Elizabeth Johnston, Investigating Minds, 11/21/97

Lecture 19: Can Machines Think?



Turing Machines and the Turing Test

Alan Turing (1912-1954) was a brilliant Cambridge mathematician who is recognized as one of the founders of computer science. He proved that any algorithm can be implemented on Turing machine: a simple device that consists of a tape, divided into squares, and a device that can write symbols on the tape, and read those symbols(Crane, 1995). The idea of a Universal Turing machine underlies modern, general purpose, digital computers.

In a 1950 article 'Computing Machinery and Intelligence' appeared in the journal 'Mind'. This provocative article begins with the sentence "I propose to consider the question 'Can machines think?'" To make this question tractable Turing proposed a novel operational definition of 'thinking', which is now known as the 'Turing test'. Turing introduces it thus:

"It is played with three people: a man (A), a woman (B), and an interrogator (C) who may be of either sex. The interrogator stays in a room apart from the other two. The object of the game for the interrogator is to determine which of the other two is the man and which is the woman. He knows them by labels X and Y, and at the end of the game he says either "X is A and Y is B" or "X is B and Y is A". The interrogator is allowed to put questions to A and B thus:
C: Will X please tell me the length of his or her hair?
Now suppose X is actually A, then A must answer. It is A's object in the game to try to cause C to make the wrong identification. His answer might therefore be
"My hair is shingled, and the longest strands are about nine inches long."
In order that tones of voice may not help the interrogator the answers should be written, or better still typewritten. The ideal arrangement is to have a teleprinter communicating between the two rooms. Alternatively, the questions and answers can be repeated by an intermediary. The object of the game for the third player (B) is to help the interrogator. The best strategy for her is probably to give truthful answers. She can add such things as "I am the woman, don't listen to him!" to her answers, but it will avail nothing as the man can make similar remarks.

We now ask the question, "What will happen when a machine takes the part of A in this game?" Will the interrogator decide wrongly as often when the game is played like this as when the game is played between a man and a woman? These questions replace our original, "Can machines think?"" (my emphasis)

Aside from the interesting feature of introducing the test in terms of gender (note that the woman is the aide to the interrogator) there are many things to note about Turing's formulation: it is a purely functional definition of 'thinking', social interaction is reduced to sending written messages, trickery or dupery is involved, and the human programmer is still in the loop.

Chess as a Restricted Turing Test

Alan Turing was an avid chess player and, along with his good friend David Champernowne, wrote (but did not implement) the first chess-playing program, called 'Turochamp'. It is appropriate, therefore, that one of the restricted Turing tests that has been passed by computers is passing as an exceptional human chess player.

How Humans Excel at Chess

A Dutch psychologist, de Groot, studied chess grandmasters intensively (de Groot (1965), 'Thought and Choice in Chess). He discovered that discovered that grandmasters do not look many moves ahead and examine all the possibilities, instead they excel at recognizing the significance of particular configurations on the board. When asked how many moves ahead he typically looked, the grandmaster, Richard Reti, replied: "One, the right one." The one reliable difference between grandmasters and novices that de Groot documented lay in their memory for chess configurations. Grandmasters are able to reconstruct more than 20 pieces from a chessboard after viewing it for only 5 seconds, whereas novices can reconstruct only 4 or 5 pieces. If the positioning of the pieces is random rather than from a real chess game, the grandmasters do no better than the novices, revealing that their well honed memory applies only to permissable chess configurations. De Groot estimated that a grandmaster holds around 100,000 patterns in his or her head.

This type of 'chunking' or well developed pattern matching expertise is not the strategy employed by programmers of current chess plating machines. The following example of the difference in human and computer chess playing was published in 'New Scientist'.

Grabbing a poisoned rook

Deep Thought II (Deep Blue's predecessor) was playing white. After lengthy searching the computer captured the black rook with its pawn. Deep Thought had no way of seeing that white's only salvation is the pawn-barrier between the white king and black's extra pieces. The barrier is recognizable to human chess players as a pattern in which interlocked white and black pawns form a "Chinese wall", impregnable as long as white does nothing to breach it. Working this out would mean looking ahead more than 50 moves but humans know simply from the telltale pattern. In the next section you will see why 50 moves ahead is too far ahead to consider even for such a computational wonder as Deep Blue.

How Computers Excel at Chess

Computer chess playing programs rely on exactly those capacities that computers are designed for: brute force calculating and searching. This is a formidable task given Claude Shannon's calculation that the possible paths through the maze of a typical chess game could number 10120.
A computer chess program assigns a numerical value to each potential move in a game based on the presence of characteristics known to give an advantage to the player, like material or positional advantage. This is done for a large tree of possible moves, then the minimax principle is then used to decide among the tree of possibilities. The logic of the minimax principle is to maximize your score and minimize that of your opponent. In this example the machine is playing white and the following table shows the difference between considering moves at only the first level, versus working through the consequences up to three levels, employing minimax.
A search using this strategy to a depth of ten levels requires the study of half a quadrillion moves on the average. Deep blue can evaluate 200 million positions per second and search to depths of 14 levels when making a move. While Richard Reti's comment about only looking one move ahead was an exaggeration, it is true that a grandmaster does not examine more than about 30 positions in a tree, and rarely more than 50. In contrast, in some cases, Deep Blue examines an astounding 50 billion or more.

Is Deep Blue thinking?

The answer to this depends upon who you ask, in a surprizing way. The programmers who know the workings of the computer too well to be impressed say not, the World Chess Champion, Gary Kasparov says yes. Last year Kasparov said of Deep Blue:"I believe that signs of intelligence can be found in the net result, not in the way the result is achieved." Before this summer's rematch he declared: "I don't care how the machine gets there. It feels like thinking." It certainly passes a restricted chess Turing test.

In fact, AI researcher Donald Michie postulates that Kasparov lost because of psychological factors. "Deep Blue came across to Kasparov as being able to read his mind, and this slowly wore him down." Deep Blue's speed mean that it can calculate the opponent probable next move while he is contemplating it. Half the time Deep Blue had correctly predicted Kasparov's move and replied instantly to moves he had contemplated for 15 minutes. When this happened Kasparov grimaced. "It was not Deep Blue that destroyed Kasparov, it was Kasparov." Michie suggests that chess playing computers should only be regarded as intelligent if they can hold their own at the press conference afterwards as they explain their winning strategy.

The Difficulty Conundrum

One of the central lessons of the last 40 years in AI research is that problems we thought were hard turned out to be fairly easy, and that problems we thought were easy have turned out to be profoundly difficult. Chess is far easier than a task accomplished apparently effortlessly by a representative three year old: understanding speech. There have been various failed attempts to program computational conversers and we will examine some of these in the next lecture.

Sources: