Describe the three stages of non-REM sleep and compare non-REM sleep w

Topics in Cognitive Science 9 (2017) 485–496 Copyright © 2016 Cognitive Science Society, Inc. All rights reserved. ISSN:1756-8757 print / 1756-8765 online DOI: 10.1111/tops.12232

This article is part of the topic “Game-XP: Action Games as Experimental Paradigms for Cognitive Science,” Wayne D. Gray (Topic Editor). For a full listing of topic papers, see: http://onlinelibrary.wiley.com/doi/10.1111/tops.2017.9.issue-2/issuetoc.

Testing Sleep Consolidation in Skill Learning: A Field Study Using an Online Game

Tom Stafford, a Erwin Haasnoot

b

aDepartment of Psychology, University of Sheffield b Department of Electrical Engineering, Mathematics and Computer Science, University of Twente

Received 6 July 2015; received in revised form 4 April 2016; accepted 11 July 2016

Abstract

Using an observational sample of players of a simple online game (n > 1.2 million), we are able to trace the development of skill in that game. Information on playing time, and player location, allows us

to estimate time of day during which practice took place. We compare those whose breaks in practice

probably contained a night’s sleep and those whose breaks in practice probably did not contain a night’s

sleep. Our analysis confirms experimental evidence showing a benefit of spacing for skill learning, but

it fails to find any additional benefit of sleeping during a break from practice. We discuss reasons why

the well-established phenomenon of sleep consolidation might not manifest in an observational study of

skill development. We put the spacing effect into the context of the other known influences on skill

learning: improvement with practice, and individual differences in initial performance. Analysis of per-

formance data from games allows experimental results to be demonstrated outside of the lab and for

experimental phenomenon to be put in the context of the performance of the whole task.

Keywords: Consolidation; Skill acquisition; Practice; Sleep

1. Introduction

1.1. Consolidation

It is widely accepted that memories are consolidated after acquisition (McGaugh,

2000)—that is, the organization and strength of habits, associations, and skills can

*Correspondence should be sent to Tom Stafford, Department of Psychology, University of Sheffield,

Western Bank, Sheffield, S10 2TP, United Kingdom. E-mail: [email protected]

improve in the gap between acquisition or practice and subsequent testing, even without

active rehearsal. Sleep is thought to be intimately involved in this consolidation process.

A first basic demonstration was by Jenkins and Dallenbach (1924), who showed that

retention of memories of nonsense syllables (following Ebbinghaus, 1885) was less

degraded after a delay which involved sleep rather than a delay of equivalent time which

did not involve sleep. Subsequent results have even shown that, for motor skills, perfor-

mance can improve after a delay involving sleep (e.g., Karni, Tanne, Rubenstein, Aske-

nasy, & Sagi, 1994). More recently, well-controlled experiments have demonstrated that

sleep conveys a crucial benefit, beyond mere disengagement from the task for a compara-

ble delay, and controlling for the known effects of practice spacing (Cohen, Pascual-

Leone, Press, & Robertson, 2005; Walker, Brakefield, Morgan, Hobson, & Stickgold,

2002; Walker et al., 2003).

Although the most consistent evidence for memory consolidation concerns procedural

memories (Stickgold, 2005; Walker & Stickgold, 2004; Walker & Stickgold, 2006), there

are good reasons to suspect this is not a phenomenon restricted to motor skills (Ellenbo-

gen, Hu, Payne, Titone, & Walker, 2007), with there being a complex interaction of sleep

and wakefulness in consolidation and reconsolidation of memories across procedural and

declarative domains (Walker, Brakefield, Hobson, & Stickgold, 2003). Other evidence

suggests that sleep may provide greater benefit for the most difficult aspects of a skill

(Kuriyama, Stickgold, & Walker, 2004).

1.2. Games

Whereas sleep consolidation has been rigorously demonstrated in experiments, it has

been difficult to validate outside the lab. We approach this problem by using a large natu-

rally occurring dataset (Goldstone & Lupyan, 2016) collected from people who play a

simple game of skill online (Stafford & Dewar, 2014).

Previously Stafford and Dewar (2014) have shown that observational data from this

game can be used to validate and extend the analysis of phenomenon previously estab-

lished in the experimental literature on skill acquisition. They show how practice amount

and practice spacing contribute to skill development.

