Dimiter
Dobrev
Institute of Mathematics and
Informatics
Bulgarian Academy of
Sciences
1113
Sofia, Bulgaria
dobrev@2-box.com
Abstract:
Two different definitions of the Artificial
Intelligence concept have been proposed in papers [1] and [2]. The first
definition is informal because it says that the programs that are cleverer than
human are acknowledge as Artificial Intelligence. The second definition is
formal because it avoids reference to the concept of ‘human’. The readers of
papers [1] and [2] are left with the impression that both definitions are
equivalent and the definition in [2] is simply a formal version of that in [1].
This paper will compare both definitions of Artificial Intelligence and will
hopefully help for the understanding of the concept.
Keywords: Artificial Intelligence, AI Definition.
What is the main idea?
The idea behind the definitions of Artificial
Intelligence in [1, 2] is as follows. If a program is intelligent it would
manage well in an arbitrary world. This is a version of the popular wisdom that
the clever person can handle any job. Certainly, the clever person will manage
not immediately but after some training (learning).
Therefore, when a program is measured how it
manages in a particular world, it should go through a certain period of
training. Only when the training is over, we should assess how well the program
is doing.
We can make an analogy with humans. The first
eighteen years of human life are considered as a period of training, if we do
wrong and even commit a crime, the punishment will be weaker, because it is
believed that we are still on training. The training period with animals is
usually shorter; this is usually associated with their shorter lives and the worse
living conditions. The rule that is often true about animals is ‘Learn fast or be
eaten!’.
How long is the training period and is it
possible to say when exactly it is over?
Here comes the first difference between the
definitions in [1] and [2]. The first definition supposes that life is infinite
and there is enough time for training. This is to say, the training period is an
arbitrary finite beginning of life, which is infinitely less than the whole
life that is infinite.
The approach of the second definition is
different. It assumes that life is limited and there is a parameter will be referred
to as ‘the maximum length of life’. The assessment of intelligence is based on
life – from its beginning to the end. The second definition is different with
view of the training period, as well. There is no period of training with the
second definition. The rule which is true is ‘Learn fast or be eaten’.
Why the period of training is zero with the
second definition? Because we cannot say how long should be this period and
when it will be completed. Therefore, it is convenient to assume that there is
no such period.
Are both definitions
equivalent?
It follows from the above that both definitions
are not equivalent because the retarded program is Artificial Intelligence
according to the first definition, however this is not true according to the
second definition. Retarded is a program that needs nearly infinite time for
training (i.e. its period of training is finite but in practice it is
infinite). Examples for retarded programs are TD1 and TD2 in paper [2].
The retarded program does not match our idea of
Artificial Intelligence. Our idea is to construct a robot that can sweep the
floor and if the robot needs 1000 years to learn how to sweep, this would not
work for us.
Are there other programs that formally satisfy the
definition but do not match the idea of Artificial Intelligence? Yes, the
extremely inefficient program is such a program, even if it satisfies both
definitions. Paper [2] describes such a program as Trivial Decision 5. This program
would work if we had an infinitely fast computer, as it makes nearly infinite
number of steps (i.e. finite many but in practice infinite) in order to
calculate only one step.
Both definitions define a set of programs. These
sets should coincide in order to make the definitions equivalent. Therefore,
both definitions are not equivalent and the example for this is the retarded
program.
There are two additional reasons why both
definitions are not equivalent:
Firstly, the definition in [1] is informal and
dependent on people. That is to say, it does not define a particular set but
rather a sort of fuzzy one. We have said that we acknowledge for Artificial
Intelligence these programs that are cleverer than human. The reasonable
question to ask with that statement is: ‘Who is the man we compare with?’ A
possible answer is that the program is cleverer than any human, however, this
would not define the set of these programs in a unique manner. For example, if
we define the chess-playing program like a program playing chess better than
the world champion, there will be still programs that will play chess better
than one particular world champion and worse than another. Therefore, a formal
and an informal definition cannot be equivalent.
Secondly, the definition in [2] is dependent on
several parameters. This is to say, that this is not the case of set of programs
but rather a function that for different parameter values returns different sets
of programs.
The assumption made in [2] is that such
parameter values exist for which the set defined by the definition is not empty
and the programs within that set match our idea of Artificial Intellect.
Certainly, not all programs within that set match
our idea. We have already mentioned the problem with the infinite ineffective
program. Another problem that appears with the definition in [2] is the ‘cramming’
program. The problem is that in [2] it is assumed that the worlds we are
interested in are finite in number. Therefore, a program can be created specially
for these worlds. We come across the same problem with the candidate students –
when people who have crammed all possible exam topics sit for the exam. Such
people will manage and will pass the exam but they cannot solve any problem out
of the range of the crammed exam topics.
The assumption is that if the shortest and the
most efficient program is taken out of the those that satisfy the definition,
then it will match our idea of Artificial Intelligence. We should limit the
programs both in length and in efficiency because the infinitely inefficient
program is rather short, whereas the ‘cramming’ programs are rather efficient.
The ‘cramming’ program is shorter than the Artificial Intellect for a small
number of worlds; however the Artificial Intelligence is a shorter program than
the ‘cramming’ one if the number of worlds is sufficiently great.
Lifetime
The first parameter which the second definition depends on is the lifetime.
Once we give up infinite life, then we should limit it and set a parameter
indicating the life expectancy. To make it simple, the life in [2] has been fixed at 100 games, each one no longer than 1000 steps.
Giving up infinite life, we get rid of the
retarded program. Another advantage is that it is no longer required that the
worlds are without fatal errors. This requirement was important in [1] when we
needed sufficient time for training, presuming that no fatal error is made that
would ruin the whole life. This is to say, we can easily make mistakes as each
one can be overcome and none is fatal.
Definition: Fatal error is a group of internal states of the world, such that when
we enter this group we can no longer exit. If there is an exit, the error would
not be fatal. Besides, the world in this group should be worse than that out of
the group (i.e. the assessments made there are relatively lower). If the world
in the group is not worse, then the fact that we happen to be there would not
be a mistake.
Another possible definition of fatal error is
the following. If we calculate for each moment of life what is the maximum
anticipated success of life (what returns the Success function), provided we
play (live) by the best possible strategy, from this moment onwards, then fatal
error will be referred as a step after which this number is decreasing.
Having assumed that life is finite, it is not
needed to assume that there are no fatal errors in the world, because our time
for training is anyway limited. When life is finite, a common mistake can be
equal to fatal, because time may not be sufficient to fix it.
Thus, it is natural to assume that life is
finite and we are looking for a program that can manage well within certain
lifetime and not with any lifetime. On the other hand, it is inconvenient that there
are parameters included in our definition. It would be better to define the
Artificial Intellect as a program independent of anything. That is one program,
irrespectively of the anticipated lifetime.
However, life
expectancy is an important parameter
that influences the strategies.
Let’s consider the human behavior in war time, natural disasters and other catastrophes.
When life expectancy becomes shorter, the behavior of people significantly
changes. This is expressed mainly in the tendency to take greater risks. You
may also notice that the young people are braver than the older. A possible
explanation of this observation is that the young are more willing to
experiment and take risks while the adults prefer the stable and the secure
because they estimate that there is no time for experimenting. Thus, we can say
that life expectancy definitely influences behavior of people and their life
strategies.
Arbitrary world
The first definition requires for the
Artificial Intellect to manage as good as human in an arbitrary world. This
requirement is so strong that it appears that there is no program that can meet
the requirement and the set of programs satisfying the definition may prove to
be empty.
Let’s try to create a world which is too
complex for any program to manage but not for the human. Imagine a world where
robots are not liked. In such world, if you are recognized as a robot you immediately
score low. However, if you are considered human you score high. It seems that
this is the world where humans will do better than robots. Let’s remember the definition
of a world; there are two functions (World and View) defining the world. They are
absolutely arbitrary functions and we can presume that they return success when
there is human living in the world and respectively lack of success in the case
of a robot. Still, the world is not God and there is no way to know if its
inhabitant is human or robot. The world will know who is who based on the acts
of the opponent. That is, if the robot behaves as human and acts appropriately
then the world will be deceived and will accept it as human. In this case, the
program is required to play the imitation game. The same game was proposed by
Turing as test for Intelligence. It is to be concluded that if a program meets the
definition in [1], then it is satisfactory to the Turing test, not immediately
but after some time of training.
