Possible variants of the definition

      Now we have the perfect CPP and the question is how to define what is a CPP. We can say that this is any program which is as good as the perfect CPP. At least the perfect CPP covers this requirement. This is not a good idea for definition because we need a program which can play in real time. Also, we do not need a player so perfect as the perfect CPP. 
      That is why we will define CPP in the following way:
      First variant of the definition: CPP will be the program which makes average success which is on smaller distance than ε from the perfect. 
      If this ε is chosen smaller enough, then the chess playing program will play no worse than a human.
      Here we have one problem. We do not know how big is the perfect result. If the starting position in chess is winning for the whites, then the perfect result is one (or 1 - δ where δ is zero). If the starting position is not winning for the whites, then the perfect result is 1 - δ, where δ is some very small number bigger than zero. We have an algorithm for calculating the value of δ, but we cannot calculate it because this algorithm will not end in reasonable time (i.e. before the end of the universe). The conclusion is that the previous definition is not convenient because it includes two constants whose value cannot be calculated (δ cannot be calculated due to the combinatory explosion and ε cannot be calculated at all because it depends on the human beings, which means that this constant has no exact value). Even if we choose a value for ε we will still not know the value of the δ+ε. This is the reason to define a CPP in a different way.
      Second variant of the definition: CPP will be the program which makes average success which is bigger than 1 - α.
      Here α is some small positive number. It must be equal to δ+ε in order to be the new variant of the definition the same as the previous variant. In any case α must be greater than δ because in the opposite case CPP will not exist.
      The second variant of the definition gives us the possibility to check it for a concrete program by the tools of the statistics. There are two problems. First, we cannot say how big should be α for the CPP to exist and in order to play no worse than a human being. Of course, such value exists but we cannot say which it is. The second problem is that the value of α is very small, which makes it extremely difficult to calculate it by the methods of the statistics.
      In order to eliminate the second problem we will change the opponent. We need a stronger player as opponent in order to have bigger expectation for δ+ε. Let us try with human being as an opponent.
      Third variant of the definition: CPP will be the program which makes average success against human being bigger than 1/2.
      The third variant of the definition is not formal, but this variant is usually used as a definition of CPP. The good side of this definition is that it can be checked by the methods of the statistics. For example, if we make ten games between one program and a human being and if the average success of the program is 60% (0.6), then we can say with big possibility that this program satisfies the definition for CPP. Really ten games are too few and for bigger certainty we may make 100 or even 1000 games. Here our expectation for δ+ε is 1/2. Really the game is not symmetric (opponent plays always with the blacks) and that is why we are not sure that human against human will make 1/2 average success. Also, you have to notice that this variant of the definition is not formal and that the results will depend on the humans we use for opponents.
 


New fixed opponent

      Anyway, we need a formal definition of CPP and that is the reason why we will use a well-defined opponent. Such opponent is the standard chess playing program which calculates N moves in the tree of the game and by Min-Max algorithm selects the best move. The problem is that this opponent is deterministic. We prefer the opponent to be nondeterministic because if both CPP and the opponent are deterministic, then the result of the game between them is also deterministic (actually, there is only one possible game between them). In this case we cannot use the methods of the statistics. Also, it is not a good idea to use deterministic opponent for the definition because all deterministic programs which have average success one will be CPP but maybe some of them accidentally made victory in this only game, which does not prove that these programs are CPP.
 

      For opponent we will use nondeterministic variant of the standard chess playing program. Let our opponent depend on one integer N and two functions - F and P. Let function F evaluate positions and on every position F return an integer. Let function P give the possibility for one move to be chosen. This means that the input of P is a list of integers which represent the values of the positions which can be obtained on the next move and the output is a list of the possibilities each of these positions to be obtained. The integer N and the functions F and P should be concrete and fixed because they will be parameters in our definition. If we change one of these parameters, then the definition will also be changed.
      How will this fixed opponent work. It will evaluate by the function F all positions which are obtainable after N moves from the current position. After this, by the Min-Max algorithm, it will calculate the values of the positions which are obtainable on the next move. After that, by P, it will calculate the possibility of each of these positions to be obtained and on the end will randomly choose one of the moves, but in such a way that the possibilities of the next positions to be the same as the ones given by P.
      This fixed opponent is nondeterministic (except the case when P gives 1 to one of the numbers and zero to the rest of them).
      After the change of the opponent we also have change in the perfect CPP. It will be constructed almost in the same way, but instead of average value it will calculate the sum of Value(i).Possibility(i) where i runs true all possible moves of the opponent. Here Possibility(i) will be given by the function P. (Of course, before applying the function P we have to use function F and Min-Max for N moves in order to calculate the values of all positions which are obtainable on this move.)
      If the number N is big enough and if the functions F and P are reasonable, then our fixed opponent is a good chess player. This means that our expectation for δ and ε will be high. Let us assume that ε is 10%. This means that we suppose that a human being will make average success against the fixed opponent 10% worse than the perfect CPP. Let us assume that the average success of the perfect CPP is 80% (i.e. that δ is 20%). If the starting position is winning for the whites, then δ is zero but our expectation is that the starting position is not wining and not losing. This leads us to the following:
      Final variant of the definition: CPP will be the program which makes average success against the fixed opponent which is bigger than 70%.
      As we mentioned, this definition depends on the fixed opponent and from the number α, which here is fixed to 30%. If we fix also the number N and functions F and P, then we will have one concrete definition. We can easily check whether one program satisfies the requirement of this definition by the methods of statistics. Really, we cannot check this for sure, but with a great percent of possibility (which percent grows with the number of the test games).
      This definition is made with one assumption:
      Assumption 1: Here we assume that the average success of the perfect CPP is about 80%. If this conjecture is true, then there exists a program which satisfies the definition (at least the perfect CPP does). If the average success of the perfect CPP is smaller than 70%, then there is no such a program (of course, in such case we can change this parameter and make it smaller than 70%).
 


Conclusions

      The first concussion is that CPP exists and that the perfect CPP satisfies the demands of the definition. Really, this is true if assumption 1 is true, but if we take the fist variant of the definition then this is true without assumption 1.
      The second conclusion is that the perfect CPP is CPP but it is useless due to the combinatory explosion. In order to obtain a real CPP we have to optimise the perfect CPP and make a program which may be a worse player than the perfect but which can play in real time. That is an easy task in the case of chess, but not so easy in the case of AI.
 

References

[1] Dobrev D. A Definition of Artificial Intelligence. In: Mathematica Balkanica, New Series, Vol. 19, 2005, Fasc. 1-2, pp.67-74. (The same in popular form)

[2] Dobrev D. Formal Definition of AI. In: IJ ITA, Volume 12, Number 3, p.277.
 
 
 

Page 2 of 2     <<Previous       AI-Project