SquirrelChess.com - How many patterns do you have to memorize to become a master?

How many patterns do you have to memorize to become a master?

Written by Sciurus

Friday, 03 November 2006

This is part 2 of the chess tactics psychology miniseries based on "The Psychology of Chess Skill" by Dennis H. Holding (and the last part because I had to return the book to the library). The first part discussed the concepts of "chunking" and "pattern recognition". In summary, chess positions are memorized by grouping the pieces together in so-called chunks of typically 2-3 pieces. Move decisions are based on the association of several chunks with pattern also stored in long-term memory. One of the main differences between chess masters and amateurs seems to be the number and size of chunks and pattern available in long-term memory. An aspiring chess beginner might therefore ask "how many patterns do I have to learn to become a master?"


Photo by Vincent van Dam

If chunks are indeed used to reduce the amount of memory needed to store a chess position in memory then one would expect that the same or at least similar chunks are used to encode a position when it is shown a second time. Experiments with both chess masters and amateurs show that this is indeed the case. Moreover, the same chunks tend to get re-used when a group of pieces is just moved to a different position on the board. This greatly reduces the number of different chunks needed to encode an arbitrary position, particularly if the same chunk can be used for both white and black pieces.

Several studies used computer programs to simulate the chunking of chess positions. If the program was given a new position, it would first try to match it with chunks it had already learned. After that, the program would try to sort the remaining parts of the position in new chunks based on piece types, defense relationship, etc. Interestingly, the majority of the chunks recognized by the computer programs were identical to those used by chess masters to recall the same position.

The success rate of a computer program for recalling a given chess position can then be used to estimate the number of chunks needed to achieve the same skill level as chess masters. However, these methods are based on too many assumptions and chess is too complex to obtain a fixed number. Furthermore, the behavior of a computer program knowing only a few hundred chunks that probably wouldn't be able to compete with human novices is extrapolated to the level of a chess master. Therefore, most studies conclude that something between 10000 and 100000 different chunks have to be memorized to perform on the same level as a chess master. At the first look, these numbers may look daunting. However, they are in the same order of magnitude as the number of english words available to a competent speaker and mastery of both a language and chess requires many years of practice. And even grandmasters are not perfect. A factor of 5 to 10 times as many chunks is estimated to be necessary for perfect performance.

The numbers cited above are based on the ability of chess masters to recall a given chess position that has been shown to them for a few seconds. It should be kept in mind that being able to recall a position is not the same as deciding which move is correct (see also the difference between chunks and pattern in part one of this miniseries). The recognition of tactical pattern that trigger a move decision is probably based on the spatial relationship between the chunks of pieces present on the board. While reading Dennis Holding's book, I came the belief that it is likely that chess mastery requires several layers of chunking and pattern recognition and that tactical ability requires more than just recognizing a simple group of pieces within a chess position. Using analogy between learning chess and a language: Becoming fluent in a language requires more than just learning the vocabulary (which might be the equivalent of chunks in chess). Fluent speakers assemble whole sentences and decipher them "on the fly" without conciously thinking about them similar to chess masters making the right moves even in complex situations.

The most important practical insight of this research for learning to play chess is that chess mastery requires to store a collection of chunks and pattern in long-term memory which can be asociated with the positions during games. The problem is, however, that nobody really knows how exactly the chunks we need to know to play good chess look like - otherwise we could "simply" try to memorize them directly. Moreover, most of the research concentrates on explaining the fascinating skills of master-level chess players and does not give a specific optimum study plan for beginners who want to improve quickly. Reading about scientific explanations on how our brain works was interesting to me. To learn play better chess, however, I still have to rely on playing many many games to get these mysterious chunks into my head.

Comments (4) Add Comment

Subscribe to this comment's feed

Show/Hide comments

How did computers locate chunks?
written by Thomas Dybdahl Ahle, November 09, 2006

You wrote that

it would first try to match it with chunks it had already learned. After that, the program would try to sort the remaining parts of the position in new chunks based on piece types, defense relationship

Did the book describe these algorithms in more details?

It would be interesting to write a chessengine, that plays chess the same way as humans.

Re: How did computers locate chunks?
written by Sciurus, November 09, 2006

Thomas:

good question, I was wondering about that, too. Unfortunately, the book did not give many details on the way the computer was used to determine which pieces are grouped together in one chunk. There seem to be relationships between pieces that make it likely that they get chunked together. Examples for that are close proximity and pieces that attack/defend each other. They cited the original work done on this but I had to return the book to the library and cannot access it anymore. Therefore, I cannot even provide you with the references - sorry.

I think I read somewhere that people actually tried to program chess engines that resemble human thinking in the sense of pattern recognition. As far as I remember (and this might be incorrect), even the best of these programs were much weaker than programs that more or less calculate variations by brute force. Seems that the human brain is after all very special and pretty hard to emulate