Our interest here is to build on this analysis, using an estimate of the players’ time-

zones. The time-zone of a player, combined with the time of each play, allows us to cal-

culate the local time of each play and so compare comparable practice histories which

are likely to contain, or not to contain, a night’s sleep. This allows us to interrogate our

dataset for the existence of the phenomenon of sleep consolidation. Our study allows us

to use a large sample to quantify the magnitude of the effect as it manifests among those

who are intrinsically motivated to learn an arbitrary task. It also allows us to put the phe-

nomenon within the context of other factors affecting skill development.

The analysis of data from games has particular advantages and disadvantages for the

cognitive scientist. Unlike so many of our experimental tasks, games are played for their

intrinsic enjoyment rather than out of obligation or for external reward (Baldassarre et al.,

2014). This allows us to look at skill development in a context where motivation plays as

486 T. Stafford, E. Haasnoot / Topics in Cognitive Science 9 (2017)

large a part as ability. This supports an expectation of generalization to skill development

outside the lab and avoids the normal confound of large variation in participant motiva-

tion (and the attendant high degree of satisficing which occurs within traditional experi-

ments Maniaci & Rogge, 2014; Oppenheimer, Meyvis, & Davidenko, 2009). Data from

games allow us to measure skill development as it occurs in a naturalistic setting, over

the course of days and weeks, rather than the mere minutes of most typical lab

experiments.

Games also present a skill development domain in which automated data collection at

a large scale is plausible. Unlike other skill development domains—for example, spoken language, playing the violin, soccer—each action taken during a game is conducted through a computer and so may be easily and unobtrusively recorded.

Games involve complex task performance. Further, they contain many elements which

exist to facilitate enjoyment of play, rather than being strictly relevant to the operations

which a cognitive scientist may be interested in. Because of this, the use of games in

cognitive science requires, and will benefit from, analysis of the whole task (as encour-

aged by Newell, 1973).

2. Data acquisition

We used anonymized-at-source data from “Axon,” an online game developed for

the Wellcome Trust by Preloaded. The game can be played at http://axon.well-

comeapps.com/. The game involves guiding a neuron from connection to connection,

through rapid mouse clicks on potential targets. A screenshot can be seen in Fig. 1

(see figure caption for description of game dynamics). Cognitively the game involves

little strategic planning, instead testing rapid perceptual decision making and motor

responding.

The analysis was approved by the University of Sheffield, Department of Psychology

Ethics Sub-Committee, and carried out in accordance with the University and British Psy-

chological Society (BPS) ethics guidelines. The data were collected incidentally and so

did not require any change in the behavior of game players, nor impact on their experi-

ence. Individuals were identified by cookie stored in their browser. For our analysis, we

have assumed a one-to-one mapping between machine and player. No identifying infor-

mation on the players was collected and so the data were effectively anonymized at the

point of collection. Location information was approximate, to the city-block level at max-

imum. For these reasons the institutional review board waived the need for written

informed consent from the participants. For further details of the dataset, see Stafford and

Dewar (2014).

The data were extracted from Google Analytics using a Python library written by Nick

Mihailovski. In contrast to Stafford and Dewar (2014), we were able to extract data for

the longer period of between March 2012 and February 2015. The original data and code

for coding, filtering, and analyzing it are available at https://osf.io/fckq8/.

T. Stafford, E. Haasnoot / Topics in Cognitive Science 9 (2017) 487

This data set comprised a total number of 1,201,515 players, the vast majority of

whom played fewer than five times. The data and code for producing the analysis and

plots presented here are also available from https://osf.io/fckq8/.

3. Analysis 1: Spacing and sleep consolidation

3.1. Aim

Our aim with this analysis was to compare subjects who took a break in their practice

of the game, against those who played a comparable number of games without a break.

This reproduces the analysis done in Stafford and Dewar (2014), which showed the bene-

fits of practice spacing, and extends it to ask if activity during gaps in practice may influ-

ence subsequent performance. To do this, we wish to compare those for whom the timing

suggests that they had probably slept between bouts of practice (e.g., someone who plays

between 8 pm and 9 pm and then again between 8 am and 9 am) against those for whom

the timing suggests that they probably did not sleep between bouts of practice, but

Fig. 1. Game screenshot. Players control the axonal branching of the white neuron. At each point, possible

synaptic contacts (the other dots) are those within the zone of expansion (the larger transparent circle), which

shrinks rapidly after each new contact is made. Non-player neurons (in red here) compete for these synaptic

opportunities. Score is total branch length in micrometers (shown bottom left).

488 T. Stafford, E. Haasnoot / Topics in Cognitive Science 9 (2017)

nevertheless did take a comparable break (e.g., someone who plays between 8 am and 9

am and then again between 8 pm and 9 pm).

3.2. Filtering

First, we only analyze players who complete a minimum of 15 games, leaving 26,727

players. Additionally we filter the data for players on which we are unable to calculate

valid longitude data or valid timing for their practice attempts. This leaves 26,291 players.

3.3. Coding

The local time for each play was calculated using the formula local- time = UTCtime + (longitude 9 24 � 360), modulo 24. This formula gives a local time which is correct in the majority of cases and almost always true within 2 h; the excep-

tions are due to irregularities in time-zone/national borders. Since our location informa-

tion is approximate anyway, there is a limit to the possible level of accuracy regardless

of the method of calculating the local time.

Next, we categorise players into four types, according to the nature of the timing of

their first 15 attempts at the game. Players who play their first 15 games with a gap of

less than 15 min between each game we categorise as “no gap” (9,388 players). Players

who have a single gap of between 7 and 12 h are categorised as resting, either in the

“sleep” or “wake” categories depending on the timing of the gap (761 and 423 players

respectively). A break that finished between 5 am and 12 pm is categorized as a “sleep”

gap (since gaps are 7–12 h, this means that the earliest rising player last played before 10 pm). A break that finished between 5 pm and 12 am is categorized as a “wake” gap. All

other players are categorized as “no category” (15,719 players). This includes people who

have medium length gaps, longer gaps, and multiple gaps.

4. Results

Results are shown in Fig. 2. We show the median scores, not means (inspection of

score distribution showed that there were a small number of very high scores which made

the results—although qualitatively the same—less consistent). The 95% and 99% confidence bounds shown are calculated using a bootstrap analy-

sis: scores from all categories resampled in sample sizes as large as the smaller cate-

gory of the “no gap,” “sleep,” and “wake” categories (for 10,000 iterations). This gives

an indication of how likely it is that samples of these sizes (or larger) would provide

medians outside of the range predicted if the scores for players in these categories were

all drawn from a common distribution. As can be seen, the “no gap” scores fall below

the level predicted by the “no category” scores, and both the “sleep” and “wake” scores

fall above.

T. Stafford, E. Haasnoot / Topics in Cognitive Science 9 (2017) 489

Subtracting the average score for “sleep” category players at each attempt from the

corresponding score for “wake” players shows there is no advantage of the “sleep play-

ers” (indeed, the scores of the “wake” group are slightly, but significantly, higher; differ-

ence = 669.6, t(14) = 2.81, p = .014).

5. Analysis 2: Putting the effects into whole task context

5.1. Aim

Following Newell’s (1973) injunction to study a whole task, we were interested to put

the effect of spacing into the context of other effects which manifest in game perfor-

mance. A disadvantage of observational data is that multiple different factors, both mea-

sured and unmeasured, simultaneously influence outcomes, but a corresponding advantage

is that the data afford the chance to gauge the importance of different factors against each

other. Hence, we ask, having established that the effect of spacing is statistically signifi-

cance, if it is also a meaningful difference.

Secondarily, the quantity of data available makes it possible to analyze in more detail

the functional “shape” of how various factors affect performance. In conventional experi-

mental work we typically compare a small number of points, typically a control and

experimental group, and analyze the contrast to reveal the effect of the manipulated fac-

tor. Here we can show how performance changes with many different levels of the factor.

This “parametric analysis” shows more than just whether a factor has an influence on

Fig. 2. Improvements in performance with practice for those who do not take breaks (“no gap”) and those

who have long breaks, either overnight (“sleep”) or during the daytime (“wake”). Uncategorized players not

shown. Black line shows median for all players and 95% (dashed line with large dots) and 98% (dashed line

with small dots) confidence limits based on samples the size of the smallest of “no gap,” “sleep,” and

“wake.”

