Memory researchers love chess. Chess experts have been studied by psychologists to try to explain the role of higher-level processing in memory. One reason chess is a great domain in which to ask these questions is because its rules and concepts provide meaning to the spatial positions of the pieces only if one knows them. As a non-chess example, imagine trying to memorize this string of numbers “31415926535” compared to this one “9201844341.” The first is much easier, but only if you know the digits of pi. If you know pi, all you have to remember is “the digits of pi” – one mnemonic element. On the other hand, one must recall all 10 digits separately for the second string because there is no underlying integrative feature.
We’ve discussed research on chess experts on this blog before, but to summarize: chess players have an advantage in remembering the positions of pieces on the board because they can “chunk” sets of pieces into meaningful groups. As a result, experts recall more pieces compared to novices from a board that could occur in an actual chess game but not from a board that consists of randomly-placed pieces. More recent research has shown that chess experts also have a small advantage in random configurations, attributed to the small amount of non-random information that is accidentally contained in those otherwise random configurations.
Still, there is a debate about how chess experts gain this advantage. One line of reasoning is that chess experts have templates: familiar patterns of chess pieces, which they can call to mind. This so-called template theory has been modeled using a computational architecture called CHREST. To model expert chess memory, the computational model learns templates from viewing a large set of chess positions. Later, when asked to recall pieces on a board, the model can determine an appropriate template, and put individual pieces into the appropriate slots. These templates are thus learned and used by being viewed a number of times and forming a familiar pattern.
But in a recent paper in the Psychonomic Society’s journal Memory & Cognition, David Lane and Yu-Hsuan Chang argue that familiar patterns alone cannot account for chess experts’ memory advantage. The researchers contend that high-level strategic chess knowledge, which does not result in similar patterns across different game scenarios, also accounts for a portion of chess experts’ advantage.
Here’s an example. Take a look at the two boards below. To my eye (full disclosure: I am a moderately competent chess player), there is no discernible pattern that is similar across these two boards. But experts may recognize that both are examples of a “bad bishop,” or a bishop on the same color as several pawns (on the left, black has a bad black bishop; on the right, black has a bad white bishop). The bishop is bad, by the way, because it is blocked behind pawns without operational space and resembling a rather big pawn.
This high-level structure can result in several seemingly dissimilar patterns, but still organizes the information on the board.
Lane and Chang argue that the high-level conceptual knowledge chess players possess results in a form of memory for chess that cannot be captured by modeling familiar patterns. In other words, understanding the bad-bishop rule results in the ability to attribute meaning to a huge variety of chess positions, including those with unfamiliar patterns.
To test this, Lane and Chang administered a chess memory task (look at a chess board, recall the positions of pieces), a chess knowledge test (consisting of questions about chess openings and strategy), a chess experience survey (estimate amount of time playing chess), and non-chess measures of fluid intelligence to 79 chess players. Replicating past work, chess experience correlated with chess memory. But, in support of the role of conceptual knowledge, chess knowledge correlated with chess memory after controlling for chess experience. In other words, a player could have played 1,000 games of chess, but never learned what a bad bishop is. That player would have worse memory performance than a player who played 1,000 games, but did learn about bad bishops.
This is shown in the figure below, which plots chess memory scores (y-axis) as a function of chess knowledge scores (x-axis), controlling for chess experience. The figure reveals the association between high-level conceptual knowledge about chess and memory for chess positions, independent of how many chess positions one has actually viewed.
These findings relate to a field of memory research about schemas, which can help structure the information in the world into a meaningful organization. But are these schemas the result of encountered patterns or the result of a conceptual understanding of chess? What distinguishes the computational model, CHREST, from Lane and Chang’s approach is that CHREST only uses encountered patterns, whereas Lane and Chang argue that knowledge of the rules and strategy of chess lead to generalization beyond what has been seen before.
This research also helps to expose the limits of models based on memorization to explain human cognition. While CHREST does a good job explaining the bulk of human behavioral performance, because it doesn’t know chess, it can’t exploit patterns of meaning (like bad bishops) that don’t correlate with patterns of pieces.
Beyond chess, it is crucial to know whether human memory extrapolates from known patterns or can generalize from higher-level information. Even in models of vision, the role of top-down processing is debated, and studying games like chess might play a key role in resolving the dispute.
Reference for the article discussed in this post:
Lane, D.M. & Chang, Y-H., A. (2017). Chess knowledge predicts chess memory even after controlling for chess experience: Evidence for the role of high-level processes. Memory & Cognition. DOI: 10.3758/s13421-017-0768-2.