Here’s a nightmare story for you: Machines, endowed with artificial intelligence, get smarter than their creators, take charge and attempt to save humans from themselves. Oops. Smarts, it turns out, are different from wisdom. Pooh-poohing a saying variously attributed to Socrates, Aristotle as well Albert Einstein — the more I know, the more I realize I know nothing — the machines push ahead and bring civilization crashing down on all of us.
This sci-fi trope could have a real-world parallel. When engineers give AI more leeway in controlling complex systems, how do they keep the machines from overstepping their bounds and putting property, and perhaps lives, in danger? For instance, what’s to keep an AI in charge of optimizing the efficiency of a generator from pushing it past the breaking point for the sake of an extra kilowatt-hour of performance?
This lack-of-wisdom problem weighs heavily on Colin Parris, head of software at GE’s Research Center. There are clear benefits in an AI algorithm that, say, continually adjusts the tilt of the blades and the speed of a wind turbine’s rotation to reach an optimal level of performance based on a simulation of what happened in the past. The turbine generates more power with less wear and tear on the equipment, which translates to lower costs.
But there’s also some risk involved. The ability of the algorithm to achieve these results is closely linked with the accuracy of the machine learning algorithm it relies on. “When you build a model, you make an implicit assumption that the data you use when you execute the model will be the same as the data when you built the model,” Parris says. “For example, if I build a model in the summer, it may not be accurate in the winter.”
The real world can also stir up a “black swan,” an event of low probability with possibly catastrophic consequences. What happens when, say, a weird weather pattern produces a once-in-a-hundred-year gust of wind that was not captured in the data originally fed into the learning algorithm? The uncertainty over what a computer running a wind farm is going to do in those rare situations poses a business risk.
The solution, says Parris, may be to program humility into AI. That doesn’t mean endowing the algorithm with feelings — though some researchers are looking for ways to code morals. It means giving the AI an awareness of the limitations of the simulation of the real world it relies on, and an alternative way of proceeding that removes any uncertainty from its behavior. That plan B is typically an old “deterministic” algorithm that sacrifices peak performance for extra caution and predictability. “[Humble] AI knows its region of competency,” Parris says. “It can be defined as a region of temperatures and pressures. When it goes outside that region, right away it says, I’m going to revert back to an old algorithm, which has been robust for many years.”
Parris and his colleagues have undertaken two pilot programs to test this concept. One is a wind farm with two turbines. An AI there is forecasting the wind speed, adjusting the pitch of the blades and other factors to catch as much wind as possible. The other pilot program involves electricity-generating gas turbines, where the algorithm controls the nozzles that mix air and gas and regulate the pressures in the combustion chamber. In both situations, GE engineers have seen improvements in energy output of 1 or 2 percent.
Parris expects this new “humble AI” to be incorporated into commercial products by the end of 2019 or early 2020. Going forward, whenever the algorithms encounter conditions beyond their realm of competency, they will ask for new data so that they can expand the situations they’re able to handle.
The humble-AI concept is important, Parris says, to encourage acceptance of this powerful new tool. The best AI in the world is not going to work if it makes people nervous. Says Parris: “Humble AI allows people to embrace AI according to rules they feel comfortable with.”