Where Does AI Fail in the Factories of the Future

By Brad Jokubaitis

When AlphaGo defeated Lee Sedol in Go (an abstract strategy board game) in 2016 4 games to 1, it was seen as a milestone in AI.  AlphaGo created a strategy that was unlike any human players before it—something novel that neither researchers nor players could predict, performing billions of careful calculations per second. Champions of Industry 4.0 (the fourth generation of factories described below) look to this milestone as a benchmark for fully automated ‘smart factories’, where a single control system governs production.

But what if machines are being overestimated?

Garry Kasparov, the chess Grandmaster that lost to IBM’s Deep Blue supercomputer in chess, said of the machines that faced him, “the assumption that humanlike was better than brute force was largely wrong.” Despite machine victory in these games, there are elements to intelligence that are missing: intuition, creativity, flexibility. Even if machines are getting better at brute-forcing solutions, that gap will always be present, no matter how small.

As the manufacturing environment gets increasingly automated with machines that inherit from AlphaGo and Deep Blue, it is in situations that require intuition, creativity, and flexibility that humans will become increasingly valuable to solve the problems that machines are ill-equipped to address.



Board games to reality—how well does it translate?

What separates AlphaGo from a singular control computer capable of emulating human thought are the rules and constraints given to operating domains. AlphaGo was trained on sets of possible moves and rigid limitations that those moves had to be checked against. It was fed data from millions of Go games and played itself time and time again to develop its algorithms in preparation for its challenger.

However, AlphaGo was given a clear definition of winning. While its in-the-moment tactics may be difficult to comprehend, the principles guiding its decision making are clearly understood. This victory signals that machines have a grasp on low and mid level planning (tactics and some strategy), but the upper end--defining an objective--is still elusive.  But what if the definition of winning needed to change? All of the preparation and optimization the machine completed would be towards the wrong goal.


Interactions in real life are rarely so clean or predictable, and winning may not be clear. In a complex factory, where some problems may not show symptoms until it’s too late, a machine would be unable to identify and solve problems back to root causes unless it had been trained on a similar problem. Oftentimes solving the symptoms of a problem is not the same as solving the problem itself. Therefore, the machine needs to know what ‘winning’ means—a clear objective needs to be defined and understood. If it changes, what happens?

Take an example where certain screws in an assembly are failing because of a supplier quality issue. If AI detects that having looser assemblies prevents screw failure, has the AI ‘won’? If there were a persistent flaw that no one knew about until after machines were trained, how could machines diagnose and achieve the objective of flawless production?

The possibility of fully automated problem solvers

For a machine to identify, diagnose, and recommend corrective action it must have the ability to capture and process the necessary data. Herein lies a difficult tradeoff: a complete sensor suite monitoring all possible measurements at all times would be both prohibitively expensive and generally overwhelming. But, any simpler system may not be capable of discovering root causes to all potential problems, and will therefore not be fully automated.

A factory which is fully automated must be prepared for all circumstances, having the ability to detect and interpret

  1. Issues that have happened and were solved

  2. Issues that are happening and are currently unsolved

  3. Issues that have yet to happen

Having sensors, measurements, and data are necessary in the problem-solving process, but erroneous data or incomplete data is as bad as having none at all.

In one of Stroud’s recent projects, a combination of mass spectrometers, flow meters, X-ray diffractors, barometers, thermometers, concentration meters and cameras was necessary to determine the root cause of a years-old problem. The costs associated with constant operation, maintenance, calibration and redundancy of an all-encompassing sensor suite that can solve every possible problem makes permanent installation absurd.

Where will human beings play a role in factories of the future?

The march towards increased automation is inevitable and ongoing. Advances in software and robotics, combined with increased design and production challenges of future products, ensures a greater role for advanced machinery in manufacturing.

On the path towards automation, there are certain hurdles that must be overcome before value can be realized. To maximize this value, and to avoid an expense headache later on, it is necessary that these systems be set up correctly. This includes knowing where and what to measure to such that the system can monitor all parts of production. Once installed, if the integrated system is unable to identify a problem, someone must be able to find the missing link, bridge the knowledge gap, and ensure that the system is prepared for similar issues in the future. If the integrated system can identify the problem but not recommend a solution or locate its root cause, someone must fix the issue while training the machine on this failure mode for the future.

It is here that problem solvers will be most needed. The responsibilities of problem solvers are to:

  • Quickly understand the physical process at hand

  • Quickly understand the control architecture and measurable variables

  • Distill the extreme influx of data to its most relevant elements

  • Aid in the creation of production and monitoring systems

  • Aid in the identification and resolution of system shortcomings

  • Mediate conversations between software developers and manufacturers

Until a machine can know everything there is to know or until it has the ability to gather any piece of information it  needs, problem solvers will remain necessary pieces of industry. But as the industry enters this new 4.0 revolution, problem solvers—those that can understand and deliver—will become more helpful than ever.

Like this article? You may be interested in: