﻿ Puccetti (Roland) - The Chess Room: Further Demythologising of Strong AI (Theo Todman's Book Collection - Paper Abstracts)
The Chess Room: Further Demythologising of Strong AI
Puccetti (Roland)
Source: Behavioral and Brain Sciences, Volume 3 - Issue 3 - September 1980, pp. 441-442
Paper - Abstract

 Paper Statistics Books / Papers Citing this Paper Colour-Conventions Disclaimer

Full Text

• On the grounds he has staked out, which are considerable, Searle seems to me completely victorious. What I shall do here is to lift the sights of his argument and train them on a still larger, very tempting target.
• Suppose we have an intelligent human from a chess-free culture, say Chad in Central Africa, and we introduce him to the chess room. There he confronts a computer console on which are displayed numbers 1-8 and letters R, N, B, K, Q, and P, plus the words WHITE and BLACK. He is told WHITE goes first, then BLACK, alternately, until the console lights go out. There is, of course, a binary machine representation of the chessboard that prevents illegal moves, but he need know nothing of that. He is instructed to identify himself with WHITE, hence to move first, and that the letter-number combination P-K4 is a good beginning. So he presses P-K4 and waits.
• BLACK appears on the console, followed by three alternative letter-number combinations, P-K4, P-QB4, and P-K3. If this were a "depth-first" program, each of these replies would be searched two plies further and a static evaluation provided. Thus to BLACK'S P-K4, WHITE could try either N-KB3 or B-B4. If N-KB3, BLACK'S rejoinder N-QB3 or P-Q3 both yield an evaluation for WHITE of + 1; whereas if B-B4, BLACK'S reply of either B-B4 or N-KB3 produces an evaluation of, respectively, +0 and +3. Since our Chadian has been instructed to reject any letter-number combinations yielding an evaluation of less than +1, he will not pursue B-B4, but is prepared to follow N-KB3 unless a higher evaluation turns up. And in fact it does. The BLACK response P-QB4 allows N-KB3, and to that, BLACK'S best countermoves P-Q3, P-K3, and N-QB3 produce evaluations of +7, +4, and +8. On the other hand, if this were a "breadth-first" program, in which all nodes (the point at which one branching move or half-move subdivides into many smaller branches in the game tree) at one level are examined prior to nodes at a deeper level, WHITE'S continuations would proceed more statically; but again this does not matter to the Chadian in the chess room, who, in instantiating either kind of program, hasn't the foggiest notion what he is doing.
• We must get perfectly clear what this implies. Both programs described here play chess (Frey 1977), and the latter with striking success in recent competition when run on a more powerful computer than before, a large scale Control Data Cyber 170 system (Frey 1977, Appendix). Yet there is not the slightest reason to believe either program understands chess play. Each performs "computational operations on purely formally specified elements," but so would the uncomprehending Chadian in our chess room, although of course much more slowly (we could probably use him only for postal chess, for this reason). Such operations, by themselves cannot, then, constitute understanding the game, no matter how intelligently played.
• It is surprising that this has not been noticed before. For example, the authors of the most successful program to date (Slate & Atkin 1977) write that the evaluative function of CHESS 4.5 understands that it is bad to make a move that leaves one of one's own pieces attacked and undefended, it is good to make a move that forks two enemy pieces, and good to make a move that prevents a forking manoeuvre by the opponent (p. 114). Yet in a situation in which the same program is playing WHITE to move with just the WHITE king on KB5, the BLACK king on KR6, and BLACK's sole pawn advancing from KR4 to a possible queening, the initial evaluation of WHITE'S six legal moves is as follows1:
Move:          K-K4 : K-B4 : K-N4 : K-K5 : K-K6 : K-B6
PREL score: 116 -  114  -  106  -  121  -  129  -  127
In other words, with a one-ply search the program gives a slightly greater preference to WHITE moving K-N4 because one of its evaluators encourages the king to be near the surviving enemy pawn, and N4 ' is as close as the WHITE king can legally get. This preliminary score does not differ much from that of the other moves since, as the authors admit, "the evaluation function does not understand that the pawn will be captured two half-moves later (p. 111)." It is only after a two-ply and then a three-ply iteration of K-N4 that the program finds that all possible replies are met. The authors candidly conclude:
The whole 3-ply search here was completed in about 100 milliseconds. In a tournament the search would have gone out to perhaps 12-ply to get the same result, since the program lacks the sense to see that since White can force a position in which all material is gone, the game is necessarily drawn, (p. 113).
• But then if CHESS 4.5 does not understand even this about chess, why say it "understands" forking manoeuvres, and the like? All this can mean is that the program has built-in evaluators that discourage it from getting into forked positions and encourage it to look for ways to fork its opponent. That is not understanding, since as we saw, our Chadian in the chess room could laboriously achieve the same result on the console in blissful ignorance of chess boards, chess positions, or indeed how the game is played. Intelligent chess play is of course simulated this way, but chess understanding is not thereby duplicated2.
• Up until the middle of this century, chess-playing machines were automata with cleverly concealed human players inside them (Carroll 1975). We now have much more complex automata, and while the programs they run on are inside them, they do not have the intentionality towards the chess moves they make that midget humans had in the hoaxes of yesteryear. They simply know not what they do.
• Searle quite unnecessarily mars his argument near the end of the target article by offering the observation, perhaps to disarm hardnosed defenders of strong AI, that we humans are "thinking machines." But surely if he was right to invoke literal meaning against claims that, for example, thermostats have beliefs, he is wrong to say humans are machines of any kind. There were literally no machines on this planet 10,000 years ago, whereas the species Homo sapiens has existed here for at least 100,000 years, so it cannot be that men are machines.

Comment:

In-Page Footnotes

Footnote 1:
• The original paper has the moves and PREL scores in tabular form, but I couldn’t be bothered to reformat the copied text.

Text Colour Conventions (see disclaimer)

1. Blue: Text by me; © Theo Todman, 2020
2. Mauve: Text by correspondent(s) or other author(s); © the author(s)