Intelligence that’s “anything but artificial” boast the ads for luxury, battery-powered cars. AI engines may be a boon to consumer marketers, who are getting better at mapping the robotic habits and appetites of their organic end users. But should we be tempering our expectations at all as we contemplate how such algorithms are applied to industrial processes?
A decade ago, our chief console operator, Chaz, witnessed the newly deployed model predictive control (MPC) application making moves in anticipation of changing measured variables and measured disturbances. He liked to call it "Spock’s Brain," which fans of Star Trek TOS (“the original series,” for the uninitiated) will recall featured a fetching alien intruder, who removes the science officer’s brain to replace a worn-out brain controlling a subterranean HVAC. The newly lobotomized Spock accompanies the usual crew, who eventually discover his new role, ensuring his captors could remain cozy and comfortable in their 1960s go-go boots.
The creator and writers of science fiction have imagined AI like the conversant ship’s computer on the Enterprise and Knight Rider’s KITT for decades. We’re experiencing it firsthand now as we interact with Siri, Hey Google and Alexa, seeking navigation guidance, movie trivia, sports updates and even tuning suggestions for pH loops.
Just as the MPC controls were just doing math, there’s no personality churning through the algorithms to answer questions on behalf of the Google or IOS user. These services are astonishing in their abilities to index and retrieve seemingly the sum of all human knowledge in seconds. However, they're balanced by the times they get it entirely wrong—no, I don’t want to call the mayor of Tunis, I want to know if Meijer supermarket has tuna.
So, when a Bayesian optimization engine recently prompted the operators to drive the unit into perilous territory, Chaz likened it to another Star Trek TOS episode, “The Ultimate Computer,” in which the brilliant scientist Dr. Richard Daystrom’s M-5 computer was determined to wipe out all the humans in pursuit of self-preservation. The moral was that even the most thoughtfully crafted logic could lead to unintended consequences. It’s incumbent on AI creators to imagine every if-then-else that might befall their programming. The engineers of self-driving cars, for example, must consider the torments their machine will encounter from humans determined to befuddle it. Star Trek enthusiasts will recall how Captain Kirk foiled the M-5 with its own logic. How prescient were the sci-fi writers and producers of the mid-20th century?
To boldly go
Data-lake mariners are also boldly going where no one has before—and have justifiable enthusiasm for the prospect of finding a new world of golden treasures in the murky depths. But, like the artificial lake made by the U.S. Army Corps of Engineers in the next county, no one bothered to clear all the stumps. Even though the lake may be vast and polluted by trash, one can hypothesize with some confidence that there are valuable insights to be found. Many providers promise their algorithms can dredge through the deposits of sediment, revealing correlations and forecasting trajectories.
Whatever you call your algorithm, George Box’s maxim still applies: "All models are wrong, but some are useful." They exist almost exclusively in cyberspace, while their real-world counterparts are often haunted by all manner of curious, unmeasured disturbances and dynamics. Twinning something as simple as a waste heat incinerator and boiler can still run across relationships that defy “useful” modeling. It’s not uncommon for carbon monoxide in the stack to react faster than the temperature sensors. Just try correlating that.
The process industries are no strangers to predictive controls and optimizers. Contemporary AI engines and methods have new promise as we channel more pervasive sensing and other flotsam into the data lake. But let’s keep in mind that robots excel at being robotic, and fantasies like Knight Rider’s KITT or the Enterprise’s guidance computer require organic intelligence to ensure their utility.
