Day after day, year after year, people keep talking about Artificial Intelligence (AI) while in fact they are only referring to Machine Learning (ML). While this was already the case with Data Science (DS) in the previous years, and with Cognitive Computing even before. The trend is to use the latest buzz word, as if something new and bigger was happening, while in fact referring to older technologies, and even sometimes only a subset of these technologies. These simplifications are plainly wrong and are the same as claiming to speak about sports but really only discussing soccer.
This post introduces the different types of intentions in Business Analytics and the different technologies applicable for Prescriptive Analytics.
Artificial Intelligence, Data Science, Machine Learning and Decision Optimization
All of these concepts are different and somewhat overlapping. Novices, and sometimes careless experts, are using them as if they were interchangeable.
Artificial Intelligence is described in Wikipedia as:
And Data Science:
We might hence consider that AI and DS greatly overlap.
Both include lots of methods, techniques, technologies, and most of these can be considered part of both AI and DS.
Before digging into techniques, methods, and technologies, let’s introduce some classification of their use cases which will help. Just as natural intelligence, Artificial Intelligence may have 3 different intentions, well known as corresponding areas of Business Analytics.
Here are simple examples:
One intention is to describe the world. This intention, Descriptive Analytics, uses known data and reports to describe what is the current situation.
The next intention, Predictive Analytics, extracts additional information from this data. It can be a forecast (the weather for this afternoon, for next week) but it can also be trends or classifications for future outcomes.
Finally, the last intention from intelligence is to prescribe the next actions to execute. Prescriptive Analytics provides suggestions for data for which values are not known and cannot be extracted from other data. In fact, this data can take different values. These are our decisions and the freedom to make these is what makes life and business complex (and interesting!).
These three areas correspond to intentions. It would be a mistake to strongly assign one technology (e.g. Machine Learning ) to one area (e.g. Predictive Analytics).
In mainstream publications and in most people’s minds, Artificial Intelligence refers to machines acting independently of humans, and hence making decisions alone. For now, let’s focus only on Prescriptive Analytics and dig into the different existing methods and technologies.
Methods and Technologies.
I have been using the following example during the last few years to introduce different Prescriptive Analytics technologies.
Recently, I have been on holidays with my family and we all had to prepare our luggage. This is a simple and common decision making problem. What should I put in my luggage? How many socks? Should I bring one book or too? Should I take shirts or sweaters?
The different approaches used illustrates nicely the different types of Artificial Intelligence which can be used for Prescriptive Analytics.
In my personal case, these are easy decisions. I more or less randomly take always the same stuff and then I manage to adapt on site. This is what we call a heuristic, and this type of technique is still used by too many people (and some businesses) to make decisions. I remember the first time I went to Chicago in winter, with -11 degrees Celsius, the adaptation was not so easy.
We might agree that making decisions randomly cannot really be considered intelligence, even if we could agree that a huge part of human decision making is based on this, even for very important decisions. Human decisions are affected by quite a few other human specific factors, as the Wikipedia definitions reminds: limited memory, limited computation capacity, and, more important, emotions and environments.
Limited memory. Humans have limited memory, and sometimes, even faced to the exact same situation that before, they are able to take the exact same bad decision. My grandmother’s most repeated recommendation was: “making a mistake is not important, but doing the same mistake twice is really stupid”. Some similar commonly heard advice is that “either we succeed or we learn”.
Limited calculations. When calculations are required (e.g. total weights of items in the luggage) and when the problem size grows, large calculations are required. Here, humans are definitely limited.
Emotions. Humans have emotions. Emotions are probably the most impacting factor for human decision making. Studies have shown that the same person can make different decisions on the same problem on different days, depending on their emotional state on those different days.
Environment. Under pressure, when tired, with severe weather, human decisions will be impacted.
In the case of my younger kid, Adrián, he just followed the rules provided by his mother: “Mum told me to put one set of socks, shorts and shirt per day, then to put one book, and then some toys if I can make it fit”
Most of our education system is dedicated to teaching our children this ability to understand instructions in the form of rules and executing them and this level is largely enough for a minimal survival in human society. At home, at school, and at work, lots of human don’t need to use any other type of intelligence.
