What is AI? Really?
AI is the hot “new” thing. Everyone wants it. It does magic. It solves problems. It replaces people. But what is AI, exactly? How does it work?
In 1994, when computers were still relatively primitive, I did my Master’s thesis in Artificial Intelligence. You might be surprised to learn that today, 25 years later, very little in the way of AI fundamentals has actually changed. Sensor technology has improved greatly, and processing speed is dramatically faster, of course, but AI mathematics remain unchanged.
A good AI engine relies on three fundamental precepts:
It must be reactive, able to respond to inputs it receives and analyze them properly.
It must be adaptive, able to change how it approaches problems and solutions, and change them according to new data.
It must be predictive, able to foresee what is likely to happen next
I would add a fourth precept, one that was not available 25 years ago, but is far more relevant today: It must be perceptive. Perceptive ability is a uniquely human trait, not readily associated with computers, but sensor technology enables computers to receive inputs that were impossible a few years ago, and those inputs give today’s AI a perceptive skill.
It is the perceptive ability that has allowed AI to become what it is today. My thesis was based on the premise of an elevator control system. The goal was to reduce aggregate waiting time. I assumed a 12-story building with 4 elevators, single-use occupancy by one company. The factor we are measuring is time, something we all want more of. Back in 1994, the inputs and sensors were limited to a few parameters: calls at the different floors, the floor buttons pressed inside the elevators, and approximately how many people were in the elevator (all elevators have weight sensors built into them, so one could estimate the number of people based on the weight the sensor was reading).
The AI engine was designed as follows: raw data was input into a fuzzy logic shell, basically a series of IF-THEN statements that set the parameters for a linear program. A more advanced version of the IF-THEN statement is to use a table to set parameters by increments. A linear program will derive the maximum or minimum of a complex function. The output would then tell the elevator which floor to go to. The data was updated several times per second, so the elevator was able to map out its projected route over the next few calls. The data was then fed into a statistical model, which would analyze things like where the most calls came from, how long people were waiting, etc. It would also analyze which times of days were busiest. When there was downtime, the elevators would park themselves on the floors that were most likely to generate the next calls, thus predictively reducing the wait time for some people to zero.
The math behind this is somewhat complex; linear modeling is a form of quantitative analysis, and is used in logistics, and has been around for over 70 years. Classic problems are the Postman problem and the Snowplow problem: what is the shortest route a postman can take to deliver the mail, or what is the shortest route a snowplow can take to clear the streets after a snowfall. These kinds of problems have very real (and large) dollar figures attached to them. However, the basic principle is fairly straightforward: what do I need to do, how do I do it most efficiently, what have I learned to make the next time even more efficient?
AI today works about the same way, only the applications are far more diverse because sensors today are far more advanced. For example, in the elevator example, a camera mounted over each elevator door could use facial recognition technology to determine exactly how many people were waiting, if one of them was an executive, or if any of them showed signs of anger. Those inputs could then be used to weight the model to get to that floor faster.
In contact and call centers today, an AI engine can route calls, emails and texts to the appropriate agent. But today’s systems can read the messages and route them more accurately based on its content and tone. Is the sender angry? Is there anything indicating that he’s been experiencing poor customer service? Those kinds of messages can be routed straight to a supervisor, again, minimizing waiting time (and frustration). Voices can be analyzed for stress or mood, also helping to route the call more efficiently.
"Any sufficiently advanced technology is indistinguishable from magic." - Arthur C. Clarke
AI isn’t really magic. It can’t solve every problem. It doesn’t just “know” stuff. Good AI requires the engine to be tuned to a specific problem set with parameters that can be defined and measured. AI can be a very beneficial tool to reduce time and the results can be spectacular, but know what results you are looking for before you just plug in an AI engine and hope for the best.