Professional Documents
Culture Documents
Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at
http://www.jstor.org/page/info/about/policies/terms.jsp. JSTOR's Terms and Conditions of Use provides, in part, that unless
you have obtained prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you
may use content in the JSTOR archive only for your personal, non-commercial use.
Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at
http://www.jstor.org/action/showPublisher?publisherCode=lebtaylorfrancis.
Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed
page of such transmission.
JSTOR is a not-for-profit organization founded in 1995 to build trusted digital archives for scholarship. We work with the
scholarly community to preserve their work and the materials they rely upon, and to build a common research platform that
promotes the discovery and use of these resources. For more information about JSTOR, please contact support@jstor.org.
Lawrence Erlbaum Associates (Taylor & Francis Group) is collaborating with JSTOR to digitize, preserve and
extend access to Cognition and Instruction.
http://www.jstor.org
COGNITION ANDINSTRUCTION, 1990,7(2),129-159
? 1990,Lawrence
Copyright Erlbaum Inc.
Associates,
DiscoveryLearningand Transfer
of Problem-SolvingSkills
Mark A. McDaniel
PurdueUniversity
Mark S. Schlager
SRI International
For the past three decades, psychologists and educators have engaged in
considerable debate concerning whether or not learning by discovery pro-
duces efficacious learning outcomes. Many educators have argued that dis-
covery learning produces better learning than expository learning (e.g.,
IN SOLVING
COMPONENTSINVOLVED
KNOWLEDGE
PROBLEMS
TRANSFORMATION
We distinguishamongtwo knowledgecomponentsinvolvedin skilledsolv-
ing of transformationproblemsand a thirdcomponentinvolvedin generat-
DISCOVERYLEARNING 131
EMPIRICALPREDICTIONS
EXPERIMENT1
Method
GOAL
FIGURE 1 Problem space of legal moves for the hobbits-orcs (5,3) problem. The let-
ter H stands for hobbit, O stands for orc, and the symbol "< > " is used to represent
the boat; the spatial location of the symbols represents the bank of the river (eft or
right) on which the objects are located. The symbols placed between the states indicate
the move that transforms one state into another state.
136
DISCOVERY LEARNING 137
Results
Training. For all analyses,the rejectionlevel was set at .05. Perform-
ance on the trainingproblemcan be simply summarized.As increasingly
more information about the solution strategywas providedto subjects
(from no informationin the SMD conditionto detailedinformationin the
ND condition),performanceimproved(see Table1 for means).This obser-
vation was supportedwith analysesof variance(ANOVAs)of trialsto last
error(i.e., numberof trials before three consecutiveminimum-pathsolu-
tions), legal moves, and illegal moves (for legal and illegal moves, mixed
ANOVAsthat includedthe first five trialsas a factorwereconducted).Sig-
nificanteffects of instructionwerefound for all threedependentmeasures:
For trials to last error, F(2, 87) = 4.20, MSE = 10.56; for legal moves, F(2,
87) = 12.94, MSE = 178.99; for illegal moves, F(2, 87) = 6.17, MSE = 9.14.
Also, performanceimprovedsignificantlyacross trials: For legal moves,
F(4, 348) = 25.59, MSE= 110.65; for illegal moves, F(4, 438) =43.42,
MSE= 6.47. Finally, significant Groupx Trials interactions-for legal
moves, F(8, 348) = 3.65, MSE = 110.65; for illegal moves, F(8, 348) = 3.19,
MSE = 6.47-indicated that differencesbetween the instructionalgroups
werereducedor eliminatedas subjectscompletedmoretrials.
Following Jeffries's (1978) technique,we identified subjects' solution
strategiesaccordingto the movesequencesthat the subjectsused in the ini-
tial part of the problem.We classifiedthe move sequencesinto one of two
basic strategies:a balancestrategyand an OSO strategy.The balancestrat-
egy was one in whichsubjectstriedto keepthe numberof hobbitsand orcs
as even as possibleon both sides of the river.This strategyresultedin sub-
jects movingfrom States5-7-9-12 (see Figure1). Also classifiedas balance
strategieswerethose instancesin which subjects'initial movesto the right
bankinvolvedmovingorcs into State3, afterwhichsubjectsmovedhobbits
to balance the orcs and continued trying to maintainthis balance with
movesequencesfrom States3-7-9 or 12, or 3-6-9 or 12. None of thesebal-
DISCOVERYLEARNING 139
TABLE1
Performanceon the FirstFiveTrainingTrialsas a Function
of InstructionalConditionin Experiment1
InstructionalCondition
Training7rial ND MD SMD
Trial1
Legalmoves 17.5 26.1 35.7
Illegalmoves 2.3 4.3 6.3
PercentageusingOSO 89.7 53.3 29.0
Thial2
Legalmoves 14.6 17.7 26.3
Illegalmoves 0.7 0.9 2.1
PercentageusingOSO 82.8 63.3 41.9
Trial3
Legalmoves 13.2 13.9 18.9
Illegalmoves 0.3 0.3 0.5
PercentageusingOSO 82.8 70.0 48.4
Trial4
Legalmoves 12.5 13.7 13.6
Illegalmoves 0.1 0.2 0.3
PercentageusingOSO 93.1 73.3 48.4
Trial5
Legalmoves 12.5 12.3 14.4
Illegalmoves 0.0 0.1 0.4
PercentageusingOSO 93.1 83.3 67.7
Numberof trialsbefore 2.4 4.0 4.8
minimum-pathsolution
Note. Subjectswho met criterionbeforeTrial5 wereassignedminimum-path
performance
for subsequenttrialsratherthanbeingdroppedfromthe analyses.