490 T. Stafford, E. Haasnoot / Topics in Cognitive Science 9 (2017)

performance, but it has the potential to show something about how a factor influences

performance.

One parametric analysis of the impact on performance that is already familiar is that

of practice, specifically in the form of the learning curve. In this same domain, Stafford

and Dewar (2014) showed that practice amount had the expected effect on performance

of a relatively rapid initial increase which slowed down as practice amount increased (this

can also be seen in the curves shown here in Fig. 2). That analysis also showed that early

performance on the task was predictive of both rate of increase and asymptotic level of

performance. We do not wish to commit on what constitutes these differences between

players—probably it is influenced by a large variety of factors, including motivation, prior experience with online games, sensory-motor function, playing environment, and

equipment as well as neuro-cognitive readiness for skill acquisition.

Here we compare three factors: practice spacing, practice amount, and initial perfor-

mance for both the size and shape of influence over performance. We note that the com-

parison is inherently limited by the arbitrary bounds of the range over which the factors

are analyzed. The effect of practice is bounded by the potential improvement in perfor-

mance due to skill (and hence also by the range over which we assess practice). The

effect of initial performance is bounded by the range within the population from whom

data are gathered. The effect of spacing is bounded by the observed delay between some

initial practice and subsequent attempts. Nonetheless, we believe it is instructive to see

the comparison, and we wish also to highlight it as an example of the way larger data

sets allow different analyses.

5.2. Filtering

As with Analysis 1, we remove all players who played fewer than 15 games, and those

for which we could not calculate longitude or timing information.

5.3. Coding

First, to perform a categorical comparison with which to gauge the size of different

effects, we split our data into high and low groups for each of the three factors we con-

sidered: spacing, practice, and initial performance.

To gauge the effect of spacing, we compared the average score on plays 11–15 for those who had no gap in their first 15 plays (i.e., the “no gap” group from Analysis 1,

n = 9,388), with those who had a single gap of between 7 and 12 h (i.e., the “wake” and “sleep” groups from Analysis 1 combined, n = 1,184). To gauge the effect of practice, we compared the average score, over all players, on plays 1–5 and on plays 11–15. To gauge the effect of initial performance, we compared the average score on plays 11–15 of those who scored in the bottom third on plays 1–5 with the average score on plays 11–15 of those who scored in the top third on plays 1–5.

Second, we sought to make a “parametric” comparison of the effect of changes in

these three factors. By this, we seek to show the way in which average scores change at

T. Stafford, E. Haasnoot / Topics in Cognitive Science 9 (2017) 491

each point along the range for which each factor can change. For practice amount, we

calculated the average score, across all players, for each of the plays numbered 1–15. For initial performance, we calculated the average score on plays 1–5 for range from lowest to highest scorers (using 16 consecutive windows, covering the 100 percentiles). For

spacing we calculated the average score on plays 11–15 according to the total gap time between plays 1 and 10 (using 16 consecutive windows, covering the range 0–60 min. The range was restricted to 0–60 min because average score does not change significantly for larger gaps). We used the median rather than the mean for all averages, since the

score distribution contains a proportion of very high scores, which disproportionately

skews mean scores.

5.4. Results

Fig. 3 shows the effect of the three factors when binary categorized. Fig. 4 shows the

parametric comparison of the three factors. Note that there is no sense in which the range

of the three factors may be compared absolutely. The initial performance line captures all

the variation present in the population, the practice line captures the variation over the

range of number of plays analyzed in this paper (1–15), while the spacing line shows a relatively short range compared to that used for the analysis shown in Figs. 2 and 3. This

is because the spacing effect doesn’t change significantly at cumulative gaps beyond

60 min.

The comparison of Figs. 3 and 4 illustrates that effects which appear to be of a compa-

rable size from a “two point” analysis can be produced by underlying functions which

have very different shapes. Practice affects performance with a decelerating function; ini-

tial performance has the opposite effect, such that the largest changes come at the high-

Fig. 3. Two-category comparison for the effects of spacing, practice, and initial performance. Standard error

bars are shown.

