Many moons ago, I had a class about artificial intelligence. On one of our first meetings, the teacher asked the group which components are required by an intelligent system. "Long term memory", some said; "short term memory" - said someone else; "the ability to learn" - others added.

I raised my hand and said that they need a randomizer. Most people did not pay attention, or just looked at me in a weird way, I was not given an opportunity to explain why that would be necessary. So here we go.

A human life can be seen as a continuous process between decision-making points. Have you ever wondered why you make the choices you make?

Here are some examples:

• you walk on a path and you reach a fork - where do you go? To the left or to the right?
• you see two apples, and you're hungry - which one do you pick? The red one or the green one? The one on the left or the one on the right?

In these cases your decision can be based on (or influenced by) one or several of the following:

• personal preference - I like green apples
• current objectives - I have to get to the town of Byron, so I go to the East (which is the left path)
• available knowledge - red apples are sometimes poisonous, so I choose the "NOT red" apple
• circumstances - the green apple is the nearest one

Each of the above is something that can be quantified, so you can make the choice by simply comparing some values. An apple is 10 cm away from me, another one is 100 cm away; 10 is less than 100, so go for 10. The need to get to Byron is more important than the need to go elsewhere - so the choice is obvious, etc.

Now, let us examine another set of examples:

• Imagine that you walk on a hypothetical path, you don't have a specific destination, you don't have a map, you don't know anything about the area. You reach a fork, either way is equally unknown and equally attractive to you. Which one will you choose?
• Imagine that a table is in front of you, the table has two identical apples on it, equally distant from you, same size, same flavour, both are your favourite kind. Which one will you choose?

This is where things get tricky, because the previous methods (preference, circumstances, etc) won't work. You cannot rely on some objective metrics, since both options are designed in such a way that they are equally attractive, by all criteria.

You may argue that the perception of the options will be different. For example, the lighting conditions or my imperfect sight can make me see an apple as "slightly bigger" or "somewhat greener" or "somewhat closer" - thus making it easy to choose.

Note: in all the examples below I will refer to a green apple and to a red one, but you should imagine that they are of the same colour. I made the distinction because it is easier to write "red" or "green" than it is to write "the first green apple of two identical apples" and "the second green apple of two identical apples".

 

So, let us bypass the senses entirely and focus on the fact that this is an imaginary scenario, and that the apples are indeed identical, and that the two paths are indeed identical, just like we can imagine things like "ideal gas", "perfectly smooth sphere", "perfect vacuum", and of course - "perfectly spherical horses moving in a vacuum" :-) Take a look at the picture below - the distances are the same, everything is the same (believe me, I measured it with a perfect ruler).

 

How do you choose then?

A perfectly rational system will continue analyzing each option in order to choose the best one. Unfortunately, there is no "best" option here. Will such a system get stuck in an infinite loop? If so, it can starve before it decides which apple is better; or maybe the apples will rot by the time a verdict is reached.

You may argue that there is no point in analyzing hypothetical scenarios, because reality doesn't work like that. But I disagree. As human beings, we spend a lot of time thinking about the future, playing with "what if?" questions, i.e. analyzing things that are simply not there. Yes, our thinking about the future is influenced by our imperfect senses, by our limited and inaccurate knowledge - hence there are no "all options are 100% identical" cases to deal with. However, by exercising our analytical skills with such edge cases, we can get better at solving regular problems.

Possible solutions to the "stuck in an infinite loop" problem

Senses are imperfect, and that is a feature, not a bug. The design of the human organism is the way it is, in order to ensure that such situations do not occur. This is a pretty good explanation, because it is much easier to build an imperfect system, than it is to build a perfect one ;-)

***

The brain has a randomizer, it uses it to resolve such cases. An engineer will ask how that randomizer works. Random numbers generated by a computer are not random, they're the result of a deterministic process - given all the input data, you can predict the number that will be generated.

If the brain relies on a randomizer in its decision-making process, it would be great if we knew how that randomizer really works. As long as it doesn't rely on some external source of randomness, then I doubt it is truly random. And if it is not truly random, as a marketing person I could predict which of the two products will be chosen by a potential customer.

In other words, if we understand the principles that govern the behaviour of our brain's randomizer - we can exploit that knowledge for fun and profit.

As an exercise, think of a few random numbers right now.

***

Data structures and data analysis algorithms determine choices. Take a look at this image, this is what your hypothetical data analysis system could see when it looks at the tables with apples (see above) :-)

The system could work by dividing the image into a grid of squares, each square has an index (ex: A0, A1, B3, etc). Then the system will analyze the grid in a specific order, for example:

• starting from the top-left corner and going clockwise;
• starting from the top-left corner and top-to-bottom, then bottom-to-the-top, and so on, until it reaches the edge;
• starting from the top-right corner and going to the left, then jumping back to the right edge, and so on.

 

Sometimes these patterns can be quite complex, but they're still patterns that are predictable. Before you say "ah give me a break, humans are not robots, you can't simplify things like that!", let me tell you that the F-pattern (lower left corner) is used by humans when studying the contents of a web-page. This has been determined empirically.

