Expertise and skill: the power law, deliberate practice and the limits of transfer

A screen that is easy in the first hour can be slow in the thousandth. This unit argues that it is a claim with evidence behind it, and that the evidence is narrower. The strongest counter is that practice explains less of performance than the argument supposes, accounting for between 26 and just over zero per cent of the variance across five domains (Macnamara et al., 2014). Skill comes from sustained practice inside one domain. The units before it described a mind with fixed limits; this one asks how practice changes them. Practice changes what a user perceives, what they hold in mind, and what attention a step costs. Those changes are specific to the practised task, so one screen cannot serve the first hour and the thousandth equally.

The improvement curve is a description, not a law

Newell and Rosenbloom (1981) gave the field the description it still quotes. The time a task takes falls as a power function of practice. It improves quickly at first and more slowly afterwards, with chunking as the proposed mechanism. Anderson (1982) added a fuller one. Facts about a domain are first interpreted declaratively and then compiled into procedures through composition and proceduralization. He argued that this produces the power-law speedup.

Heathcote, Brown and Mewhort (2000) fitted power and exponential functions to 40 data sets. Those covered 7,910 learning series from 475 participants in 24 experiments. The exponential fitted better in all of the unaveraged data. Averaging had biased the fit towards the power function. Improvement is fast then slow, and that shape is what a designer can rely on.

Work the curve on one task. A user enters records into a new system. The first record takes two minutes, the tenth under one, and the hundredth about forty seconds. The case shows most of the gain arrives early. It does not license a number, since no fit was made here.

What practice builds is knowledge of one domain

Chase and Simon (1973) showed what practice builds. Three players of different strength rebuilt positions seen for five seconds. On a middle-game position the master placed about sixteen pieces correctly on the first attempt, the class A player about eight and the beginner about four. On random positions from the same pieces all three managed about three or four. The master was seeing structure that the beginner was not.

Gobet and Simon (1996) revised the mechanism. With up to five positions shown for five seconds each, recall fell with every added board, but it stayed above the original chunking prediction. One master reached nine boards at better than 70 per cent accuracy, explained by templates held in long-term memory. Chi, Feltovich and Glaser (1981) showed that the change is in the representation. Sorting physics problems, experts grouped them by the principle that would solve them, and novices by the objects named.

A display organised around surface features matches a novice's reading and is noise to an expert. Design rule: choose the surface deliberately.

The practice thesis overreaches its arithmetic

Ericsson, Krampe and Tesch-Roemer (1993) made practice the centre of the explanation. Violinists estimated the hours of practice alone accumulated in each year of study. By age 18 the best had 7,410 hours, the good violinists 5,301 and the music teachers 3,420. The authors concluded that deliberate practice is necessary for expert performance and sufficient to account for it. They defined it narrowly, as structured activity with immediate feedback. The definition builds in a learner with a teacher and the hours to spend, which is not the general case.

Macnamara, Hambrick and Oswald (2014) tested the strong reading. Their meta-analysis found that deliberate practice explained 26 per cent of performance variance in games, 21 per cent in music and 18 per cent in sports. It explained 4 per cent in education and under 1 per cent in professions. The two sides differ in their evidence: the sufficiency claim rests on recounted hours, which cannot separate cause from consequence, while the counter rests on pooled correlations. The expertise a designer can depend on is expertise in this system.

One screen, two users, and one reversal

Fitts (1964) described the change the curve measures. Perceptual-motor skill learning runs through three phases. In the cognitive phase the learner works out what to do. In the associative phase the components are linked and the wrong responses suppressed. In the autonomous phase the skill runs fast with little attention. Attention that a step consumes at the start is released later. That release is the resource an expert has for something else.

Sweller (1988) supplied the mechanism that makes the release matter. Conventional problem solving consumes working-memory capacity that would otherwise build schemas. Sweller, van Merrienboer and Paas (1998) set this in a cognitive architecture of limited working memory and schema construction. Kalyuga, Ayres, Chandler and Sweller (2003) reviewed the experiments that give the unit its central claim. Support that helps novices can lose its value and harm experienced learners once it becomes redundant.

Work the reversal on one interface. A dashboard ships with a step-by-step panel that names each filter. First-week users follow it and complete the task correctly. An analyst who filters daily has read the panel hundreds of times. On Kalyuga's evidence the same support is now redundant, and the panel that guided the novice competes for the attention the data needs. The case does not come from a study that faded guidance in a working interface, which no source here reports.