Question: Is it possible that the world recognizes the robot (while it is still
on training and has not started to act like a man) and starts scoring low from
that first moment of recognition till infinity? The answer is: No, because only
worlds without fatal errors are considered and this world does not meet this
requirement.
Does that mean that the definition in [1] is
equivalent to the Turing test? Not if we train the program in [1] to pretend being
a human, then it will satisfy the Turing test but only after training. Is it
possible that the program satisfying the Turing test to be trained to manage in
an arbitrary world? The answer is: rather not. If the program can pretend to be
human, then it can be trained. However, it should rather pretend to be stupid
and hide its intelligence otherwise, it will betray the fact that it is a robot
and not human. If it is forbidden for the examiner to punish the excessive intelligence,
then the definition in [1] and the Turing test will be equivalent.
Impossible World
Is it possible that the world is so complex
that there is no program that can understand it? Yes, it is. For example, let
the world generate an infinite row of zeros and ones. The Artificial
Intelligence has been given a task to have a guess what comes next (zero or
one). Let the function describing this infinite row is not computable. Then,
there is no way for the program to calculate and make a guess which number will
follow. This is true about the human, as well. Nevertheless, the program and
the human will find different dependencies. For instance, such that the zeros
are more than the instances of having one, that it is more likely that one
comes after zero than zero and etc.
It is not necessary that the Artificial
Intelligence should understand the world at 100 %. What is important is to
understand the world better than the human.
IQ (Intelligence Quotient)
The first definition has compared the
intelligence of the program with that of human. The second definition cannot
allow making the comparison with man (because we want the definition to be formal).
Therefore, it is necessary to introduce an independent assessment of IQ by
which we could define the Artificial Intelligence. We will say that for AI we
acknowledge those programs whose IQ is above certain value. This value was
decided to be 0.7 in [2] but this choice has been largely arbitrary. It is
rather correct to say that certain IQ exists and the programs more intelligent
than this level are acknowledged as Artificial Intelligence.
We introduce the function Success which returns
a number in the interval [0, 1] for each particular life. This number makes assessment
of the device success in the particular life. Afterwards, the IQ is calculated
selecting a set of test worlds, running the program to live a life in each one
of these worlds and calculating the average success of the program in all its
tested lives.
Thus, the IQ is the average value of Success
function calculated based on the set of tested worlds.
World Complexity
Another substantial difference between both
definitions is that the first considers all possible worlds; whereas the second
limits the sets of the worlds to a finite number of tested worlds (the
assumption is that the tested world is computable with fixed level of
complexity). This fixed level of complexity is the next parameter of the
definition.
Why did we select the set of the tested worlds
to be what it is?
Something similar has been done in paper [1].
It has proposed to prepare a test consisting of finite or countable number of
tested worlds. The idea is to acknowledge as Artificial Intelligence the
program that can manage in all these worlds. Paper [1] has proposed that these worlds are prepared by human but we want
to be maximum formal in paper [2] and therefore we will define the set of
tested worlds in such a way that it would not depend on human’s choice. The
other difference is that in [1] we want the program to pass all exams, i.e. to
manage in all tested worlds, whereas in [2] we want the average success (i.e.
IQ) to be greater than 0.7. Why we want to have less in [2] than in [1]?
Because if the problems are preliminarily prepared we would want that the
program will solve them all, but if the problems are randomly generated then
there will be such that are not solvable and therefore we cannot rely that the
program will do them all.
What will be the set of worlds that we will use
for the calculation of the IQ of an arbitrary program?
The first natural possibility is to take the
set of all worlds. This set is infinite, even uncountable and seems too large
(it is not clear what should be the weight that different worlds will
participate with). The first thing we find is that many of the worlds are indistinguishable
(i.e. their tree of the world is equal).
That’s why we resort to the next idea that is to take the quotient set i.e. the
set of all possible trees of the world and make it our set of tested worlds.
This set is again uncountable, but considering the fact that we limited the
lifetime, we see that the set of these trees is even finite (more precisely,
the set of the trees of determined worlds is finite. It is finite with the
undetermined worlds, as well, because the branches are equally probable to
happen – see the definition of TM_W in [2]).
This set is not suitable (although it is
finite) because anything is possible in such a case! How will the world of the
next step continue to be? It may continue as it likes to. Anything is possible,
indeed, but far from anything is likely to happen. If we accept this set as a tested
one, then any continuation will be equally probable and what has happened so
far will be of no importance. This totally contradicts our idea of Artificial
Intelligence which says that the device gathers experience and is on training. Thus,
what has happened so far is important.
This is the time to apply the principle known
as ‘The Occam’s razor’ stating that the simpler model is more probable than the
complex one. Therefore, the simpler world is more probable than the complex
one. If we are to discuss the complexity of the world, then we would introduce
a description of the world and define the complexity of the world as the length
of the shortest possible description.
We have used the Turing’s machines in [2] to
describe the worlds. This is not the most suitable model in case you try to
make a real program which satisfies the definition. However, it works just as a
theoretical model of computability. Still, we do not want to be limited within
the set of the determined worlds and this is why we have introduced
undetermined Turing’s machines. Thus, our tested worlds are the computable
worlds generated by the undetermined Turing’s machines.
The next suggestion is to select the set of
tested worlds to be the set of the undetermined Turing’s machines taking on
board all such machines regardless of their size. Could we chose a particular
size and do only with these machines? The answer is: rather, yes.
If we take all Turing’s machines, then we
should give them different weights. As there is no way that infinitely many machines
are of equal weight and the sum of their weights to be one. Having decided what
will be the weights of the different machines, there are two options to go for:
either the average size (length) of the machines is a particular number, or the
average size is infinity (depending on the weights we have finally chosen). If
the average size is finite, then we can assume that instead of having all Turing’s
machines, only those with average size will be the tested ones. This is not
same but is almost the same. If the average size is infinity, then there is an
N and all machines with greater size would almost make no influence. Therefore,
is we chose an e that seems to us small enough to be ignored, then such N exists that
the machines longer than N will influence less than e of the average success. Then we can decide that N is the size of the
tested machines and the result will be closer to that which would get at, if we
consider all machines with their respective weights.
Next question: If we have decided on a
particular N, should our tested machines be these of size smaller or equal to N
or those with the size of N exactly. The answer is: there is no need to include
the shorter machines because each machine with size N-1 has many equivalent
machines with size N (because we can add a state that is not necessary).
So far, so good. We have decided that the
tested worlds will be the computable worlds that are computable by an
undetermined Turing’s machine having the size of N. This makes the next
parameter that our definition depends on. We decided on a particular value of
20 for this parameter in [2]. We decide that all tested machines will
participate with equal weights (this is possible because the set is finite).
Does the set chosen in this way correspond to
the principle of Occam? Are the simpler worlds more probable than the complex ones?
The answer is: yes. Indeed, all machines participate with equal weights but the
simpler machines have a great number of equivalent machines (such that compute
the same world), whereas there is not a single equivalent machine for the most
complex ones (certainly, of the same complexity, in this case with complexity
value of 20). This is to say that the simpler the world is, the more machines
with size 20 can compute it and the more this world would influence the average
value of Success function. We refer to this average value as IQ.
Which is the suitable model?
We have already said that the Turing’s machines are not the suitable
model to describe the world. We would like to have simple dependencies in the
world that are on the surface and easy to be discovered, whereas more and more
complex dependencies to be discovered the deeper we go. The Turing’s machine is
a dependence that may appear to be rather complex but once you understand it, you
have understood the world. The case of the undetermined machines is better
because the randomness is an infinitely complex dependence. Therefore, we will
never understand this dependence because once we understand it; it will become pseudorandomness
(take the example of pseudorandom numbers generated by the computer).
Is it possible that a complex Turing’s machine is partially described by
means of simpler dependencies? This is possible but it is not typical for the Turing’s
machine model. In this model usually, you either understand the whole world or
you do not understand anything.
If we are looking for a world model of the type of а determined machine,
then very soon (i.e. after very short life experience) it would turn out that
the first Turing’s machine corresponding to that life experience is so complex
that virtually it is impossible to find it. The advantage of the undetermined machines
is that we will always be able to find such world model (no matter how long is
the life experience). It is another point how adequate is this model and how good it will work for us, because the undetermined machine does not say what
will happen with the next step but it rather says that this or that can happen.
