![]()
Chapter 1
Presenting the Actors
At the beginning of a book, it is natural to present its actors. Of course, the first actor is the book itself. Then I will examine the qualities and drawbacks of two important families of actors: human beings and artificial beings. I do not forget the computer, an essential support for an artificial being. Next, I will present my reasons for developing my present research, and introduce the subject of this research: my colleague CAIA, which is an artificial scientist. Studying its behavior and its mechanisms will supply most of the examples of this book. Finally, I look at the domains where CAIA is carrying out research.
1.1. The book
The goal of this book is to examine whether an artificial being can have some capacities similar to those that consciousness or conscience give to human beings. Over many centuries, a huge amount of work has been done on human consciousness and conscience, but sometimes I feel that I still do not understand them. In relation with my work, to understand a hypothesis on the working of our brain, I need to see how the hypothesis can be implemented in a computer program. Naturally, scientists who lived before the computer era could not express themselves in such a precise way because they did not know of computers and how we can program them. This is a difficult and unnatural way of thinking, and even now, we have to realize several Artificial Intelligence systems before we can easily feed computer programs with attractive ideas. When I read these early books, I can see a dozen ways to implement each of their ideas and so I feel confused: among all these possibilities, which one was seen by the author? For instance, I have never understood how the qualia could be exactly represented. The qualia are at the core of many theories of consciousness, but philosophers agree neither on their meaning, nor on their properties. We can roughly say that their goal is to represent in our brain sensations like redness, but when the authors want to be more precise, each one has his own definition. How could we use this concept in a computer program?
Over several years, some outstanding books and papers, such as those of Marvin Minsky, Daniel Dennett, Gérard Sabah or François Anceau, have been written by scientists who know exactly how a computer works, and the reader can understand how their ideas might be implemented. However, these ideas have not always been effectively implemented. An AI scientist knows that it is impossible to foresee in a program all the elements that would be necessary to give excellent results: either some mechanisms are forgotten, or their description is not sufficiently accurate. We have to experiment with the system, and modify it to work better. Thus, a description without an implementation is an indispensable starting point, but it is not sufficient, although it contains many interesting and promising ideas. Moreover, the goal of most of the research has been to improve our understanding of consciousness and of conscience for human beings. Usually, they were not concerned to see if these faculties could be expressed by artificial beings, which have different ways to process information. However, there are some exceptions; in particular, John McCarthy has written a theoretical paper on the possibility of making robots conscious of their mental states. My book complements his work and I do not consider the theoretical point of view, but rather its realization by a practical computer system. His paper must therefore be read by all those who are interested in this approach. Several papers have been published in the last years on the realization of various aspects of consciousness. They contribute new and interesting ideas, but this is not enough: intelligence is a whole, and we cannot examine only how some of its aspects can be implemented, we have also to study what kind of consciousness has been given to artificial beings which have to perform difficult tasks.
Examining existing artificial beings
Another approach is to examine an existing artificial being, and consider whether some of its possibilities could be related to consciousness and conscience as they exist for human beings. Since, in a program, everything has to be defined, we can find which mechanisms generate a behavior similar to our own behavior. Besides, when an artificial being uses these mechanisms, it may have possibilities out of reach for us, whose performance depends on the structure of our brain and the characteristics of its basic element, the neuron. Then, we are no longer interested in the understanding of human cognition, but in the study of an artificial cognition. Its main goal is the realization and the understanding of artificial beings, for them to become as efficient as possible. In that situation, we have a huge advantage over the psychologists: we can examine all the programs that make up an artificial being, thus we can exactly know the reasons for its actions and for its limits. However, there is a practical problem because we may have to examine programs with several hundreds of thousands instructions: we may misinterpret these programs if we are not their author. Papers in journals may give a sketchy description, but thousands of pages would be necessary for an accurate description. Thus, a paper, and even a book, can only give a general idea of the methods used by an artificial being; the details cannot be included, although they are often essential to understanding its performance.