492 T. Stafford, E. Haasnoot / Topics in Cognitive Science 9 (2017)

end of the distribution of that variable. The effect of spacing is a non-monotonic function,

with an optimal point in the middle of the range (presumably reflecting a trade-off

between the memory benefits of spacing-based consolidation and the memory costs of

forgetting).

6. Discussion

These analyses show that there is a clear spacing effect. The psychological mecha-

nisms by which this is produced may be assumed to be some combination of rest/recov-

ery and active consolidation of memory. Analysis 1 suggests that, contrary to

experimental results, breaks in training which contain sleep do not provide a superior

benefit to equally long breaks which do not contain sleep. There could be many reasons

for this. One possibility is that our task and/or analysis is insensitive to any additional

effect of sleep consolidation. Although our large data set suggests this would not be due

to a lack of statistical power, it might be that the nature of our task, or the ranges over

which we conducted our analysis, fall outside the operating realm of the effect (in con-

trast to experimental results, which we might presume are carefully designed to capture

the effect). If this is so, it is interesting to note that, whereas other learning phenomena

such as practice or spacing effects do manifest, sleep consolidation does not here.

Other results suggest that the benefit of sleep consolidation is larger for more complex

tasks (Ellenbogen et al., 2007; Kuriyama et al., 2004). It may be that our task was not

complex enough for a sleep consolidation effect to manifest. Fig. 4 could be viewed as

lending support to this idea—there is no additional benefit on performance of gaps longer than 15 min, with the spacing effect appearing as gap in practice lengthens from no gap

to 15 min. This is a relatively short window compared to the size of many spacing effects

Fig. 4. Parametric comparison for the effects of least to most spacing, shortest to longest practice, and low-

est to highest initial performance. Standard error bars shown for practice and spacing curves.

T. Stafford, E. Haasnoot / Topics in Cognitive Science 9 (2017) 493

(Cepeda et al., 2009) and compared to the duration over which benefits of sleep consoli-

dation are typically seen.

The lack of experimental control over players’ behavior may be involved in the failure

to observe sleep consolidation. Suppose that the phenomenon operates in concert with

some other factor such as fatigue and amount of information needing consolidation. 1 Indi-

vidual players may automatically calibrate their practice so that they are resting as and

when they need to with respect to these factors, so that there is no additional benefit of

sleep consolidation. In contrast, experimental studies dictate when participants practice

and when they rest, which both controls for spacing effects and which may allow a bene-

fit of sleep consolidation to be isolated.

It is striking that the benefit that comes from spaced practice is comparable to the ben-

efit of players tripling their amount of practice (Fig. 3). Both of these effects are

swamped by the range in aptitude for the game, as measured by initial performance (this

importance of initial aptitude has been found elsewhere; Destefano, 2010; Huang, Yan,

Cheung, Nagappan, & Zimmermann, in press; Stafford & Dewar, 2014). Two important

caveats are, first, that although the amount and nature of our practice can be brought

under an individual’s control, it is less clear how initial performance can be controlled.

This means that while differences in acquisition due to initial performance may be larger,

it is not clear that they are more important for anyone wishing to infer how to improve

rate of acquisition. Secondly, in this study we define aptitude entirely phenomenologi-

cally—that is, it is a simple effect read off from the data by dividing players according to their initial scores. Although this shows how players vary in the initial scores, it leaves

completely unexplored why players vary. No doubt a constellation of factors contribute to initial ability, some of which are indeed mutable (for discussion of the contribution of

initial ability to expertise development, see Detterman, 2014).

Games offer an opportunity to investigate learning in a naturalistic context, under con-

ditions of intrinsic motivation, as well as bringing with them the advantages of easy col-

lection of large data sets. We attempted to show here how one particular game can be

used to study long established phenomenon. In particular we show ordered effects of

practice amount, and a predicted effect of practice spacing, in a simple game. In con-

trast, the predicted benefit of rest periods that involved sleep was not observed. We also

attempted to put these effects into mutual context, contrasting both the “size” of the

effect—admittedly with arbitrarily defined ranges—and the parametric “shape” of the effects. In this way we hoped to show that the large amount of data available in the

study of games does not just augment statistical power but makes possible new ways of

analyzing behavioral data.