Design rule: support should be present, then fadable, then gone.

Expertise is bound to its domain, and can block a better answer

Sala and Gobet (2017) set the boundary. Meta-analyses of chess instruction, music instruction and working-memory training found small to moderate effects on children's cognitive and academic skills, smaller as study design improved. The case shows why an interface cannot be eased by improving the user in general. It does not show that no transfer exists, only that the measured effects are small and shrink as the method improves.

A second limit is that the schema itself can obstruct. Bilalic, McLeod and Gobet (2008) recorded the eye movements of expert chess players facing a familiar but inferior solution and a better, less familiar one. Every player reported searching for the better solution once the first was found. Their gaze stayed on the squares of the familiar one. The schema that familiar features activate directs attention towards what is consistent with itself, without the expert noticing, so a fast familiar route can make an alternative invisible.

Contested question: how much does practice explain, and is the curve a law?

Two disagreements run through the unit. Ericsson and colleagues (1993) hold that deliberate practice is a sufficient account of expert performance. The mechanism literature supports them by showing that practice changes memory and representation in ways that would produce skill (Newell and Rosenbloom, 1981; Anderson, 1982). Macnamara and colleagues (2014) hold that the effect is real but modest and varies by domain. Sala and Gobet (2017) add that what is learned transfers little beyond the practised domain. The settling evidence would be prospective: measure hours as learning proceeds, randomise assignment to practice regimes, and use active control groups and a pre-specified effect size.

The second concerns the curve. The textbook position treats the power function as a law (Newell and Rosenbloom, 1981). Heathcote and colleagues (2000) argue that it is an artefact of averaging, and that individual learning is better described by an exponential. The settling study fits both model classes to unaveraged data from individual learners. It reports each fit with its free parameters, and compares the rivals by a nested-model test rather than by the shape of an averaged plot.

The expertise reversal rests on experiments with instructional materials (Kalyuga et al., 2003), and the cognitive-load framework predicts it (Sweller, 1988). No study cited here, however, fades guidance in a working interface for users who have outgrown it. The test is an interaction, not two main effects. Novices and experts meet the same interface, with and without support, and the study reports the size of the interaction.

While the dispute is open, a designer can act on what both readings allow. Expect most of the gain early, and do not assume it continues. Make support fadable rather than fixed, so that a user who has outgrown it can set it aside, because that is the one design move the evidence supports without the missing interaction study.

Where the argument is heading

The practice debate has moved from whether practice matters to how much and where (Ericsson et al., 1993, to Macnamara et al., 2014). The curve literature has moved from the shape of an average to the behaviour of an individual learner, tested by model comparison (Heathcote et al., 2000). The thread that has not moved far enough is design, where the surviving claim is that guidance should be adapted as expertise grows (Kalyuga et al., 2003), and the open problem is measuring how much support a user still needs.

