Where is research into the creation of intelligent systems heading? ?
Since 1956, the term Artificial Intelligence has been echoing around the world. There are many definitions of this technology. Although they all have more or less the same meaning, it is in the approaches that they differ. There are two main areas of research, each with its own approach. The study of thought (reasoning), and the study of behaviour, which takes a human and rational approach.
Here are the four approaches that are driving forward research into Artificial Intelligence:
Should an artificial intelligence think like humans? act like humans? think rationally? or act rationally?
It's true that you might think it's the same thing. But it's not. Research evolves differently depending on the approach.
The human approach
Machines must act like humans
The first person to think about "how to create an intelligent machine" was Alan Turing. His theory was that if a machine could be intelligent, it had to act like humans. So he created the famous Turing test.
For a machine to be considered intelligent, it must pass this test with flying colours. Without going into the details of the test, it involves a (human) questioner and a (human or virtual) respondent. If, at the end of the test, the questioner is unable to tell whether the respondent was a human or a machine, then the test has been passed.
This test is very well known and is still relevant 60 years later. The full version of the test includes a video signal to the machine to test perceptual abilities. To pass the full test, the machine must have these 6 functions:
- Mastering natural language
- Knowledge representation (storage of what it knows or hears)
- Automated reasoning
- Learning (adaptation + detecting invariants and extrapolating them)
- Artificial vision
- A robotic body
We will see later that these 6 functions remain essential in another approach.
But researchers prefer to study the principles underlying intelligence rather than trying to imitate it. We didn't learn to fly by imitating birds, but by studying aerodynamics. Which brings us to the theory that machines must "think like humans".
Do machines have to think like humans?
Before trying to make a machine think like a human being, there's a question that needs to be ... how humans think ? A big question, isn't it? We don't really know. But what we do know for sure is that it's the brain that provides us with the spirit (at the time we thought it came from the heart or the rat).
Today there are three techniques for trying to determine how thought works:
- Introspection (getting to grips with your own thoughts)
- Psychology (observing an individual's actions)
- Brain imaging
Once we have a fairly clear vision of the mind, thanks in particular to all the research that has been done on the subject, this theory represents the brain as a computer programme. All we need to do is find the code in our brain and we can apply it to a computer.
To confirm this theory, the actions performed by a machine and those of a human were compared for the same problem. And when the results matched, we deduced that the machine was thinking like a human. Moreover, in the early days of AI, people confused "thinking like a human" with "acting like a human", because they thought that thinking justified action. Today we know that action can be equal with a different system of reasoning.
Here again, we are trying to imitate humans. But it is important to go deeper into the cognitive system in order to try and reproduce it. Humans are rational (with limited rationality, of course). So understanding the process of rationality seems to be a good entry point into the development of AI.
The rational approach
The machine must think rationally
This approach is known as the 'law of thought'. Aristotle was one of the first to recognise that certain things are always true. He therefore sought to codify "right thinking", i.e. irrefutable reasoning procedures. To quote his example: "Socrates is a man, all men are mortal, Socrates is therefore mortal". This system of thought is known as 'logic'.
And these laws of thought were supposed to govern the entire functioning of the human mind. With this in mind, the first computer programmes capable of solving logical problems were born. These systems are already considered "intelligent" because they can assist humans in tasks that require rational reasoning.
This approach has two drawbacks:
- It is very difficult to isolate informal knowledge so that it can be used formally, especially if the knowledge is not 100%
- There is a big difference between solving a problem "in theory" and "in practice". To put it more clearly, if the system is not guided in its choice of priorities, calculation capacity can quickly be exhausted in the search for the "best way".
And of course, we all agree that certain things are not always true at 100%. Or they may vary according to each person's 'free will'. That's why this approach is doomed to failure, and why rational thinking is not always the best way to operate.
This leads us to the final approach, which asks the machine to act rationally.
The machine must act rationally
This is the rational agent approach. To be rational, the agent must theoretically be capable of :
- Operate independently,
- Perceiving your environment,
- Persist for a prolonged period,
- Adapting to change,
- Pursuing objectives
The rational agent must therefore always act in the best way (in the case of perfect rationality), or, if the environment is uncertain, give the best foreseeable solution.
This approach gives rise to two cases:
- The agent logically concludes that the action he is about to take will enable him to achieve the objectives set, and then infers (to infer means to act with knowledge of the facts).
- If the agent cannot be sure at 100% that the action he is about to take will enable him to achieve his objectives, he must evaluate the best solution and act despite his uncertainty.
In human beings, there are situations in which we act rationally without being inferential. Survival reflexes, for example. If you put your hand on a hot stove, the reflex to quickly remove your hand is very rational because it prevents you from burning yourself. On the other hand, the inference is not verified because we act too quickly to assess the best solution to follow.
This is where we return to the Turing test. Possessing all the characteristics needed to pass the test would enable the agent to act rationally.
This approach has two advantages over the others:
- It is more general, giving the validity of inferences (and therefore of the logical system) a single process for achieving rationality (unlike the "law of thought"),
- The rational approach (and not the human approach) gives research a simpler scientific approach because it is more easily verifiable and therefore provable.