Sometimes these patterns can vary. For example, a left-handed person may prefer to start scanning from the top-right corner. Statistics show that when entering an unknown area, a right-handed person that doesn't know where to go - goes to the right.

Or maybe the starting point of the scan is determined by the culture we grew in. It is possible, but the thing is that at the end of the day - a human is still an animal, so we can make a reasonable assumption - other animals do it in a similar fashion. We know that animals don't have culture, yet they still have to find and eat their apples, don't they? Therefore, I guess that the basic scanning patterns are an innate feature, that can be altered by education, but they are good enough at their default settings as well.

Let's think about what happens once the perceived image was analyzed. The found objects have to be memorized somehow, before we can play with them. The way they are stored can influence our choices. Let's consider two simple examples - a queue and a stack.

• Queue - the first object that was discovered is the first one to be processed (first in, first out). Picture this as a group of people buying tickets - whoever was there earlier got served first.
• Stack - the first object to be discovered is the last one to be processed (first in, last out; or last in, first out - same thing). Picture this as a set of books placed one on top of the other. The last book you put on top is the first one you will take out; if you try to pick a book from the bottom - the pile will collapse. Yes, you can do that, but you'll get dirty and the pig will like it :-)

If we use the "start at top-left, go right, then left, then right..." method, our senses will see the red apple first, and then the green one.

 

When it comes to making a decision about the apples, we'll process the green one first - if the discovered objects were stored in a stack; and we'll process the red ones if the data structure we use for storage is a queue.

Of course, there are many other types of data structures, much more complex, governed by more elaborate rules, but let us stick to these examples because they are easy to grasp.

It is clear that if we know what kind of data structures are used, we can predict the choice made by a person. When I say "predict", I do not necessarily mean a 100% correct prediction; any research that can make our "guess better than chance" is a step forward.

Where does this bring us in our quest? The conclusion is simple - if the brain cannot apply factors such as preference or available knowledge when facing a choice, it will begin to process the options in an order determined by the data structures and by the applied scanning pattern.

No matter how massively-parallel the brain's modus operandi is, no matter how convoluted the data structures are - there's still a starting point, there's still a way in which the memories are kept in the brain, etc.

Note: the ideas are heavily biased by my academic background. As someone for whom programming is a frequent activity - I tend to think in terms of such data structures and scanning patterns. When the only instrument you have is a hammer, every problem looks like a nail :-)

***

The somatic marker hypothesis is an alternative answer to this question; or perhaps one of the pieces of the puzzle. I am not yet fully familiar with it, so I will only tell you what I know so far.

• The hypothesis is devised by Antonio Damasio, you can read about it in his book - "Descartes' error".
• The idea is that when options are considered, you have a gut feeling that makes some options more attractive and others less attractive. "Soma" means "body" in Greek, so the idea is that the body (physically and emotionally) can prefer some options and "mark" them as preferred ones. This is how the name "somatic marker" came to be.
• People who lack the ability have some parts of their brains damaged (or removed), they are unable make good decisions, because their choice-making ability is impaired.
• Examples of such people are Phineas Gage - the guy who lost a part of his brain when a metallic rod pierced his skull and got out at the other side, taking a part of the brain with it; a patient named Eliot - a successful family man and businessman who lost everything in a short time after a part of his brain was removed (to prevent seizures). You can read more about this case from the book.

On one hand, the somatic marker hypothesis can be seen as a "high level feature", which gives preferences as a function of data structures and scanning patterns. For example, given two identical apples - my gut feeling will tell me to like one of them more than the other.

The gut feeling itself relies on a simple principle - what's the first object in the queue? Or what's the last object in the stack? That's the one you have to prefer.

The trouble is that this doesn't explain why people fail to make good decisions after a certain part of the brain is removed. These people still have vision, they still employ a certain scanning pattern and they do keep their scanned objects somewhere - but this is not sufficient to make their gut feeling make good guesses.

My conclusion is that the data structures and scanning patterns could be a "worst case scenario" mechanism the organism can fall back to, when the features that deal with this task better are not available. After all, people who have damaged brains still make decisions, therefore they do base their choice on something, don't they?

Another argument is that it is much easier to extend a system (including a biological one) by slightly modifying an existing feature, or by applying an existing feature for purposes other than those it was originally designed for. This means that relying on data structures and scanning patterns when facing a choice makes sense.

Besides that, biological evolution often favours the simple and easy to implement approach. The best example I can provide in this context is the tit-for-tat strategy (the biological applications of which are pointed out by Richard Dawkins; Robert Axelrod also describes how it works, in "Evolution of cooperation").

***

Although this article doesn't come to a definite conclusion, I hope it gave you some things to think of.

p.s. Remember when I asked you to think about random numbers? Well, my guess is that your randomly chosen number was an odd number, and most likely an "ugly number" - i.e. something that cannot be divided by 2 or by 5 easily :-) 17, 13 or 7.

Try to conduct this experiment with your friends, and write down the numbers they thought of. Do you notice any patterns? If so, then do you think those numbers were truly random?