The sources this unit rests on

• Anderson, J. R. (1982). Acquisition of cognitive skill. Psychological Review, 89(4), 369-406. doi:10.1037/0033-295x.89.4.369 Proposed that facts about a domain are first interpreted declaratively and then compiled into procedures by composition and proceduralization, a mechanism he argued produces the power-law speedup. It is the unit's mechanism for the curve.
• Bilalic, M., McLeod, P. and Gobet, F. (2008). Why good thoughts block better ones: the mechanism of the pernicious Einstellung (set) effect. Cognition, 108(3), 652-661. doi:10.1016/j.cognition.2008.05.005 Recorded expert chess players' eye movements as they faced a familiar but inferior solution and a better one. Players said they searched for the better solution while their gaze stayed on the familiar one. Expertise can block a better answer unnoticed.
• Chase, W. G. and Simon, H. A. (1973). Perception in chess. Cognitive Psychology, 4(1), 55-81. doi:10.1016/0010-0285(73)90004-2 Had three players of different strength rebuild positions seen for five seconds. The master replaced about sixteen pieces on a middle-game position, the class A player about eight and the beginner about four, and on random positions all three managed about three or four. Expertise is chess knowledge, not memory.
• Chi, M. T. H., Feltovich, P. J. and Glaser, R. (1981). Categorization and representation of physics problems by experts and novices. Cognitive Science, 5(2), 121-152. doi:10.1207/s15516709cog0502_2 Had experts and novices sort physics problems: experts grouped them by the principle that would solve them, novices by the objects named. Practice changes a problem's representation.
• Ericsson, K. A., Krampe, R. T. and Tesch-Roemer, C. (1993). The role of deliberate practice in the acquisition of expert performance. Psychological Review, 100(3), 363-406. doi:10.1037/0033-295x.100.3.363 Asked violinists to estimate their accumulated practice; by age 18 the best had 7,410 hours, the good violinists 5,301 and the music teachers 3,420. It states the strong practice thesis the unit then qualifies.
• Fitts, P. M. (1964). Perceptual-motor skill learning. Categories of Human Learning (Melton, A. W., ed.), Academic Press, 243-285. doi:10.1016/b978-1-4832-3145-7.50016-9 Set out three phases of perceptual-motor skill learning: a cognitive phase in which the learner works out what to do, an associative phase in which the components are linked and the wrong responses suppressed, and an autonomous phase in which the skill runs fast with little attention. The phases name the change to design across.
• Gobet, F. and Simon, H. A. (1996). Templates in chess memory: a mechanism for recalling several boards. Cognitive Psychology, 31(1), 1-40. doi:10.1006/cogp.1996.0011 Had players from class A to grandmaster recall up to five positions shown for five seconds each. Recall fell with every added board but stayed above the original chunking prediction, and one master reached nine boards at better than 70 per cent accuracy, explained by templates in long-term memory.
• Heathcote, A., Brown, S. and Mewhort, D. J. K. (2000). The power law repealed: the case for an exponential law of practice. Psychonomic Bulletin and Review, 7(2), 185-207. doi:10.3758/bf03212979 Fitted power and exponential functions to 40 data sets covering 7,910 learning series from 475 participants in 24 experiments, and the exponential fitted better in all unaveraged data. Averaging biased the fit towards the power function.
• Kalyuga, S., Ayres, P., Chandler, P. and Sweller, J. (2003). The expertise reversal effect. Educational Psychologist, 38(1), 23-31. doi:10.1207/s15326985ep3801_4 Reviewed experiments giving the same instructional support to learners of different levels: support that helps novices can lose its value and harm experienced learners once it becomes redundant. It is the direct evidence that one screen must serve two users.
• Macnamara, B. N., Hambrick, D. Z. and Oswald, F. L. (2014). Deliberate practice and performance in music, games, sports, education, and professions: a meta-analysis. Psychological Science, 25(8), 1608-1618. doi:10.1177/0956797614535810 Pooled studies from five domains and found deliberate practice explained 26 per cent of performance variance in games, 21 per cent in music, 18 per cent in sports, 4 per cent in education and under 1 per cent in professions. It makes the practice thesis arguable.
• Newell, A. and Rosenbloom, P. S. (1981). Mechanisms of skill acquisition and the law of practice. Cognitive Skills and Their Acquisition (Anderson, J. R., ed.), Lawrence Erlbaum, 1-55. doi:10.4324/9780203728178-6 Proposed that performance improves as a power function of practice and offered chunking as the mechanism, drawing on tasks that improve quickly at first and more slowly later. It is the description the field treated as settled.
• Sala, G. and Gobet, F. (2017). Does far transfer exist? Negative evidence from chess, music, and working memory training. Current Directions in Psychological Science, 26(6), 515-520. doi:10.1177/0963721417712760 Reported meta-analyses of chess instruction, music instruction and working-memory training: effects on children's cognitive and academic skills were small to moderate and smaller as study design improved. It closes the route of easing an interface by improving the user in general.
• Sweller, J. (1988). Cognitive load during problem solving: effects on learning. Cognitive Science, 12(2), 257-285. doi:10.1207/s15516709cog1202_4 Argued that conventional problem solving, and means-ends analysis in particular, requires processing capacity that is then unavailable for schema acquisition. It supplies the mechanism behind the expertise reversal.
• Sweller, J., van Merrienboer, J. J. G. and Paas, F. G. W. C. (1998). Cognitive architecture and instructional design. Educational Psychology Review, 10(3), 251-296. doi:10.1023/a:1022193728205 Set out the cognitive-load framework, connecting limited working memory, schema construction in long-term memory and design decisions that add or remove avoidable processing. It states the position behind this unit's fading-guidance argument.