In the best case, it provides the probability of having this or that
happening.
Well, which is the suitable model of the world? We should think of the
world as a union of different factors that may be related but are largely
independent. Certainly, we will need a better model if we want to create a
particular program satisfying the definition of Artificial Intellect, but this
is not important for definition itself.
Work of other researchers
The occasion to write this paper is publication
[5] where two scientists from Switzerland
have tried to generalize the definitions in [1, 2]. Their idea was to get rid
of the parameters, which the definition in [2] depends on and to get to a new
concept of IQ independent of any parameters.
They have removed the limitation on lifetime
and have assumed that life is infinite. In their representation the beginning
of life is the most important part of life. They assume that the rewards become
lighter at any next step having been multiplied by the coefficient of discount.
Surely, the coefficient of discount is a
parameter, as well, and they have simply replaced one parameter with another.
What is more, their presentation is contrary to the idea that the beginning of
life is not important. What is important is what happens having the program already
been trained. Their presentation assumes the beginning of life as the most
important part. The life is so much discounted that from a moment onwards, in
practice, it is not important what the program is doing. Certainly, there is a
moment (the maximum length of life) in [2] from which onwards there is no point
what the program is doing, but at least until that moment all rewards are of equal
importance. This is to say, that ‘it runs, runs and stops’ in our case, whereas
‘it fades, fades, fades and like this to infinity’ in their case.
We have to acknowledge that the authors of [5]
have understood that the coefficient of discount is a parameter which their
definition depends on and have, therefore, proposed a second version. It would
have been better if the second version have not been proposed, at all, as it has
resulted in meaningless outcome compromising the whole paper. More, about the second
version, has been written further below (striking mistakes 3 and 4).
The other parameter, that the authors of [5]
have tried to get rid of, is the world complexity. We have decided on
particular complexity (i.e. number of states of the Turing’s machine which
generates the world). They have preferred to sum up all complexities having
used a coefficient of discount 1/2. This results in having average complexity
of 2 in their case (this parameter has been picked to be 20 in our case). This
is to say, that if they want to allow for higher values of their average
complexity, they will have to replace the number 1/2 with a different parameter.
Therefore, they replace one parameter with another, again.
They get rid of the parameter 0.7 by omitting
to say what Artificial Intelligence is. They say what is IQ but they do not say
how big the quotient should be in order to acknowledge a program as Artificial
Intelligence.
Unfortunately, our Swiss colleagues have
omitted to quote the Bulgarian source. Another problem is the fact that they
have not managed to understand many details of the original papers and,
therefore, there are many mistakes and wrong things in the resultant text.
Striking mistakes
These are six of the most striking mistakes
made in [5].
1. Their paper has written that the world is
computable according to the thesis of Church. It is true that it has been
written in [1] that it follows from the thesis of Church that the Artificial
Intellect is a program but not that the world is a program, too. Is the world
computable or not, is it determined or not, these are questions whose answers
we do not know and will never know. This is something that cannot be verified because
there is no such experiment that can result in the answer of these questions.
(The question is the world determined has been considered in details in [3].
The question is the world computable is analogical).
2. They have defined the IQ in such a way that it
is infinity for each program. It was not envisaged that the number of programs
grows exponentially with the increase of their length. This can be regarded as
an oversight mistake, moreover that it is clear how it can be fixed. (in their
case, they sum up the programs’ success achieved in different lives, i.e. the
values of function Success. This sum does not require that each addend is
multiplied by 1/2 raised to the power of the complexity degree, but would
rather take the average for the respective complexity and multiply it by the
same). It has been said above that the average complexity of the world in [5]
is 2, having assumed that this mistake has been fixed. If what has been written
remains unchanged, then the average complexity is infinity and the sum that we
refer to as IQ is infinity, too. Thus, if the mistake is not fixed, the IQ
concept becomes meaningless.
3. The most serious problem in [5] is the second
version that was proposed in order to avoid the coefficient of discount. The
set of worlds, in this version, is different and respectively the Success
function is different (this function that says for each life what success has
been achieved in this life).
Let’s summaries the outcome in [1], [2] and both
versions in [5].
An infinite life in a set of worlds without
fatal errors is considered in [1]. It is assumed that life is finite in [2].
The first version proposed in [5] describes life as infinite but fading which
is the same, as if the life is finite. The second version proposed in [5]
assumes infinite life in a world where all errors are fatal. The problem, in
this case, is not that there could be a fatal error (the fatal errors do not
interfere in [2] and will not interfere in this case, as well). The problem is
that all errors are fatal. This is to say, that there are not fixable errors.
Humans do not learn from their fatal errors, perhaps, they have learned from
the fatal errors of others but not from their own. That’s why this concept
contradicts the idea of learning.
The Success function is monotonically increasing in the second version proposed
in [5]. It was set to be the sum of all rewards which are
numbers in the interval [0, 1]. It is natural to consider that the Success
function can be increasing and decreasing during the lifetime. It can be
considered as a fixable mistake, when it starts to decrease. The authors of [5]
have preferred to have this function monotonically increasing which means that the device cannot make fixable errors. The only
possible mistake is of the type ‘lost profits’ and this mistake is always fatal
because once the profits are lost there is no way to bring them back. Besides,
there is no feedback with the lost profits. Thus, when the Success function
changes, the device will know but there is no way to know when the profits are
lost. Eventually, it may know in future, but life is infinite and therefore, it
will not know even in future. The device will always hope that the profits are
not lost and will soon emerge.
4. The decision of the authors of [5] to limit
the sum of the rewards is very strange and illogical. It reminds me of one of
my teachers, for who the students were telling that he had a limit on the A
grades and you should be among the first to be examined because the A grades
will end. No matter how much knowledge you demonstrate if you are at the end of
the exam you will not be A-graded because of his limitation.
This limitation is imposed, in their case, so
that the Success function will be in the interval [0, 1]. Instead of distorting
the world in such a horrible way, it would be better if the Success function is
the arithmetical mean of the rewards (as it was done in [2]) instead of being the
sum of the rewards. This would have made the Success function being in the
interval [0, 1], and it would not be monotonically increasing.
The conclusion is that the authors of [5] use
worlds where training is impossible in their second version of the IQ
definition. The success of the device in such a world depends solely on being
lucky, however when there are many worlds the luck ceases to act. Thus, all
programs are equally intelligent.
The question to ask is whether the authors of
[5] have managed to understand that both [1] and [2] consider a device that is
being trained and will achieve good success as a result of the training, or
they rather believe that the device was born trained. As a matter of fact, the
beginning of [5] says that the device should be given sufficient time for
training, however, later they propose two versions of the definition which
contradict this idea (especially the second version).
It is true that with the definition of Turing,
there is a device that was born trained, but this concerns a particular world.
The device can be born trained for a particular world but there is no way that
it was born trained for any world.
5. We can consider for a mistake the fact that the definition of a world
in [5] has been changed. The world in [1] has got a set of internal states and
a function indicating how to transfer from one state to another. As you know, there
is a tree of the world corresponding to each world. It was the tree of the
world that has been regarded for the definition of the world in [5]. This is
the same as if we do not consider functions in mathematics but only graphs of
functions. We can somewhat justify the authors of [5] because they make their
attempt to improve somehow the definition of AI, however their change implies
that they have not understood our main idea. We suppose that the world has got
some structure and the device tries to understand that structure. They deny the
structure of the world and this is a mistake.
They act analogically when they define the device, as well. The device
is a program for us, whereas for them it is a strategy. Certainly, there is a
strategy corresponding to each program (but not vice versa). Still, to consider
the device as a strategy is a mistake, as thus we presume that it has not got
internal states, i.e. no memory. This mistake is very common among researchers working
in the area of AI. Many of them look for the AI in the set of the functions
meaning, that for them, AI is a device without memory. The strategy is also a
function, whose input argument is the entire life experience. At first sight,
it looks as if we do not need the memory if we have available the entire life
experience, but this is not true.