So, the only person who can accurately describe the properties of a system is its author. Several artificial beings have already used interesting mechanisms which give them some capacities related to consciousness or to conscience, one of the most impressive is Lenatâs EURISKO. Unfortunately, I cannot describe them as much as I would because I cannot examine their programs in detail, I cannot know how some points have been dealt with, I cannot make experiments in order to evaluate their possibilities accurately. For this reason, I will mainly take my examples from the CAIA (Chercheur Artificiel en Intelligence Artificielle: an Artificial Artificial Intelligence Scientist) system, which I have experimented with for more than 20 years. CAIA has several attributes in common with human scientists, who discover new methods for solving problems, and perform many experiments to do so. However, CAIA does not have to deal with other important roles of current human scientists: CAIA does not publish scientific papers yet, CAIA does not search for scientific grants or funds, CAIA does not interact with the scientific community or policy makers, CAIA does not manage or advise junior human researchers.
CAIA is a step toward the realization of an artificial AI scientist. For the present, its main research domain is solving problems defined by a set of constraints, which must be satisfied by the solution. This family of problems includes many applications, and we often have to solve such problems. For instance, when we are choosing a meal for guests, we have to take into account a lot of constraints: their likes, the contents of the fridge, our cooking tools, our budget, the food supply at the nearest supermarket, etc. Crosswords and Sudoku problems are also in this category. My goal was not to develop an artificial consciousness or conscience; it was to realize a system able to learn to solve problems without the need to imitate the human behavior. Once CAIA was successful, I looked for and analyzed the mechanisms related to consciousness and conscience. When they were present, they were necessary for the success of the artificial scientist.
Plan of the book
This book comprises ten chapters. In Chapter 2, we describe some possible meanings for consciousness and conscience. Then in Chapter 3 we show that the concept of an individual is different for a human being and an artificial being. Chapters 4 to 6 examine what the consciousness of an artificial being could be: we describe several ways it can observe itself, why it is useful to observe its own behavior and how that can be implemented. Chapters 7 and 8 show that an artificial being can and must have a conscience. In Chapter 9, we examine some problems related to the conscience such as the importance of emotions, the difficulty in modifying its own conscience, and also the consequences of the existence of artificial beings on the human conscience. Chapter 10 discusses what the future has in store for CAIA. Appendix 1 describes the methods used for solving the problems given to CAIA while Appendix 2 gives more details of the implementation of CAIAâs consciousness.
1.2. Human and artificial beings
Two of the most amazing features of human behavior are consciousness and conscience, which philosophers and psychologists have tried to explain for many centuries. How is it possible to observe a part of the processes that occur in our brain? How can we acquire and use the knowledge of what is right and wrong? These problems are difficult because we do not have access to the processes which are executed in our brain: if we could observe our mechanisms, it would be easy to answer these questions. Besides, no one can observe what happens in the brain of another person. Finally, although some experiments might help to understand what happens in the brain of a person if some of its parts were destroyed, they are naturally forbidden for ethical reasons. Therefore, studies on the working of the brain are very limited, and this helps explain why there are so many theories, which are often contradictory. However, everybody agrees that these processes are essential for satisfactory behavior. Socratesâ motto was âKnow Thyselfâ, which can be realized because we are conscious. Furthermore, societies can survive only because the conscience of their members incites them to perform useful acts, and forbids them from committing actions dangerous for the future of the society.
Realizing artificial beings
We, AI scientists, realize computer systems, which are artificial beings. For the same reasons that consciousness and conscience are useful for human beings, we are led to give them the possibility to behave as if they were conscious and as if they had a conscience. The goal of some scientists is to model how the human brain works so that they will have a better understanding of human behavior. Indeed, if a mechanism generates a behavior similar to ours, it is possible that we are also using it. On the other hand, if the two behaviors are completely different, our brain certainly does not use the same methods as those implemented in the model. This approach is very interesting, but the goal of other scientists, and I am one of them, is not to understand how we work, but to realize systems that are as efficient as possible: we no longer claim to model the human brain. Naturally, a useful heuristic is to start from a hypothesis on how our brain works, but we are not obliged to follow it if we find better ways for performing the tasks given to our system. Naturally, once a system has been completed, it is interesting to analyze its methods, and perform some experiments on human beings to find whether or not we are using these same methods. It is also important to analyze its results for building an artificial cognition, which includes the description of new ways for implementing consciousness and conscience.
