TITLE: Search, Abstractions, and Big Epistemological Questions AUTHOR: Eugene Wallingford DATE: September 11, 2015 3:55 PM DESC: ----- BODY: Andy Soltis is an American grandmaster who writes a monthly column for Chess Life called "Chess to Enjoy". He has also written several good books, both recreational and educational. In his August 2015 column, Soltis talks about a couple of odd ways in which computers interact with humans in the chess world, ways that raise bigger questions about teaching and the nature of knowledge. As most people know, computer programs -- even commodity programs one can buy at the store -- now play chess better than the best human players. Less than twenty years ago, Deep Blue first defeated world champion Garry Kasparov in a single game. A year later, Deep Blue defeated Kasparov in a closely contested six-game match. By 2005, computers were crushing Top Ten players with regularity. These days, world champion Magnus Larson is no match for his chess computer.

a position in which humans see the win, but computers don't

Yet there are still moments where humans shine through. Soltis opens with a story in which two GMs were playing a game the computers thought Black was winning, when suddenly Black resigned. Surprised journalists asked the winner, GM Vassily Ivanchuk, what had happened. It was easy, he said: it only looked like Black was winning. Well beyond the computers' search limits, it was White that had a textbook win. How could the human players see this? Were they searching deeper than the computers? No. They understood the position at a higher level, using abstractions such as "being in the square" and passed pawns like splitting a King like "pants". (We chessplayers are an odd lot.)

When you can define 'flexibility' in 12 bits,
it will go into the program.

Attempts to program computers to play chess using such abstract ideas did not work all that well. Concepts like king safety and piece activity proved difficult to implement in code, but eventually found their way into the programs. More abstract concepts like "flexibility", "initiative", and "harmony" have proven all but impossible to implement. Chess programs got better -- quickly -- when two things happened: (1) programmers began to focus on search, implementing metrics that could be applied rapidly to millions of positions, and (2) computer chips got much, much faster.

The result is that chess programs can beat us by seeing farther down the tree of possibilities than we do. They make moves that surprise us, puzzle us, and even offend our sense of beauty: "Fischer or Tal would have played this move; it is much more elegant." But they win, easily -- except when they don't. Then we explain why, using ideas that express an understanding of the game that even the best chessplaying computers don't seem to have. This points out one of the odd ways computers relate to us in the world of chess. Chess computers crush us all, including grandmasters, using moves we wouldn't make and many of us do not understand. But good chessplayers do understand why moves are good or bad, once they figure it out. As Soltis says:

And we can put the explanation in words. This is why chess teaching is changing in the computer age. A good coach has to be a good translator. His students can get their machine to tell them the best move in any position, but they need words to make sense of it.

Teaching computer science at the university is affected by a similar phenomenon. My students can find on the web code samples to solve any problem they have, but they don't always understand them. This problem existed in the age of the book, too, but the web makes available so much material, often undifferentiated and unexplained, so, so quickly. The inverse of computers making good moves we don't understand brings with it another oddity, one that plays to a different side of our egos. When a chess computer loses -- gasp! -- or fails to understand why a human-selected move is better than the moves it recommends, we explain it using words that make sense of human move. These are, of course, the same words and concepts that fail us most of the time when we are looking for a move to beat the infernal machine. Confirmation bias lives on. Soltis doesn't stop here, though. He realizes that this strange split raises a deeper question:

Maybe it's one that only philosophers care about, but I'll ask it anyway: Are concepts like "flexibility" real? Or are they just artificial constructs, created by and suitable only for feeble, carbon-based minds?

(Philosophers are not the only ones who care. I do. But then, the epistemology course I took in grad school remains one of my two favorite courses ever. The second was cognitive psychology.)

We can implement some of our ideas about chess in programs, and those ideas have helped us create machines we can no longer defeat over the board. But maybe some of our concepts are simply be fictions, "just so" stories we tell ourselves when we feel the need to understand something we really don't. I don't think so, the pragmatist in me keeps pushing for better evidence. Back when I did research in artificial intelligence, I always chafed at the idea of neural networks. They seemed to be a fine model of how our brains worked at the lowest level, but the results they gave did not satisfy me. I couldn't ask them "why?" and receive an answer at the conceptual level at which we humans seem to live. I could not have a conversation with them in words that helped me understand their solutions, or their failures. Now we live in a world of "deep learning", in which Google Translate can do a dandy job of translating a foreign phrase for me but never tell me why it is right, or explain the subtleties of choosing one word instead of another. Add more data, and it translates even better. But I still want the sort of explanation that Ivanchuk gave about his win or the sort of story Soltis can tell about why a computer program only drew a game because it saddled itself with inflexible pawn structure. Perhaps we have reached the limits of my rationality. More likely, though, is that we will keep pushing forward, bringing more human concepts and abstractions within the bounds of what programs can represent, do, and say. Researchers like Douglas Hofstadter continue the search, and I'm glad. There are still plenty of important questions to ask about the nature of knowledge, and computer science is right in the middle of asking and answering them. ~~~~ IMAGE 1. The critical position in Ivanchuk-Jobava, Wijk aan Zee 2015, the game to which Soltis refers in his story. Source: Chess Life, August 2015, Page 17. IMAGE 2. The cover of Andy Soltis's classic Pawn Structure Chess. Source: the book's page at Amazon.com. IMAGE 3. A bust of Aristotle, who confronted Plato's ideas about the nature of ideals. Source: Classical Wisdom Weekly. -----