24 October 2010

More on Computer Assistance

While I'm on the subject of computer assistance in chess960 (see my previous post, Advanced Chess960 @ Chess.com) here are a few more points of view. In Fischer Announces Fischerandom, a portion of the press release announcing Fischer's version touched on computer analysis.
With many people wondering about the future of chess after the IBM computer Big Blue beat Garry Kasparov earlier this year, Fischer's statement that computers would be at a considerable disadvantage in Fischerandom Chess received a great deal of attention. He stated that without access to databases of the millions of opening variations in traditional chess, computers do not really play chess all that well.

As an aside, don't read too much into the sentence that 'IBM computer Big Blue beat Garry Kasparov earlier this year'. The press conference took place in June 1996, so the Kasparov - IBM match would have been the first, played in February 1996. Kasparov suffered his only loss in the first game, but won the match by winning the last two games. It wasn't until the following year that Deep Blue beat Kasparov in a match (see a page on my World Championship site, Kasparov vs. IBM's Deep Blue, for a few details about the two matches).

Fischer's thoughts on computers drew attention from the experts. In The birth of Fischer Random Chess by Eric van Reem, which first appeared in 2001, the well known chess journalist wrote,

Fischer stated, that without access to databases of the millions of opening variations in traditional chess, computers do not really play chess all that well. However, Matthias Wüllenweber, one of the founders of ChessBase, has a completely dífferent opinion on that subject. Last year, when "Fritz on Primergy" played two Shuffle Chess games against German number 1 Artur Jusupov [Yusupov], the software specialist said: "When playing F.R Chess unusual patterns come up on the board. Knowledge of these patterns, however, is one of the main weapons for humans in their battle against computers.

Wüllenweber refers to a test his partner Frederic Friedel did with Hungarian Grandmaster Andras Adorjan. Friedel showed Adorjan several positions for a period of ten seconds. The Hungarian could recall those "normal" postions far better than amateur players did. Humans remember so-called "chunks" e.g. they do not remember pawn on f2, g2 and h2, King on g1 and Rook on f1, they remember the chunk "Castling Kingside".

If you build up a position without those patterns, but try to put up a position that really doesn´t make sense, with pawns on the first and eigth rank for example, there is hardly any difference in memorization between amateurs and grandmasters. According to Wüllenweber this 'thinking in chunks' is the main difference between humans and computers and the difference in ELO is some hundreds of points. A computer can play with 3 knights or 5 rooks, no problem.

This last point is in accordance with a sentiment I reported in A Few Novel Ideas (with links to the ICCF.com forum for the context of the discussion).

"For serious correspondence chess, as opposed to casual correspondence chess, playing chess960 games is a step in the WRONG direction. The reason is simple : the human knowledge effect in the games will be further reduced since the engines that are already affecting classic correspondence chess have zero problems adapting to chess960."

So who is right, Fischer or his critics? Since Fischer's heyday occurred when computer chess was still in its infancy (the first World Championship for Computer Chess took place in 1974), it's easy to conclude that the experts who came later were far more knowledgeable. Fischer, however, surprised us many times in the past and he might well surprise us in the future, long after his death.

1 comment:

HarryO said...

Hi there
Yes these are some interesting thoughts. For sure we humans use chunks to remember ever deeper and deeper patterns. The point is that chunk memory does not just extend to patterns of the pieces on the board, but also to chunks of "principals". Chess960 shatters the framework of stable "principals" that we come to know in SP518. We go from a few chunks of principals in SP518, to literally hundreds of principals in Chess960.

In Chess960 these "principals" are actually "exception" principals, where we have a firm principal in mind that causes our memory to form "chunks" of exception principals that apply to that "principal". There are many relationships that are formed in our mind and this is the heart of the "chunking" memory framework of Chess960.

The issue is that we are all Chess players. We have habituated to a different familiarity of "chunking" memory. It will be the next generation of dedicated Chess960 kids that will think in "principals" chunking methods. It is exactly the same process. Ultimately our memory is object oriented but objects exist as properties and methods that are chunked together through a process of familiarity.

So we see a chunk of pawns clustered in our SP518 familiarity. That chunk of pawns is actually an object in our mind that has properties and methods. It would be no different if we were to form relationships of invisible opening principals and exceptions in our mind for all 960 positions in Chess960.

I suggest to Chess960 players out there to do what I am currently doing. I am gradually playing through all Chess960 SP's starting from position 1, 2, 3 and so on consecutively. When you do this patiently and diligently, the small changes from SP to SP become like variations on a theme. The experience is like listening to Shostakovitch's 24 preludes and fugues or Bach's Goldberg variations. That is what it is like to play Chess960.

We tend to play Chess960 positions randomly. This is necessary for competition but for training it is precisely the wrong thing to do. We must work through consecutive subset's of positions. When we do this, we begin to instinctively feel the way that clusters of SP's relate to each other. It is wrong to study one SP in isolation. This is a bad habit from Chess SP518 thinking.

For example I am working through positions 73 to 79 at the moment. I call these the "Bat" positions because over the last few days "Bats" have been flying over my house in the thousands depositing seeds on my roof like hail as I play them. The "Bat" positions are characterized by two undefended pawns on A2 and B2. They are creepy tactical positions that are especially good for tactical training on knight movements.

As I play through them making mistake after mistake, I am learning principals and exceptions to those principals that will apply to other batches of SP's down the track. The structure in my mind is invisible but it is no different to seeing visible structures of pawns in SP518. I have an older mind and so this process of familiarity works more slowly than for unfettered minds.

These batches of consecutive Chess960 SP's will be the heart of how future generations learn how to play Chess960. These batches of SP's have moods and colours associated with them in our mind. It is like variations on a theme in music and is an indescribable "feeling".

It is just a process of familiarity. We would be comfortably be beating Chess960 computers (only at classical time controls) had we had a generation of kids through the system playing Chess960 from childhood, because they would be better at manipulating and extending patterns of "principals and exceptions" in their mind in this era of technology since Bobby, so long as the silicon is not allowed to refer to an opening database. In effect, Bobby was absolutely correct. That is my belief.