Discussion
The equivalenttransferfor SMD instructionand MD instructionwas con-
sistent with the second predictionderivedfrom the frameworkpresented
earlierin this article. This framework,however,would also anticipatethat
subjectswho had to discovera move sequenceto instantiatethe requisite
strategy(i.e., SMD and MD conditions)would be better able to generate
new moves to apply the strategyon the transferproblem,relativeto sub-
jects who weregiven a move sequenceduringtraining(i.e., the ND condi-
tion). This hypothesiswas not supported.
The most strikingfinding was the good transferperformanceof all of
the groups,especiallyin light of the relativelylong retentionintervaland a
transferproblemwith surfacefeaturessomewhatdissimilarto the training
problems.Transferbetweenproblemisomorphsis characteristicallydiffi-
cult to obtain (e.g., Catrambone& Holyoak, 1989;Reedet al., 1974).Ac-
cordingly,a central challenge to problem-solvingtheorists has been to
explain the obstaclesto positive transfer(e.g., Bassok & Holyoak, 1989;
Novick, 1988;Reedet al., 1974;Stein, Way,Benningfield,& Hedgecough,
1986). One salient differencebetweenmost of the extantresearchand the
presentwork is the amount of exposuregiven to the trainingproblem.In
most other studies, subjectswereinstructedto solve the trainingproblem
DISCOVERYLEARNING 141
TABLE2
Trialsin Experiment
on FirstTwoTransfer
Performance 1
InstructionalCondition
Trial1
Legalmoves 12.7 12.9 13.2
Illegalmoves 1.0 1.0 1.1
Strategyclassification(in percent)
OSO 17 13 15
Paired 41 40 39
Mixed 24 30 26
Trial2
Legalmoves 12.6 11.1 11.4
Illegalmoves 1.0 0.5 0.4
Strategyclassification(in percent)
OSO 21 50 32
Paired 28 30 29
Mixed 28 20 26
only once before they attempted to solve a transfer problem. In the present
experiment, subjects continued their attempts to solve the training problem
until they performed the task perfectly. Perhaps one important ingredient
for effective transfer of a strategy or schema (Catrambone & Holyoak,
1989; Sweller & Cooper, 1985) is complete learning of the solution strategy
during training. Apparently, the degree to which training involves discovery
of the strategy is irrelevantas long as the strategy is well learned and as long
as the learned strategy is applicable for the transfer problem.
EXPERIMENT2
2Traininga balance (i.e., a paired strategy) strategy and transferring to a problem solved
only with OSO seems, at first glance, reasonable. Novices, however, are overwhelmingly biased
toward using a balance strategy (see McDaniel, 1980). Thus, one would not expect that explic-
itly providing subjects with the balance strategy would produce any differences in acquisition
relative to the SMD group, which was not explicitly provided with the balance strategy.
DISCOVERYLEARNING 143
Method
Design. The design was 2 x 2 x 2 between-subjectsfactorialincluding
the variablestype of instruction(MD, SMD), type of transferproblem
(near,remote),and test interval(immediate,48-hrdelay).All subjectswere
trainedon 16 three-jarwater-jarproblemspartitionedinto four sets; each
set exemplifieda differentstrategy.The first set containedfour problems
that were governedby a strategyof filling the largestjar, subtractingthe
volume of anotherjar twice, and adding the volume of the remainingjar
(1 - 2 + 1). The second set of four problemswas governedby a strategyof
filling the largestjar, subtractinganotherjar three times, and subtracting
the remaining jar twice (1 - 3 - 2 or 1 - 2 - 3). The strategy for the third
set involved filling the largestjar, subtractinganotherjar once, and sub-
tracting the remaining jar twice (1 - 1 - 2 or 1 - 2 - 1). Finally, the fourth
set was governedby a strategyof fillingthe largestjar, subtractinganother
jar twice,and addingthe remainingjar four times (1 - 2 + 4; see Appendix
B for the completeproblemset).