Imagine, that at some point (based of your whole life experience) you decide
to go to the fridge and grab a beer. Ten seconds later, you see that you are on
your way to the fridge but you do not know if you are going to get a beer or
milk. It is true that you can rely on your whole life experience but this will
not answer the question. If you were told to fetch a beer, this can be
extracted from your life experience, but if you have, yourself, decided to grab
a beer, you would not remember it (because you have no memory) and there is no
way that you extract it out of the life experience. This is to say, the memory
is needed. We have to note that it is absurd that the device will make a
decision at any step based on its whole life experience (because this is a huge
amount of information). It is more logical to assume that it decides based on
its internal state and the immediate input received at the last step.
6. The last comment to make with reference to
[5] is the strange reasoning whether the rewarding is to be part of the world
or part of the device. The rewarding has been confidently treated as part of
the world at the beginning of [5] and the reasoning about its belonging at the
end of the paper is rather odd. The authors answer this question themselves
saying that if the students are allowed to mark their knowledge themselves they
will all have excellent results.
Nevertheless, the confusion of Shane Legg and
Marcus Hutter about the belonging of the rewards is reasonable. They ask
themselves the question, if the rewards, in the case of human, is the feeling
of pain and pleasure (with food, sex, music and other pleasure sources). The
answer we would give is that human has not built-in reward. The success with human
is evaluated by the world through the evolution. According to the evolution meaning,
these who survive in the world are those who manage to survive and to reproduce.
Human does not receive the evolution meaning by default. Actually, humans do never
receive it. If one has lived in compliance with the evaluation meaning, that he
is not aware about, then he will survive long enough and will be inherited in
the next generation. This is the reason why people are looking for the meaning
of life all their lives (they are looking for it, because they do not know it).
As long as the feeling of pain and pleasure are concerned, this is not the
meaning of life but an instinct. Thus, humans are born with some knowledge.
They instinctively know that pain is bad, whereas pleasure is good. These
instincts should not be trusted blindly. For example, the feeling of pain when
the dentist pulls a tooth is misinform (as far as the dentist pulls the right
tooth, of course). The bitter taste in food is instinctively perceived as bad
but with time people get to like the taste of the coffee and the beer, for
example. The feeling of pleasure is also often misinform.
Despite all the remarks we have made, the
papers of Shane Legg and Marcus Hutter are very valuable to us, because they
are the first to acknowledge the definitions in [1, 2]. Furthermore, the
analysis of the mistakes made by our Swiss colleagues is also useful because it
indicates what has not been explained sufficiently well. It is obvious that
Shane Legg and Marcus Hutter have done serious work on this subject and it was useful
to us to study their experience and the attempt to improve our definition.
Where they have not understood the definitions of [1, 2], this is our fault, as
we seem to have poorly explained, before. This paper was prepared on the basis
of this analysis and we hope that it will help them and other researchers
working in the field of Artificial Intelligence.
Acknowledgements
I want to thank professor Dimiter Skordev and
professor Tinko Tinchev for their comments and advice to me regarding this
paper. Furthermore, I would like to thank profession Skordev that he challenged
me at a time to write the definition [2] saying the following regarding
definition [1]: ‘This definition is not formal, at all. You claim that it can
be formalized but what is not clear is how it can be done!’ This was a
constructive criticism and its result was definition [2].
References
[1] Dobrev D. D. AI - What is this. In: PC
Magazine - Bulgaria, November'2000, pp.12-13 (www.dobrev.com/AI/definition.html).
[2] Dobrev D. D. Formal Definition of
Artificial Intelligence. In: International Journal "Information Theories
& Applications", vol.12, Number 3, 2005, pp.277-285. (www.dobrev.com/AI/).
[3]
Dobrev D. D. AI - How does it cope in an arbitrary world. In: PC Magazine - Bulgaria,
February'2001, pp.12-13 (www.dobrev.com/AI/).
[4] S. Legg and M. Hutter. A Universal Measure
of Intelligence for Artificial Agents, In
Proc. 21st International Joint Conf. on
Artificial Intelligence (IJCAI-2005), pages
1509–1510, Edinburgh,
2005.
[5] S. Legg and M. Hutter. A formal measure of
machine intelligence. In Proc. 15th Annual Machine Learning Conference of
Belgium and The Netherlands (Benelearn'06), pages 73–80, Ghent, 2006.
[6] S. Legg and M. Hutter. Universal
Intelligence: A Definition of Machine Intelligence, Minds & Machines, 17:4
(2007) pages 391-444.