Acknowledgments

We are very grateful for the feedback and discussion provided by Wayne Gray, Walter

Boot, and two anonymous reviewers.

494 T. Stafford, E. Haasnoot / Topics in Cognitive Science 9 (2017)

Note

1. Although we note that Stafford and Dewer (2014, Fig. 4) provide evidence for a

true consolidation effect in these data, and not just a “relief from fatigue” effect

References

Baldassarre, G., Stafford, T., Mirolli, M., Redgrave, P., Ryan, R. M., & Barto, A. (2014). Intrinsic

motivations and open-ended development in animals, humans, and robots: An overview. Frontiers in Psychology, 5.

Cepeda, N. J., Coburn, N., Rohrer, D., Wixted, J. T., Mozer, M. C., & Pashler, H. (2009). Optimizing

distributed practice: Theoretical analysis and practical implications. Experimental Psychology, 56(4), 236–246.

Cohen, D. A., Pascual-Leone, A., Press, D. Z., & Robertson, E. M. (2005). Off-line learning of motor skill

memory: a double dissociation of goal and movement. Proceedings of the National Academy of Sciences of the United States of America, 102(50), 18237–18241.

Destefano, M. (2010). The mechanics of multitasking: The choreography of perception, action, and cognition

over 7.05 orders of magnitude. Unpublished doctoral dissertation, Rensselaer Polytechnic Institute.

Detterman, D. K. (2014). Introduction to the intelligence special issue on the development of expertise: Is

ability necessary? Intelligence, 45, 1–5. Ebbinghaus, H. (1885). €Uber das ged€achtnis: untersuchungen zur experimentellen psychologie. Leipzig:

Duncker & Humblot.

Ellenbogen, J. M., Hu, P. T., Payne, J. D., Titone, D., & Walker, M. P. (2007). Human relational memory

requires time and sleep. Proceedings of the National Academy of Sciences, 104(18), 7723–7728. Goldstone, R. L., & Lupyan, G. (2016). Discovering psychological principles by mining naturally occurring

data sets. Topics in Cognitive Science, 8, 548–568. doi:10.1111/tops.12212 Huang, J., Yan, E., Cheung, G., Nagappan, N., & Zimmermann, T. (in press). Master maker: Understanding

gaming skill through practice and habit from gameplay behavior. Topics in Cognitive Science. Jenkins, J. G., & Dallenbach, K. M. (1924). Obliviscence during sleep and waking. The American Journal of

Psychology, 35, 605–612. Karni, A., Tanne, D., Rubenstein, B. S., Askenasy, J., & Sagi, D. (1994). Dependence on rem sleep of

overnight improvement of a perceptual skill. Science, 265(5172), 679–682. Kuriyama, K., Stickgold, R., & Walker, M. P. (2004). Sleepdependent learning and motor-skill complexity.

Learning & Memory, 11(6), 705–713. Maniaci, M. R., & Rogge, R. D. (2014). Caring about carelessness: Participant inattention and its effects on

research. Journal of Research in Personality, 48, 61–83. McGaugh, J. L. (2000). Memory–A century of consolidation. Science, 287(5451), 248–251. Newell, A. (1973). You can’t play 20 questions with nature and win: Projective comments on the papers of

this symposium. In W. G. Chase (Ed.), Visual information processing (pp. 283–308). New York: Academic Press.

Oppenheimer, D. M., Meyvis, T., & Davidenko, N. (2009). Instructional manipulation checks: Detecting

satisficing to increase statistical power. Journal of Experimental Social Psychology, 45(4), 867–872. Stafford, T., & Dewar, M. (2014). Tracing the trajectory of skill le

Collepals.com Plagiarism Free Papers

Are you looking for custom essay writing service or even dissertation writing services? Just request for our write my paper service, and we'll match you with the best essay writer in your subject! With an exceptional team of professional academic experts in a wide range of subjects, we can guarantee you an unrivaled quality of custom-written papers.

Get ZERO PLAGIARISM, HUMAN WRITTEN ESSAYS

Why Hire Collepals.com writers to do your paper?

Quality- We are experienced and have access to ample research materials.

We write plagiarism Free Content

Confidential- We never share or sell your personal information to third parties.

Support-Chat with us today! We are always waiting to answer all your questions.