The MD subjectswereinformedthat they would be shown a strategyat
the beginningof each set of trainingproblemsthat would help them solve
all problemsin that set. The strategiesfor solving these problemstook the
form of mathematicalexpressionsin whichthe jar labelswereleft out. For
example,the strategy1 - 2 + 1 could be appliedto a problemin which Jar
C is the largestand Jar A is the smallest, Jar B the largestand Jar C the
smallest, and so on. Thus, these subjects had to discoverthe particular
moves to implementthe strategy.Subjectsweretold to take their time be-
fore startingthe first problemin orderto memorizethe strategy,becauseit
wouldbe presentedonly once. An exampleof the mathematicalformof the
strategies(i.e., as displayedearlier)was also provided.
The SMD subjectswerenot given the strategiesbeforeeach set but, in-
stead, wereinformedthat each problemin a set must be solved using the
samegeneralstrategy.Theyweretold the strategywouldbe a moreabstract
form of the specific equationsrepresentingthe solution of each problem,
and they werealso told to try to discoverthe strategyfor each set, because
it would help them solve the problemsmore easily.
Two types of transferproblemswere employed. One type represented
144 McDANIEL AND SCHLAGER
for failure to complete the training phase, 4 were lost in the control condi-
tion for failure to solve both problems, 3 were terminated by computer fail-
ure, and 6 were excluded for other reasons (1 due to illness, 3 chose not to
continue, and 2 due to extreme fatigue).
Resultsand Discussion
Acquisition. For all analyses,the rejectionlevel was set at .05. Mean
solution times and total movesare given in Table3 in parentheses.The to-
tal-movemeasurereflects the sum of the legal moves, illegal moves (i.e.,
movesresultingin states that violate the generalwater-jarproblemrestric-
tions, e.g., overflowinga jar), and restartmoves (a commandto eraseall
previousmovesfrom the screenand startthe problemover).Measurements
of solutiontime werecollectedbecausethe simpleadd and subtractnature
of the movesin this domainmay haveallowedsubjectsto covertlyconsider
and rejectmovesthat werenot enteredinto the computer.Presumablyany
such covertactivitywouldbe reflectedin solutiontime.
Three-way2 (InstructionalLevels)x 4 (TrainingSets) x 4 (Problemsin
Each Set) mixed ANOVAswereperformedon log-normalizedtransforma-
tions of each of the just-mentionedperformancemeasures.Thereweretwo
significanteffects (for each dependentmeasure)involvingthe instructional
variable,both of which were interactions.A two-wayinteractionbetween
instruction and the problem number in each set (i.e., first problem, second
problem, etc.) indicatedthat SMD subjectsrequiredmore time and took
more movesthan MD subjectsprimarilyfor the first problemin each set,
F(3, 354) = 28.95, MSE = .05, and F(3, 354) = 4.20, MSE = .06, respec-
tively. A significant three-wayinteraction(among instruction, problem
number,and set number)furtherindicatedthat this resultwas limited to
Problem Sets 3 and 4, F(9, 1062)= 10.02, MSE= .05, and F(9,
1062)= 2.77, MSE= .06, for time and moves, respectively.Indeedfor the
first problemin ProblemSet 1, SMD subjectswerenominallyfaster and
took slightlyfewermovesthan MD subjects.
Solution times and moves significantlyvariedas a functionof problem
set, F(3, 354) = 565.03, MSE = .07, and F(3, 354) = 163.25, MSE = .09, re-
spectively.This effect was due in part to the fact that the numberof moves
for minimum-pathsolutionvariedacrossproblemsets. The particularpat-
tern of this effect, however,suggeststhat an additionalcomponentwas in-
volved(presumablya learningcomponent).Subjectsspentlongerand took
moremoveson Set 1 problemsthan any otherset, despitethe fact that Set 1
problemscould be solved in just four moves(vs. six moves for Set 2, four
for Set 3, and sevenfor Set 4).
DISCOVERYLEARNING 147
TABLE3
TotalNumberof Movesand SolutionTime for Each Training
Problemas a Functionof Instructionin Experiment2
Moves Timea
Set I
There was also a significant main effect of problem number for solution
time and moves, F(3, 354) = 335.23, MSE = .05, and F(3, 354) = 37.85,
MSE = .06, respectively, such that subjects experienced more difficulty on
the initial problems (in a set) than the later problems, in general. This effect
significantly interacted with problem set for time and moves, indicating
that the difference between initial and later problems was obtained almost
exclusively in Set 1, F(9, 1062) = 39.04, MSE= .05, and F(9, 1062)
= 18.79, MSE = .06, respectively.
