is the ability to formulate our question as a game. Without this translation work, we cannot expect to take advantage of new technologies.
Individuals and organizations that manage to incorporate these new requirements into their business model will most likely not complain about requests in the years to come. It could be a very lucrative business idea to develop a platform that allows its customers to independently and quickly transfer their business processes with this approach.
Referenz: “Physics-informed reinforcement learning optimization of nuclear assembly design” von Majdi I. Radaideh, Isaac Wolverton, Joshua Joseph, James J. Tusar, Uuganbayar Otgonbaatar, Nicholas Roy, Benoit Forget und Koroush Shirvan, 5. Dezember 2020, Nuclear Engineering and Design.