The quality of the decision is mostly impacted by the quality of the rules and the ability to execute them correctly.
This type of intelligence corresponds to Business Rules Management Systems.
My older kid, Hugo, has gone to the next level. He has learnt on his own based on looking at his mum packing luggage several dozens of times.
Indeed, in our complex world, you cannot explicitly provide all the rules on everything to your kids, so you expect them to learn by looking at you and looking after other models you tell them are good. You eat healthy food, do sports, and read books, expecting your kids to look at you and reproduce these good habits.
When humans can just learn on their own, this is another level of intelligence. Some humans (not all) use such kind of intelligence. But humans are capable of learning on their own, as many people have likely learned from having their children remotely educated for portions of this past year. A very limited type of learning is at least not to repeat the same mistake twice (see my grandmother’s advice above).
Hugo had to see his mother pack luggage dozen of times before he could be autonomous. The quality of his packing highly depends on whether he learned from someone or alone, and who he learned it from. If he would have learned from looking at me, then the outcome could have been quite worse.
Also this is interesting to note that as he has seen his mother pack many times for summer holidays (to the beach), so he would still be in trouble if we would have to pack luggage to go skiing.
This type of intelligence corresponds to Machine Learning.
The third type of intelligence is what my wife, Mónica, uses. She knows that the airline luggage weight limit is 23kg, she knows what items are mandatory or not, and take into account the weather at the destination to set preferences between optional items.
Based on these constraints and preferences, she mentally calculates that this item plus this item would not comply with the weight constraints, or that this set of items offers a better global satisfaction than this other set.
There are not so many decisions, constraints and objectives with these decisions so that she can’t optimize the decisions mentally.
Underlying assumptions for the choices in packing could be wrong, for example, either a bad prediction of weather at the destination, or a misunderstanding on airline regulations.
This is how Decision Optimization works.
In summary, to each of the two intentions (predictive and prescriptive) correspond many possible methods and technologies.
It is interesting to notice that Machine Learning can be used for predictive and also, in some cases, for prescriptive. May be this is what made it so famous, and what is creating so much confusion.
Real world and conclusions
In the news, and unfortunately also in many technical publications, we do see that AI is made equal to ML. Even if not correct, we might accept this simplification if all or almost all AI applications would be using ML. This is not the case, by far.
As shown above, for different type of AI intention, there are different methods and techniques. While millions of “decisions” are prescribed using ML (e.g. which next movie you should watch on your preferred VOD network), there are also millions of “decisions” prescribed using Decision Optimization every day (e.g. electricity production mix, planes and trains schedules, plant production plans, etc).
It is hard to measure, but I would claim that in terms of business value, and even in 2021, every day, Decision Optimization has more impact on the real world than ML. (Measuring this with data would be an interesting topic)
Each technique might have its own set of benefits and inconveniences. I will not dig into this here, and that will be the topic of another post focusing on relative pros and cons of ML and DO when doing Prescriptive Analytics.
Meanwhile, let’s be careful when writing or reading about AI Ethics, AI Fairness, AI Robustness, AI Explanability, etc.. Check whether this is about AI in general or only about ML. The well-known concerns about ML fairness and explanability do not apply to Decision Optimization. Let’s not, at the end, conclude that AI does not work as it promised, if, in fact, only one particular technique or technology is not working as it promised.
Intelligent humans know when to use each level of intelligence. Sometimes, deciding does not require to develop an optimization model (e.g. decide where to go out for dinner tonight), but sometimes it really helps (e.g. to decide how to invest your savings for retirement).
With Artificial Intelligence, a similar situations occurs, where different approaches, which still not always easily combine, may have different benefits and inconveniences. This is why this is critical to understand these different techniques, their benefits and limitations and be able to identify when each one would better apply and when they can combine efficiently. AI Platforms should also be created and used taking into accounts the benefits of the different available technologies.
Note: this post reuses some content from an older post : (Artificial) Intelligence, which one?