Constraints coming from the use of neurons
We must not blindly imitate human behavior because computers are tools which work differently from our brain. An obvious advantage is their speed, the number of operations executed in one second may be approximately the same as the number of operations that a human will execute in its whole life; besides, the computer will not make the thousands of mistakes made by a human. This is true when we are working in a serial mode, that is we cannot make several operations simultaneously, as is the case in arithmetic where we cannot make two divisions at the same time. However, the brain can also work in parallel, where it processes much data at the same time, for instance when we are perceiving a picture; for such application, the advantage of computers is not so large. Yet, the slowness of the neurons is not the only drawback coming from their use: we can neither observe their state, nor create new neurons. Another drawback is that it is not easy to store specific information such that it could be correctly used immediately, we have to use it several times before it is operational. I know that cars drive on the left in England, but I must be careful when I am in Great Britain because this information is not available each time that I cross a street. After a few days, this becomes automatic; unfortunately, when I return to the continent, it will be almost as long to acquire new reflexes. For a computer program, it is sufficient to swap ârightâ and âleftâ in a program, and everything is adapted in the new environment immediately; it is also easy to return to the previous behavior by doing the inverse swap.
Constraints coming from the structure of our brain
However, some characteristics of human intelligence do not come from using neurons, but from the structure of our brain. For instance, our working memory can hold around seven elements, which is a serious restriction for many tasks. We are not aware of the consequences of the small size of this memory because every human has this handicap; we avoid putting other people in situations where a larger working memory would be necessary. For instance, we speak in such a way that our interlocutor can process our sentences. We could possibly generate sentences such as:
The mouse that the cat that the dog that the woman that the soldier loves pets hates devours.
However, our interlocutor cannot understand it, while he can easily understand the following equivalent sentence:
The soldier loves the woman who pets the dog which hates the cat which devours the mouse.
We cannot stack in our working memory enough items so that we can wait for the many verbs that are at the end of the first sentence. Our brain evolved so that our ancestors, hunters and gatherers, could survive, breed and raise their children. The natural selection has created a wonderful tool for all of these activities, where a large working memory is not essential. Therefore, evolution has not developed this kind of memory, and it is the same for many other activities such as proving mathematical theorems, or managing large organizations. Our intellectual powers were useful for our ancestorsâ activities, and it happens that some of these powers are also very useful for other activities. However, there is no reason why they would be optimal for these new applications: the structure of our brain has not been optimized for them. Consciousness and conscience are among those useful aptitudes which evolution has developed for our ancestors, but it is likely that it would be possible to improve them so that we would be more efficient for most of our new activities. In fact, with the restrictions due to the neuron and those due to the structure of our brain, we are handicapped in comparison with artificial beings which are faster and which can simulate a huge variety of mechanisms.
Therefore, when we want to realize an efficient artificial being, we do not restrict its performance by imitating our way of thinking, thus enforcing constraints that are not essential. On the contrary, we have to take advantage of their extraordinary possibilities so that we obtain the best results. The goal of this book is to show that several features of consciousness and of conscience may be improved in that way. However, some of the human capacities have not been given to CAIA because they did not seem useful for the tasks which it is doing at present. Naturally, for future developments, it will certainly be necessary to give CAIA some of these missing capacities.
1.3. The computer
The computer is the extraordinary tool that enables an AI scientist to implement and experiment with his ideas. The methods which will be given to an artificial being are first defined as programs and data, then they are given to a computer which will execute these programs and give the desired results. We do not need exceptional computers for performing most AI research. I use an ordinary PC, such as those which are sold in their millions each year. It is not ultra-fast (1.8 GHz), and it only has 512 MB of memory.
However, a computer is almost useless on its own, we need a program to manage its working, which is the operating system. These systems did not always exist: the first computers were used without an operating system, but the present computers are so complex that it is almost impossible to use them without an operating system. It manages the computer resources, it decides which programs will be executed (a program being executed is called a process), it allocates them memory chunks, it deals with the inputs made through the mouse and the keyboard, the outputs such as those on the printer or the monitor, the network connections with other computers, and so on. It also checks that several parameters necessary for a satisfactory use of the computer are correct, for instance it will start a fan if it finds that the temperature inside the processi...