These results suggest that, on the first problem set, the MD subjects were
148 McDANIELAND SCHLAGER
Transfer. The mean solutiontimes and the mean numberof moves for
each transferproblemare presentedin Table4. The data werefirst submit-
ted to three-waybetween-subjects analysesof covariance(ANCOVAs),with
instruction(SMD, MD), type of transferproblem(near,remote),and test
interval (immediate,delayed) as the independentvariables. These AN-
COVAswere performedon log-normalizedtransformationsof the total-
move and solution-time measures. Total solution time for Training
Problems1, 5, 9, and 13 was covariedwith the transfer-performance vari-
ables to reducethe possibilitythat any transfereffects would be obscured
becauseof individualdifferencesin problem-solvingskill amongsubjectsin
the experimentalgroups.
Comparedwith remote-transferproblems,near-transferproblemswere
solved in significantly fewer moves (9.0 vs. 19.0), F(1, 228) = 31.99,
MSE = .20, and requiredsignificantlyless time to solve (116.7 sec vs. 282.1
sec), F(1, 228) = 41.09, MSE = .21. This advantagefor near-transferprob-
lems was expected,giventhat the near-transferproblemsweremore similar
to the trainingproblemsthan the remote-transfer problems.The near-trans-
fer problemsalso appearedto be somewhateasierthan the remote-transfer
problems,regardlessof prior training(see Table4 for the control group
means).Therewas a significantmaineffect of instructionfor solutiontime,
with SMD instruction producing faster solutions than MD instruction
(163.8 sec vs. 235.0 sec), F(1, 228) = 7.50, MSE = .21. The SMD instruc-
DISCOVERYLEARNING 149
TABLE4
TotalNumberof Movesand SolutionTimeas a Functionof Type
of TransferProblem,Instruction,and Time of Test in Experiment2
Moves Time"
Transfer
Instruction7Tpe Immediate Delay Mean Immediate Delay Mean
Near
Three-jar
MD 13.5 22.8 18.2 187.5 247.2 217.4
SMD 8.9 14.6 11.8 139.4 220.1 179.8
Control 26.7 941.4
TWo-jar
MD 3.3 2.7 3.0 39.0 34.8 36.9
SMD 2.5 3.3 2.9 30.6 34.8 32.7
Control 4.1 105.3
Remote
Three-jar
MD 24.3 33.2 28.8 378.5 552.0 465.2
SMD 25.1 24.9 25.0 279.6 406.6 343.1
Control 38.0 1115.8
TWo-jar
MD 16.8 13.4 15.1 233.8 207.4 220.6
SMD 8.8 5.4 7.1 94.5 104.4 99.4
Control 4.8 167.5
tions also tended to produce fewer total moves than MD instruction (11.7
vs. 16.3), F(1, 228) = 3.64, p < .06, MSE = .20.
Although the instructional variable did not interact significantly with the
transfer-problem variable, inspection of the means (see Table 4) reveals that
the magnitude of the difference between the instructional groups was
greater for the remote-transferproblems, especially for solution time. These
possible interactions between instruction and problem type may not have
reached statistical significance because of the nonnormal distributions of
scores created by forming conceptually similar groups (near and remote
transfer) from problems that varied dramatically in difficulty. Table 4 gives
an indication of how greatly performance varied between the two problems
in each transfer category; the difference in performance probably resulted
in an inflated error term. This observation motivated separate two-way
(instruction and test delay) between-subjects ANCOVAs for each problem.
These analyses directly test the expectations outlined at the outset of the
experiment.
For both near-transfer problems, as expected, there were no main effects
of instruction or interactions between instruction and delay for either of
the dependent variables (Fs < 1). (Upon inspecting Table 4 with regard to
150 McDANIELAND SCHLAGER
Theoretical implications. The key finding was that SMD training was
advantageous, relative to MD training, when the transfer task required use
of a strategy that was different from that used to solve training problems.
As mentioned earlier, one interpretation of this finding is that, during
training, the SMD subjects had to utilize processes for generating and test-
ing alternative strategies, whereas the MD subjects did not. Having to use
these general processes during training apparently practices, refines, and/or
develops them so that they can be used more efficaciously in later novel
problems within the domain. Other research has shown that active hypothe-
sis-testing processes can play a role in learning solution strategies for trans-
formation-type problems (Lewis & Anderson, 1985), and it may be that
differential hypothesis-testing processes (across the instructional groups)
were responsible for the present transfer results.
Alternatively, differential transfer for the SMD versus the MD subjects
might have been due to differential acquisition of particular solution strate-
gies, which may have guided selection of particular move sequences, during
training that were common to the training and transfer problems. At first
glance, it might appear that this common elements explanation would not
be viable in the present context, because the solution strategies for the re-
mote-transfer problems differed explicitly from the strategies governing the
training problems. However, the solution protocols indicated that, during
training, subjects entertained specific erroneous strategies that produced
dead ends. Although these strategies produced dead ends in the training
problem, they still could be useful for the remote-transfer problems. That
is, solving a remote-transfer problem might be accomplished by applying a
strategy that produced a dead end in the training problems. Further, it is
possible that such deadend strategies, common to those required to solve
the remote-transfer problems, were tried more frequently by SMD subjects
during training. If so, enhanced remote transfer by SMD subjects might
have been due to direct application of these common strategies to the trans-
fer problems.
To distinguish between these two explanations, we selected an easily de-
tected strategy that was ineffective in solving the training problems but was
common to the strategy required by the two-jar remote-transfer problem.
The strategy involved simply adding the contents of the medium and small
jars together. We classified subjects who displayed this strategy during
training as "add" subjects, and subjects who did not as "no-add" sub-
jects.4 The first important finding was that both SMD and MD subjects in
failed, and began to look for new strategies. The SMD subjects were faster
than MD subjects at developing or generating alternative strategies.
GENERAL DISCUSSION
Educational Implications
We conclude with some reflections on the educational implications of the
present research. Clearly, the increased effort and training time associated
with discovery modes of instruction will not always be beneficial. Specifi-
cally, discovery learning per se does not seem to provide subjects with more
insight into the information targeted for learning (e.g., a particular strat-
egy). Once learned, the information (e.g., a strategy) seems to be transfer-
red with equal facility, regardless of whether or not it was discovered or
explicated for the learner. More particularly, having to discover problem-
relevant information (e.g., a domain-specific strategy) does not appear to
produce better generalizability of that strategy to other problems in the do-
main for which the strategy can be applied (see also, Guthrie, 1967; Norton,
Graham, & Merrill, 1977). And discovery learning per se does not seem to
enable the learner to identify more quickly the one useful strategy from a
repertoire of several different strategies acquired in training (e.g., near-
transfer performance in Experiment 2).
DISCOVERYLEARNING 155
ACKNOWLEDGMENTS
REFERENCES
APPENDIX A
Orcs-Switch-Orcs
This strategycan be describedas:
Move wives either for four moves or until a boatload (or nearlyfull
boatload) of husbands can be brought across, move husbands to
achievea balance,returna husband-wifepair,movethe remainderof
the husbands,movethe rest of the orcs across.
Balance
This "paired"strategymerelyinvolvestransportingtwo pairs of husbands
and wives acrossthe river,returninga pair, and continuingin like fashion
until the problemis solved.This will not producea minimum-pathsolution
for the transferproblemused.
Orcs-Switch-Orcs-Balance
This "mixed"strategyis a hybridof the OSO and the balanceapproaches.
The movesrepresentingthis strategyare like those for OSO until the inter-
change,then husband-wifepairs are moved(as in the balancestrategy)to
finish the problem;or husband-wifepairsare moveduntil the interchange,
then an OSO sequenceis used to finish the problem.
DISCOVERYLEARNING 159
APPENDIXB
Experiment2: ExperimentalTaskSet
JarA JarB Jar C Goal Formula
ProblemSet
Set 1 123 24 10 85 1A - 2B + 1C
33 89 34 54 1B - 2C + 1A
73 20 21 51 1A-2C+ 1B
7 36 111 46 1C-2B+ 1A
Set 2 3 108 23 33 1B - 3C - 2A
136 32 3 34 1A- 3B - 2C
21 11 122 37 1C-3A-2B
15 158 11 91 1B- 3A- 2C
Set 3 7 24 98 60 1C - 1B - 2A
78 10 21 37 1A- 1C- 2B
93 40 3 47 1A - 1B - 2C
28 138 3 104 1B - 1A- 2C
Set 4 97 23 4 67 1A - 2B + 4C
43 91 18 77 1B-2A + 4C
13 46 174 134 1C- 2B + 4A
129 3 35 71 1A - 2C + 4B
7ypeof TansferProblem
Three-jar
Near 117 22 18 59 1A - 1B - 2C
Remote 15 48 166 124 1C - 6A + 1B
Two-jar
Near 67 32 129 97 1C- 1B
Remote 34 117 8 42 1A + 1C