What is Artificial Intelligence?
This week we sat down with a new client, and the following question - not an uncommon one - came up: “Okay, so with regards to what you do - scheduling, route and fleet optimization - what is artificial intelligence, and how does it compare to machine learning, or other technologies?”
It’s a common question, and so in the next few posts, we’re going to explore the topics of artificial intelligence, machine learning, mathematical modeling, and how they relate to operations research and optimization - what we do.
We wrote a bit about this in another post, which you can check out here.
Definition of Artificial Intelligence
We’re going to talk about some specific applications of artificial intelligence (AI) in later posts, but for now we can use the following general definition:
- Artificial intelligence: the theory and development of computer systems able to perform tasks that normally require human intelligence, such as visual perception, speech recognition, decision-making, and translation between languages.
So, in other words: getting computers to do things humans can.
Artificial intelligence was founded as an academic discipline in 1956, and has since gone through many periods of excitement and disappointment as various technologies have or haven’t panned out. The areas typically included in the field of artificial intelligence have also changed somewhat, with technologies that are successfully developed usually becoming standard and less interesting to those in the field, one example being optical character recognition (OCR).
Most recently, machine learning, a sub-domain of artificial intelligence, has been generating much of the current hype, and we will talk specifically about machine learning in a follow-up post.
While we won't dig into the distinction between "hard-AI” and "soft-AI", let's just say that hard-AI is still in the realm of science fiction. Soft-AI, on the other hand, is usually very good at executing systematically repetitive tasks that require a significant amount of information. If a task is not repetitive, it's probably a poor candidate for AI application, though it’s not because it's repetitive that it can be automated with AI.
Why is Artificial Intelligence Important?
The hope for artificial intelligence has long been that we will be able to develop systems that supercede human ability. The promise for such technologies has widespread implications in all sorts of fields - driverless cars, better medical diagnosis and treatment - almost all industries from finance to entertainment have the potential to be transformed by AI.
At the moment, successful applications of AI are limited to fairly specific, narrow domains. Examples include voice recognition, image recognition, cybersecurity, data science and finance, and we’ll talk a bit more about how many of these systems work when we examine machine learning.
However, most people who have used these technologies - Siri, Alexa and Google Assistant being common examples - will readily admit they’re far away from natural, and have many limitations compared to humans. Often the domains in which they work well are limited (a series of test images, as opposed to a large real-world sample of images where lighting, color, etc. changes).
One might point out Google’s recent testing of AlphaGo Zero, their computer program to play the Chinese game ‘Go’ as another domain where AI has actually exceeded humans. The Google technology defeated the best human players and played in a way not previously seen but it’s useful to note that in this example, the technology used was actually a combination of machine learning (an AI technique) and mathematical modeling, which is something different that we will talk about in a future post.
How Does AI Apply to Scheduling, Routing, and Other Areas of Optimization?
Currently, the most valuable applications of AI are business problems that are repetitive, but still require a great deal of intelligence. Scheduling is one such example. Schedule automation systems overall have struggled to beat humans, but such tasks are heavily repetitive. Unlike a computer, the more repetitive a task is, the more likely a human is going to make a mistake due to fatigue. Integrating recent discoveries in AI into scheduling automation systems is now breaking that tradition, and beginning to exceed humans.
In a scheduling context, there are a number of ways which artificial intelligence could be applied, but they are generally limited, compared to hybrids of other technologies with an AI component.
The following are some examples and common limitations in applying AI in the logistics/transportation optimization context.
One could use AI to attempt to generate crew schedules by providing a large number of historical schedules, or try to predict charter flight demand by inputting a large amount of data on historical demand.
You could try to predict the impact of impending weather by analyzing the historical impact of various types and severities of weather.
All these scenarios have a few common limitations, and the first that we see in a real-world context is that often the level of data maturity (in other words, the quality of the data) is not useful.
For example, if you want to predict the impact of various types of weather, you need data on both the weather, and the unexpected impact that can be attributed to weather. Typically, it’s rare for us to see solid records of planned and operated crew schedules, vehicle routes, flights, etc., let alone good data on why things changed.
The quality of your predictions will depend on the quality and quantity of historical data, so if you want to get good results, you need good data.
Compared to some techniques, like mathematical modeling, applying artificial intelligence is also resource-intensive. Once you have a successful, working model, not so much, but before that happens, there is typically a lot of training and data-processing required.
One Final Example
One final example we will return to: a good application of artificial intelligence would be examining all of Alex Rodriguez’s career hits, and attempting to predict where he will hit the ball in the next game. What happens, however, if you don’t have that data?
In the next posts in this series, we're going to explore what machine learning is, what mathematical modelling is, and the differences between the two, particularly as they relate to scheduling and logistics/transportation optimization.
What questions would you like answered about artificial intelligence? Let us know in the comments!
Get a PDF version of this blog post here: