Artificial Intelligence Review, (1990) 4, 251-277

I nt el l i gent Tutoring Systems: an

overvi ew
Hyacinth S. Nwana
Depart ment of Comput er Science, Uni versi t y of Li verpool ,
Li verpool L69 3BX, UK
Abstract, This is a non-expert overview of Intelligent Tutoring Systems
(ITSs), a way in which Artificial Intelligence (AI) techniques are being
applied to education. It introduces ITSs and the motivation for them. It
looks at its history: its evolution from Computer-Assisted Instruction
(CAI). After looking at the structure of a ' typical' ITS, the paper further
examines and discusses some other architectures. Several classic ITSs are
reviewed, mainly due to their historical significance or because they best
demonstrate some of the principles of intelligent tutoring. A reasonably
representative list of ITSs is also provided in order to provide a better
appreciation of this vibrant field as well as reveal the scope of existing
tutors. The paper concludes, perhaps more appropriately, with some of the
author' s viewpoints on a couple of controversial issues in the intelligent
tutoring domain.
1 Ar t i f i c i a l I n t e l l i g e n c e a n d E d u c a t i o n
The i ncor por at i on of Art i fi ci al I nt el l i gence (AI) t echni ques i nt o educat i on in or der
to pr oduc e educat i onal l y usef ul c omput e r art efact s dat es back to t he ear l y 1970s. By
t he ear l y 1980s r esear cher s in t he al r eady vi br ant fi el d had cl ear l y spl i t i nt o t wo
unequal camps wi t h t he emer gence of t wo school s of t hought . The first and smal l er
of t he t wo gr oups advocat ed ' expl or at i on e n v i r o n me n t s ' - - e nvi r onme nt s whi ch
encour age di s cover y l ear ni ng (i.e. l ear ni ng by doing). Per haps t he mos t f amous is
t he LOGO l anguage (Papert , 1980) whi ch i nt r oduces st udent s to t he wor l d of
geomet r y t hr ough t he use of robot ' t ur t l es ' and ' t ur t l e gr aphi cs ' t echni ques, i.e. t he
s t udent l earns by di r ect pr ogr a mmi ng r at her t han by i ndi r ect i nst r uct i on. Paper t
(1980) proj ect s t hat ' c omput e r pr es ence wi l l enabl e us to modi f y t he l ear ni ng
e nvi r onme nt out si de t he cl as s r oom so t hat much, if not all, of t he knowl edge
school s pr es ent l y t ry to t each wi t h such pai n, expens e and l i mi t ed success, wi l l be
l ear ned as t he chi l d l ear ns to t al k, pai nl essl y, successf ul l y and wi t hout i nst r uct i on. '
He goes on to concl ude t hat ' s chool s as we know t hem t oday have no pl ace in t he
f ut ur e. ' Cl earl y, Paper t ' s dr eam is qui t e r evol ut i onar y: hence, he and hi s advocat es
in t he LOGO c a mp are oft en referred to as r evol ut i onar i es.
251
252 H. S. Nwana
The second and larger of t he camps is t he "intelligent t ut ori ng' group who refer to
t hemsel ves as reformi st s as t hey prefer a gradual i mpr ovement (i.e. evol ut i on) in t he
pr esent qual i t y of educat i on usi ng AI t echni ques. They advocat e a par adi gm wher e
t he comput er acts as a t ut or, i.e. st udent s largely l earn by bei ng told.
Nat ural l y, t he appr oach Paper t champi ons was bound to be greet ed wi t h con-
si derabl e scorn and cr i t i ci sm as it pr oposes radi cal l y to change t he status quo. It
appear s i nconcei vabl e t hat such a r api d change is feasible even if it wer e t hought
desi rabl e. On t he ot her hand, t he i nt el l i gent t ut ori ng appr oach enj oys t he pri vi l ege
of bei ng cl oser to cur r ent t radi t i onal cl assroom i nst ruct i on. As a resul t , t he LOGO
camp, per haps unf or t unat el y, is now oft en br acket ed wi t h its much larger intelli-
gent t ut or i ng count er par t . In any case, this paper over vi ews t he latter (i.e. intelli-
gent t ut or i ng systems).
2 Introduction to i ntel l i gent tutoring systems
Int el l i gent t ut or i ng syst ems (ITSs) are comput er programs t hat are desi gned to
i ncor por at e t echni ques from t he AI communi t y in or der to pr ovi de t ut ors whi ch
know what t hey t each, who t hey t each and how to t each it. AI at t empt s to pr oduce
in a comput er behavi our whi ch, if per f or med by a human, woul d be descr i bed as
' i nt el l i gent ' : ITSs may si mi l arl y be t hought of as at t empt s to pr oduce in a comput er
behavi our whi ch, if per f or med by a human, woul d be descr i bed as ' goad t eachi ng'
(Elsom-Cook, 1987). The desi gn and devel opment of such t ut ors lie at t he i nt ersec-
t i on of comput er sci ence, cogni t i ve psychol ogy and educat i onal research; this
i nt er sect i ng area is nor mal l y referred to as cogni t i ve sci ence (see Fig. 1). For hi st or-
ical reasons, muc h of t he r esear ch in t he domai n of educat i onal soft ware i nvol vi ng
AI has been conduct ed in t he name of ' ICAI' , an acr onym for ' Int el l i gent Comput er -
Ai ded I nst r uct i on' . Thi s phrase, in t urn, evol ved out of t he name ' Comput er - Ai ded
I nst r uct i on' (CAI) oft en referri ng to t he use of comput er s in educat i on. Never t he-
less, to all i nt ent s and pur poses, ITSs and ICAI are s ynonymous . However , t hough
some r esear cher s still prefer 'ICAI' (e.g. Self, 1988a, uses it in t he title of his r ecent
book), it is now oft en r epl aced by t he acr onym ' ITS' (Sl eeman & Brown, 1982b). The
latter, whi ch is also t he aut hor ' s per sonal preference, is cer t ai nl y gai ni ng suppor t ,
as conf i r med by t he i nt er nat i onal conf er ence on Int el l i gent Tut or i ng Syst ems hel d
in Mont real , Canada, as r ecent l y as June 1988. Thi s pr ef er ence is mot i vat ed by t he
cl ai m that, in many ways, t he si gni fi cance of t he shift in r esear ch met hodol ogy goes
beyond t he addi ng of an T to CAI (Wenger, 1987). However , some r esear cher s are
under s t andabl y hesi t ant to use t he t er m ' i nt el l i gent ' , i nst ead opt i ng for labels such
as ' Knowl edge- Based Tut or i ng System' (KBTS) or ' Adapt i ve Tut or i ng Syst em'
(ATS) (e.g. Streitz, 1988): Wenger (1987) prefers t he label Knowl edge Communi ca-
t i on Syst ems. Nevert hel ess, most researchers appear to be r easonabl y cont ent wi t h
t he acr onym ITS. Thi s is fine as long as ever yone i nvol ved wi t h t he area under -
st ands t hat t he usage of t he wor d ' i nt el l i gent ' is, st ri ct l y speaki ng, a mi snomer . Thi s
does not appear to be t he case, resul t i ng in some ver y ambi t i ous goals/claims,
par t i cul ar l y in t he mor e t heoret i cal parts of the l i t erat ure: this also appear s to be a
val i d cri t i ci sm of t he ent i re AI l i t erat ure.
Intelligent Tutoring Systems: an overview 253
The fact t hat ITS r esear ch spans t hr ee di fferent di sci pl i nes has i mpor t ant i m-
pl i cat i ons. It means t hat t her e are maj or di f f er ences in r esear ch goals, t er mi nol ogy,
t heor et i cal f r amewor ks , and e mpha s e s amongst ITS r esear cher s. Thi s wi l l be c ome
a ppa r e nt l at er in t hi s paper . ITS r esear ch al so requi res a mut ual under s t andi ng of
t he t hr ee di sci pl i nes i nvol ved, a ver y st ressful de ma nd gi ven t he pr obl ems of
keepi ng abreast wi t h even a si ngl e di sci pl i ne t oday.
However , s ome r esear cher s have st ood up to t he chal l enge. As a result, a great
deal has been l ear nt about how to desi gn and i mpl e me nt ITSs. A numbe r of
i mpr es s i ve ITSs descr i bed in t hi s paper bear t es t i mony to t hi s fact.
sci ence
Fig. 1. ITS domains.
3 Mo t i v a t i o n
Why do r esear cher s bot her to pr oduce such comput er - bas ed t ut ors? Ther e s eem to
be t wo ma i n mot i vat i ng fact ors.
(1) Research needs. On t he pur e r esear ch l evel , t here is a need to under s t and mor e
about t he pr ocesses whi ch cont r i but e to an educat i onal i nt er act i on (El som-Cook,
1987). Si nce ITS r esear ch lies at t he i nt er sect i on of t hr ee mai n di sci pl i nes, it
pr ovi des an excel l ent t est - bed for var i ous t heor i es f r om cogni t i ve psychol ogi st s. AI
sci ent i st s and educat i onal t heori st s. For exampl e, a pr i mar y r eason why t he f amous
Car negi e Mel l on psychol ogi st John Ander s on came i nt o t he area, was to t est out hi s
var i ous t heor i es of l ear ni ng ( Ander son, 1987). Hence, t he desi gn of an ITS wi l l
cont r i but e to t he di s cover y of mor e accur at e t heori es of cogni t i on (Burns & Capps,
1988).
(2) Practical needs. On t he mor e appl i ed l evel , t here are a numbe r of usef ul r esul t s
whi ch can be achi eved usi ng ITSs whi ch cannot be achi eved wi t h h u ma n t ut or s for
economi c and soci al r easons (El som-Cook, 1987). A pr i ma r y advant age of ITSs is
t he possi bi l i t y for pr ovi di ng one- t o- one t ut ori ng. Ther e is a cons ens us on t he vi ew
t hat i ndi vi dual t ui t i on, t ai l or ed to t he needs of t he st udent , is t he mos t effect i ve
f or m of educat i onal i nt er act i on, at l east for most domai ns. Bl oom (1984) in hi s
c ompa r i s on of pr i vat e t ut or i ng wi t h cl as s r oom i nst r uct i on of car t ogr aphy and
254 H. S. Nwana
pr obabi l i t y f ound t hat 98% of t he st udent s wi t h pri vat e t ut ors per f or med bet t er t han
t he average cl assr oom st udent , even t hough all st udent s spent t he same amount of
t i me l earni ng t he t opi cs. Ander s on et al. (1985a, b) also r ecor ded a f our - t o- one
advant age for t he pri vat e t ut or, as measur ed by t he amount of t i me for st udent s to
get to t he same l evel of pr of i ci ency. Si nce our educat i onal syst ems have, of necessi -
ty, become geared t owar ds group t eachi ng, many of t he advant ages of one- t o- one
t ut or i ng have been lost. ITSs can pr ovi de such t ui t i on wi t hout necessar i l y losing
t he advant ages of t he group t eachi ng envi r onment (e.g. by pr ovi di ng one ITS per
st udent in a class), t her eby getting t he best of bot h worl ds. The ITS coul d pr ovi de
i mmedi at e f eedback to t he st udent on t he task bei ng per f or med. Thi s i ndi vi dual -
i zed and i mmedi at e f eedback is cruci al because t ut or i ng is most effect i ve when
occur r i ng in di rect r esponse to t he need of t he st udent .
4 Hi s t ori cal r e v i e w
4.1 Introduction to revi ew
Comput er - assi st ed i nst r uct i on/ l ear ni ng (CAI/CAL) has evol ved consi der abl y si nce
its i ncept i on in t he 1950s wi t h Ski nner i an t ype ' l i near programs' . Thi s has hap-
pened despi t e bei ng set off in t he wr ong di r ect i on by Ski nner ' s i nsi st ence t hat
st udent s' r esponses coul d be i gnored in l i near programs (O' Shea & Self, 1983). The
cent r al pr obl em wi t h earl y syst ems was that t hey wer e unabl e to pr ovi de ri ch
f eedback or i ndi vi dual i zat i on, because t hey were not desi gned to know what t hey
wer e t eachi ng, who t hey wer e t eachi ng or how to t each it. In or der to sol ve t hi s
pr obl em, CAI/CAL syst ems have evol ved over t he past t hree and a hal f decades i nt o
what are now usual l y t er med ' Int el l i gent Tut or i ng Syst ems' (ITSs). Al t hough we
may still be far f r om t r ul y i nt el l i gent t ut ori ng systems, most woul d agree consi der -
able progress has al r eady been made.
4.2 From CAI to ITSs: major stages
Ther e wer e some maj or stages in t he met amor phosi s of t he l i near programs of t he
1950s into t he ITSs of t he 1980s (see Fig. 2). The pat h has s panned a per i od of
al most f our decades. It began in t he 1950s wi t h si mpl e ' l i near pr ogr ams' whi ch
wer e based on t he pr i nci pl e of oper ant condi t i oni ng. The mai n pr oponent of such
l i near programs was t he psychol ogi st B. F. Ski nner (1954, 1958). Mat eri al whi ch
had been sel ect ed and arranged to take t he st udent step by step t owar ds t he desi r ed
behavi our was pr es ent ed in a series of ' frames' . Most frames had ver y si mpl e
quest i ons (e.g. i nvol vi ng onl y t he filling in of a mi ssi ng space or two), and t he
st udent was t ol d i mmedi at el y whet her t he answer was ri ght or wrong. The syst em
pr oceeded to pr esent t he next frame regardl ess of t he correct ness of t he st udent ' s
response. To be fair, Ski nner hel d t hat st udent s shoul d not be al l owed to make
mi st akes because this gives negat i ve r ei nf or cement . If t he desi gner s ucceeded i n
this aim, all t he r esponses woul d be correct and so coul d l egi t i mat el y be i gnored.
Unf or t unat el y, exper i ence s howed t hat such an i deal si t uat i on was usual l y not
at t ai nabl e. The maj or l i mi t at i ons of l i near programs t hen became gl ari ngl y appa-
I nt el l i gent Tutoring Sy s t e ms : an overview 255
rent: they did not provide individualization, which meant that all students,
irrespective of their abilities, background, or previous knowledge of the domain,
received exactly the same material in exactly the same sequence; neither did they
provide feedback, as the students' responses were ignored. This style of CAI has
been dubbed a d - h o c frame-orientated (AFO) CAI by Carbonell (1970) to stress its
dependence on author-specified units of information. Carbonell concluded that 'in
most CAI systems of the AFO type, the computer does little more than what a
programmed text book can do, and one may wonder why the machine is used at all
when teaching sequences are extremely simple, perhaps trivial, one should consid-
er doing away with the computer, and using other devices or techniques more
related to the task' (Carbonell, 1970, pp. 194,201). Overcoming these limitations
prompted the chain of events which has culminated in today' s ITSs.
Linear programs
t 950s
Bronchincj
progrorns
1960s
]
Generative CAI
~o~e t960~
Fig. 2. CAI to ITS metamorphosis.
I TSs1980s
Crowder (1959) overcame some of the limitations of Skinnerian systems by
ceasing to ignore students' responses. He proposed using them to control the
material shown to the student. The 'branching programs' that resulted still had a
fixed number of frames, but were able to comment on a student' s response and then
use it to choose the next frame, possibly repeating an earlier one. Pattern-matching
techniques allowed alternate answers to be treated as acceptable or partially accept-
able rather than as totally correct or incorrect as demanded by Skinnerian systems.
However, the teaching material became too large to be manageable through straight-
forward programming and so a special breed of programming languages, called
' author languages', were developed for creating CAI material.
In the late 1960s and early 1970s, 'generative systems' came into being (also
called ' adaptive systems'). These emerged from the recognition of the fact that the
teaching material could itself be generated by the computer. A generative system
has the ability both to generate and solve meaningful problems. In some domains
like arithmetic, researchers realized they could do away with all the pre-stored
teaching material, problems, solutions and associated diagnostics, and actually
generate them. The potential advantages, if exploited, were enormous. They in-
cluded drastically reduced memory usage and the generation and provision of as
many problems (to some desired level of difficulty) as the student needed. Most
notably, Uhr and his team implemented a series of systems which generated
problems in arithmetic that were 'tailor-made' to a student' s performance (Uhr,
1969). Suppes (1967) and Woods & Hartley (1971) produced systems with similar
abilities. Wexler (1970) describes a system which combines generative CAI with
frame-orientated CAI in which the course-author must specify certain question
formats. The system generates parameters for these formats and searches the data-
base to determine the correct answer. Nevertheless, a major shortcoming was the
256 H. S. Nwana
r est r i ct i on to dri l l -t ype exerci ses in domai ns as wel l - st r uct ur ed as mat hemat i cs.
Onl y par amet r i c summar i es of behavi our were used to gui de pr obl em generat i on,
r at her t han an expl i ci t r epr esent at i on of t he st udent ' s knowl edge ( Sl eeman &
Brown, 1982b). Sl eeman (1983) not es t hat in t he i ni t i al ver si on of t he Leeds
Adapt i ve Ar i t hmet i c Syst em ( whi ch later evol ved to t he Leeds Model l i ng Syst em,
LMS, Sl eeman & Smi t h (1981) and from t her e to t he PIXIE syst em (Sl eeman, 1987)),
t he model of the st udent consi st ed mer el y of an i nt eger to i ndi cat e t he l evel of t he
st udent ' s compet ence.
Gener at i ve CAI was t he mai n pr ecur sor of ITSs. Al t hough i ndi vi dual i zat i on and
f eedback had been i mpr oved, t her e was a rat her shal l ow knowl edge r epr esent at i on.
Yazdani (1986) not es t hat ' none of t hese syst ems has human- l i ke knowl edge of t he
domai n it is t eachi ng, nor can it answer t he seri ous quest i ons from t he st udent s as
to " why" and " how" t he task is per f or med. ' Such sent i ment s have also been echoed
by ot her researchers (e.g. O' Shea & Self, 1983). Hence, many probl ems r emai ned
unsol ved. Sl eeman & Br own (1982b) and Hawkes et al. (1986) poi nt out t hat t hese
syst ems wer e f ound to be l acki ng for a number of reasons.
(1) They at t empt ed to pr oduce total courses rat her t han concent r at i ng on bui l di ng
syst ems for more l i mi t ed t opi cs.
(2) They had severe nat ural language barriers whi ch rest ri ct ed users' i nt er act i on
wi t h t hem.
(3) They had no ' knowl edge' or ' under st andi ng' of t he subj ect t hey t ut or ed or of t he
st udent s t hemsel ves; this is somet i mes referred to in t he l i t erat ure as t he ' Eliza'
syndr ome. Consequent l y, t hey t ended to assume too much or t oo little st udent
knowl edge, and t hey coul d not concept ual i ze so as to di agnose a st udent ' s mi scon-
cept i on wi t hi n hi s/ her own framework.
(4) They were ext r emel y ad hoc. Bui l di ng t ut ori ng syst ems was not r ecogni zed to be
a non-t ri vi al task - - a task requi ri ng det ai l ed psychol ogi cal t heor i es of l earni ng and
mi sl earni ng. Anyone wi t h a knowl edge of comput i ng at t empt ed to bui l d a t ut or.
Consequent l y, t here was little or no co- oper at i on among educat ors, psychol ogi st s
and comput er sci ent i st s ill t he devel opment phase of t he tutors.
(5) They t ended to be static rat her t han dynami c. Ther e was little exper i ment at i on
wi t h syst ems in or der to i mpr ove t hem. Human t ut ors l earn about t hei r st udent s
and about t he subjects t hey t each ever y day, and so shoul d machi ne tutors.
In r esponse to t he pr obl ems CAI faced, Self, in his i nt erest i ng and classic paper,
argued t hat a comput er t ut ori al pr ogr am shoul d have a r epr esent at i on of what is
bei ng t aught , who is bei ng t aught and how to t each hi m/ her (Self, 1974). Carbonel l
(1970) argued that a sol ut i on to t hi s pr obl em coul d not be achi eved wi t hout t he use
of AI t echni ques. Jaime Carbonel l ' s i mpor t ant cont r i but i on to cogni t i ve sci ence is
best summar i zed i n t he title of his first-rate 1970 publ i cat i on AI in CAl. He want ed
to put AI i nt o CAI syst ems. He dr eamed of a syst em whi ch had a dat abase of
knowl edge about a subj ect mat t er and general i nf or mat i on about language and
pr i nci pl es of t ut ori al i nst ruct i on. The syst em coul d t hen pur sue a nat ural l anguage
di al ogue wi t h a st udent , somet i mes f ol l owi ng t he st udent ' s i ni t i at i ve, somet i mes
t aki ng its own i ni t i at i ve, but al ways generat i ng its st at ement s and r esponses in a
nat ural way from its general knowl edge. Such a syst em s har pedl y cont r ast ed wi t h
Int el l i gent Tut ori ng Syst ems: an overvi ew 257
exi st i ng CAI syst ems at t he t i me in whi ch a rel at i vel y fi xed sequence of quest i ons
and possi bl e r esponses had to be pr e- det er mi ned for each topic. He const r uct ed an
earl y ver si on of his dream, a classic syst em he cal l ed SCHOLAR (Carbonell, 1970,
1971), but he di ed before it was ful l y real i zed. Carbonel l ' s i nt r oduct i on of AI i nt o
CAI mar ked t he begi nni ng of t he era of ITSs whi ch t herefore emer ged to pr ovi de
answer s to t he l i mi t at i ons of generat i ve CAl. (ITSs are al t er nat i vel y r ef er r ed to as
Int el l i gent Comput er - Assi st ed I nst r uct i on/ Lear ni ng (ICAI/ICAL) syst ems as men-
t i oned in Sect i on 2.) It is cl ai med t hat ITSs combi ne AI, psychol ogi cal model s of t he
st udent and t he expert , and educat i onal t heory. The psychol ogi cal model s al l ow for
si mul at i ons of st udent per f or mance and can be exper i ment ed upon unt i l t hey
cl osel y r epr esent t he behavi our exhi bi t ed by t he st udent .
Summar i zi ng, pr ovi di ng a t r ul y ' i nt el l i gent ' syst em was r ecogni zed to be a non-
t ri vi al task whi ch needed expert s from several ot her di sci pl i nes; a need whi ch
coul d be pr ovi ded by t he AI r esear ch communi t y, whi ch cont ai ns comput er scien-
tists, psychol ogi st s and educat i onal i st s. Most of t he pr esent - day wor k in ITSs is
Fig. 3. General ITS architecture.
258 H. S. Nwana
bei ng carri ed out by AI researchers or ent husi ast s, and AIs i nf l uence has been so
great t hat Yazdani (1983) concl udes t hat ' i nt el l i gent t ut or i ng syst ems are AI' s
answer to CAL packages. ' However , it is wor t h maki ng t he poi nt t hat most so-cal l ed
ITSs still do not escape t he bul k of t he cri t i ci sms di scussed above, so t hat t here is
consi der abl e scope for f ur t her work. Some of t he maj or di fferences bet ween ITSs
and CAI programs are (or shoul d be):
(1) ITSs pr ovi de a cl ear ar t i cul at i on of knowl edge for a l i mi t ed domai n;
(2) ITSs have a model of st udent per f or mance whi ch is dynami cal l y mai nt ai ned
and is used to dri ve i nst ruct i on;
(3) t he ITS desi gner defi nes t he knowl edge and t he i nf er ence rules, but not t he
t eachi ng sequence, whi ch is der i ved by t he program;
(4) ITSs pr ovi de det ai l ed di agnost i cs of errors rat her t han si mpl y drill and pract i ce;
(5) st udent s can pose quest i ons to an ITS (this is t he mai n charact eri st i c of ' mi xed-
i ni t i at i ve tutors' ).
5 St ruct ure of ITSs
5.1 General Architecture
Exi st i ng ITSs vary t r emendous l y in archi t ect ure. In fact, it is al most a rari t y to fi nd
t wo ITSs based on t he same archi t ect ure. Thi s resul t s from t he exper i ment al nat ur e
of t he wor k in t he area: t her e is yet no cl ear-cut general ar chi t ect ur e for such
syst ems (Yazdani, 1986, 1987). Previ ousl y, t here was consi der abl e consensus in t he
l i t erat ure t hat ITSs consi st of at least t hr ee basic component s (Barr & Fei genbaum,
1982; Bonnet , 1985).
(1) The expert knowl edge modul e.
(2) The st udent model modul e.
(3) The t ut or i ng modul e.
However , more r ecent r esear ch (Wenger, 1987; Burns & Capps, 1988; Ma n d l &
Lesgold, 1988) has i dent i f i ed and added a f our t h component to t he list.
(4) The user i nt erface modul e.
Fi gure 3 i l l ust rat es t he general f or m of an ITS archi t ect ure, whi ch will serve as a
basis for di scussi on. It does not r epr esent any par t i cul ar known syst em.
The expert knowledge module compr i ses t he facts and rul es of t he par t i cul ar
domai n to be conveyed to t he st udent , i.e. t he knowl edge of t he expert s. In t he
t r ansi t i on from CAI to ITSs, such knowl edge has been t he first aspect of t he
t eacher ' s exper t i se to be expl i ci t l y r epr esent ed in systems. It has al r eady been
ment i oned that in t radi t i onal CAI, t he exper t i se to be communi cat ed is cont ai ned in
pre-st ored pr esent at i on mat eri al cal l ed ' frames' , whi ch are desi gned by t he exper t
t eacher and si mpl y di spl ayed to t he st udent under gi ven condi t i ons. Such i mpl i ci t
r epr esent at i on of knowl edge has si nce been r ecogni zed to be i nadequat e: in fact a
maj or l esson l earned from all t he r esear ch on exper t syst ems is t hat any exper t
modul e must have an abundance of speci fi c and det ai l ed knowl edge, der i ved f r om
I n t e l l i g e n t Tutoring S y s t e ms : an o v e r v i e w 259
people who have years of experience in a particular domain. Consequently, in ITSs,
much effort is expended in discovering and codifying the domain knowledge, i.e.
distilling years of experience into a knowledge representation.
KnowIedge eticitation and codification can be a very time-consuming task, espe-
cially for a complex domain with an enormous amount of knowledge and interrela-
tionships of that knowledge. Thus, investigating how to encode knowledge and
how to represent it in an ITS remains the central issue of creating an expert
knowledge module. In effect, this process aims to make the knowledge stored in
this module more explicit. So, in current ITSs, expert knowledge is represented in
various ways, including semantic networks, frames and production systems. It
must not only include surface knowledge (e.g. the descriptions of various concepts
that the student has to acquire), but also the representational ability that has been
recognized to be a critical part of expertise. Expert knowledge must include the
ability to construct implicit representational understanding from explicitly repre-
sented information (Mandl & Lesgold, 1988).
Expert knowledge modules can be classified along a spectrum ranging from
completely opaque or 'blackbox' representations, whereby only final results are
available (e.g. in tutors like SOPHI E I, reviewed in the next section), to fully
transparent or 'glassbox' ones, where each reasoning step can be inspected and
interpreted (e.g. SOPHI E IlI, also reviewed later).
The expert knowledge module or domain expert, as it is alternatively termed,
fulfils a double function. First, it serves as the source of knowledge to be presented
to the student, which includes generating questions, explanations and responses.
Secondly, it provides a standard for evaluating the student' s performance. For this
latter function, it must be able to generate solutions to problems in the same context
as the student, so that respective answers can be compared. The module must also
be able to detect common systematic mistakes, and if possible identify any gap in
the student' s knowledge that may be the cause of this. If the ITS is to monitor
students in solving problems, the expert module must also be able to generate
sensible, and possibly multiple, solution paths so that intermediate steps can be
compared.
Also, in its function as a standard, the expert knowledge module can be used to
assess the student' s overall progress. To achieve this required the establishment of
some criteria to compare knowledge. This type of comparison is possible only if the
knowledge has been explicitly represented. Hence, ITSs differ considerably from
traditional CAI programs in that the knowledge in the latter is implicitly repre-
sented within its code.
It is also worth recognizing that the expert knowledge module by necessity
embodies a specific view of the domain - - that of the designer. Thus, tutoring can
be compromised if the student does not understand the system' s instruction or
because the system cannot interpret the student' s behaviour in terms of its own
view of the knowledge (Wenger, 1987). Of course, human teachers also have their
own views, but they do also have the incredible ability to adapt them accordingly in
order to perceive that of the student. This issue touches on the central problem of
knowledge representation in AI; a problem which is often said to be the main
2fi0 H. S. Nwana
bot t l e- neck to AI' s success. The sol ut i on to this bot t l e-neck is par t i cul ar l y r el evant
to ITSs because of its deep pedagogi cal i mpl i cat i ons.
The student model module refers to t he dynami c r epr esent at i on of t he emergi ng
knowl edge and skill of t he st udent . No i nt el l i gent t ut or i ng can take pl ace wi t hout
an under s t andi ng of t he st udent . Thus, al ong wi t h t he i dea of expl i ci t l y r epr esent -
ing t he knowl edge to be communi cat ed came t he idea of doi ng l i kewi se wi t h t he
st udent , in t he f or m of a st udent model . Most researchers agree t hat an ITS shoul d
have a st udent model (Hart l ey & Sl eeman, 1973; Self, 1974, 1987a, 1987b, 1988b;
Rich, 1979; Sl eeman, 1985: Tobias, 1985; Zissos & Wi t t en, 1985; Ross et al., 1987;
Gi l more & Self, 1988: Wachsmut h, 1988). Ideally, this model shoul d i ncl ude all
t hose aspect s of t he st udent ' s behavi our and knowl edge t hat have possi bl e reper-
cussi ons on hi s/ her per f or mance and learning. However , t he task of const r uct i ng
such a compl et e model is not onl y non-t ri vi al but, probabl y, i mpossi bl e, especi al l y
consi der i ng t hat t he communi cat i on channel , whi ch is usual l y t he keyboar d, is so
rest ri ct i ve. Human t ut ors woul d nor mal l y combi ne data from a vari et y of ot her
sources, like voi ce effects or facial gestures. They may also be able to det ect ot her
phenomenol ogi cal fact ors such as bor edom or mot i vat i on whi ch are also cruci al in
l earni ng.
In a r ecent revi ew, Self (1988b) i dent i fi ed t went y di fferent uses t hat had been
f ound for st udent model s in exi st i ng ITSs. From anal ysi ng this list, he not es t hat t he
f unct i ons of st udent model s coul d be general l y cl assi fi ed i nt o six t ypes.
(1) Correct i ve: to hel p er adi cat e bugs in the st udent ' s knowl edge.
(2) Elaborative: to hel p cor r ect ' i ncompl et e' st udent knowl edge.
(3) Strategic: to hel p i ni t i at e si gni fi cant changes in t he t ut ori al strategy ot her t han
t he t act i cal deci si ons of 1 and 2 above.
(4) Diagnostic: to hel p di agnose bugs in the st udent ' s knowl edge.
(5) Predi ct i ve: to hel p det er mi ne t he st udent ' s l i kel y r esponse to t ut ori al act i ons.
(6) Eval uat i ve: to hel p assess t he st udent or t he ITS.
In t he aut hor ' s view, Sel f' s preceding list is still far from comprehensive; it coul d
be nar r owed down much furt her, and t he st udent model coul d be seen to fulfil a
doubl e f unct i on. On t he one hand, it acts as a sour ce of i nf or mat i on about t he
st udent . On t he ot her hand, it serves as a r epr esent at i on of t he st udent ' s knowl edge.
Wenger (1987) also suppor t s this vi ew. The next few paragraphs will at t empt to
j ust i fy this af or ement i oned vi ewpoi nt .
In its f unct i on as a sour ce of i nformat i on, it infers unobser vabl e aspect s of t he
st udent ' s behavi our from t he model . Such an i nf er ence coul d pr oduce an inter-
pr et at i on of hi s/ her act i ons and also lead to a r econst r uct i on of t he knowl edge t hat
gave rise to t hese act i ons. Such knowl edge is vital for t he pedagogi c c ompone nt of
t he ITS as it coul d be used in ei t her of t he six ways not ed by Self.
The st udent model is also l i kel y to be f or med out of t he syst em' s r epr esent at i on of
t he target knowl edge in t he exper t knowl edge modul e. Accor di ngl y, t he st udent
model can i ncl ude a cl ear eval uat i on of t he mast ery of each uni t of knowl edge in
t he exper t modul e (the f unct i on of t he st udent model here is eval uat i ve). Thi s
Intelligent Tutoring Systems: an overview 281
al l ows t he st udent ' s state of knowl edge to be compar ed wi t h t he expert knowl edge
modul e, and i nst r uct i on woul d t hen be bi ased t owar ds por t i ons of t he model
s hown to be weak (thus, t he st udent model ' s f unct i on her e is elaborative). Thi s f or m
of st udent model l i ng is referred to as ' over l ay' model l i ng (Goldstein, 1982), because
t he st udent ' s state of knowl edge is vi ewed as a subset of t he expert ' s. Again, this
shows how t he st udent model acts as a sour ce of i nformat i on.
However , i ncor r ect or subopt i mal behavi our does not al ways resul t f r om i ncom-
pl et e knowl edge. It coul d also be due to i ncorrect versi ons of t he target knowl edge.
Ther ef or e a mor e f or mat i ve st udent model shoul d also pr ovi de expl i ci t r epr esent a-
t i ons of t he st udent ' s i ncor r ect versi ons of t he target knowl edge for r emedi at i on
pur poses (clearly, t he f unct i on of t he st udent model her e is di agnost i c and correc-
tive). In such a capaci t y, it serves as a r epr esent at i on of t he st udent ' s knowl edge.
Thi s appr oach to model l i ng is cal l ed t he ' buggy' appr oach (Brown & Burt on, 1978a).
St udent model s are also expect ed to be execut abl e or r unnabl e. Thi s al l ows for
exact pr edi ct i on about a par t i cul ar st udent in a par t i cul ar cont ext (thus, t he func-
t i on of t he st udent model her e is predi ct i ve). The t ut ori ng modul e woul d also be
maki ng use of such execut abl e r epr esent at i ons for pedagogi c pur poses (i.e, t he
knowl edge r epr esent ed in t he st udent model is bei ng used in a strategic way). The
st udent model is also servi ng her e as a r epr esent at i on of t he st udent ' s knowl edge.
In concl usi on, st udent model s coul d be seen to per f or m t wo ' super ' funct i ons:
act i ng as a sour ce of i nf or mat i on about t he st udent , and servi ng as a r epr esent at i on
of t he st udent . In achi evi ng t hese f unct i ons, t hey act in roles i ncl udi ng correct i ve,
el aborat i ve, strategic, di agnost i c, pr edi ct i ve and eval uat i ve.
The tutoring module is t he part of t he ITS that desi gns and regul at es i nst r uct i onal
i nt er act i ons wi t h t he st udent . In ot her archi t ect ures, this modul e is referred to as
t he t eachi ng st rat egy or t he pedagogi c modul e. It is cl osel y l i nked to t he st udent
model , usi ng knowl edge about t he st udent and its own t ut ori al goal st r uct ur e to
deci de whi ch pedagogi c act i vi t i es will be present ed: hi nt s to over come i mpasses in
per f or mance, advi ce, suppor t , expl anat i ons, di fferent pract i ce tasks, tests to con-
f i r m hypot heses in t he st udent ' s model , etc. (Self, 1988b). The t ut ori al component
is t hus t he sour ce and t he or chest r at or of all pedagogi c i nt ervent i ons. The deci si ons
i nvol ved are subtle. The or der and manner in whi ch t opi cs are t reat ed can pr oduce
ver y di fferent l earni ng exper i ences. In t ut ori al , it is somet i mes mor e effect i ve to let
t he st udent f l ounder for a whi l e before i nt errupt i ng; somet i mes, t he st udent will get
st uck or lost if left compl et el y to hi msel f/ hersel f (however, no good human t ut or
will dest r oy a st udent ' s per sonal mot i vat i on or sense of di scovery). Consequent l y,
t he t ut or i ng in exi st i ng ITSs can be classified al ong a spect r um rangi ng f r om
syst ems t hat monitor t he st udent ' s ever y act i vi t y ver y cl osel y, adapt i ng t hei r ac-
t i ons to t he st udent ' s r esponses but never r el i nqui shi ng cont rol , to guided-
discovery learning syst ems wher e t he st udent has al most full cont rol of t he act i vi t y,
and t he onl y way t he syst em can di rect t he cour se of act i on is by modi f yi ng t he
envi r onment . In t he mi ddl e are mixed-initiative syst ems wher e t he cont rol is
shar ed by t he st udent and t he syst em as t hey exchange quest i ons and answers. The
exi st ence of this spect r um cl earl y hi ghl i ght s the fact that t ut or i ng is an art t hat
262 H. S. Nwana
requi res great versat i l i t y whi ch is still ext r emel y di ffi cul t to art i cul at e and r epr esent
in an ITS. Nevert hel ess, some progress is bei ng made, albeit l i mi t ed.
ITSs also ai m at expl i ci t l y r epr esent i ng t he knowl edge f ound in t he t ut or i ng
modul e. Thi s creat es t he pot ent i al to adapt and i mpr ove strategies over t i me (as in
t he case of sel f-i mprovi ng tutors), and for t he same strategies to be used for ot her
domai ns. Once mor e this is cont rast ed wi t h t radi t i onal CAI syst ems wher e such
pedagogi cal knowl edge is deepl y bur i ed in t he vari ous pi eces of code t hat cont r ol
t he t ut ori al i nt eract i on.
The user i nt erf ace modul e is t he communi cat i ng component of t he ITS whi ch
cont r ol s i nt er act i on bet ween t he st udent and t he syst em, as depi ct ed in Fig. 3. In
bot h di rect i ons, it t ransl at es bet ween t he syst em' s i nt ernal r epr esent at i on and an
i nt erface language t hat is under st andabl e to t he st udent . Because t he user i nt erface
can make or break t he ITS, no mat t er how ' i nt el l i gent ' t he i nt ernal syst em is, it has
become cust omar y to i dent i f y it as a di st i nct component of its own. In fact, it woul d
be a mi st ake to consi der it a secondar y component of t he ITS for t wo mai n reasons.
First, whe n the ITS pr esent s a t opi c, t he i nt erface can enhance or di mi ni sh t he
pr esent at i on. Si nce t he i nt erface is t he final form in whi ch t he ITS present s itself,
qual i t i es such as ease of use and at t ract i veness coul d be cruci al to t he st udent ' s
accept ance of t he syst em. Secondl y, progress in medi a t echnol ogy is i ncr easi ngl y
pr ovi di ng more and mor e sophi st i cat ed tools whose communi cat i ve power heavi l y
i nf l uences ITS design.
Cur r ent ITSs pr ovi de user i nt erfaces whi ch, for t he i nput , range from t he use of
fi xed menus wi t h mul t i pl e- choi ce answers to a fairly free t r eat ment of a pseudo-
nat ural language. For t he out put , t hey range from t he mer e di spl ay of pre-st ored
t ext s t ypi cal of CAI, to t he use of fairly compl i cat ed generi c frames. Wi t hi n t hese
t wo ends of t he spect r um t here is also a varyi ng fl exi bi l i t y (Wenger, 1987).
Some ITSs are maki ng t hei r i nt er act i on wi t h t he st udent mor e ' user f r i endl y' by
subst i t ut i ng pi ct ures and poi nt i ng for t ext and t ypi ng. In some, t he t r eat ment of t ext
is onl y s uppl ement ed by pi ct ur es and graphics; much mor e cogni t i ve r esear ch i nt o
t he use of such i nt erfaces is still requi red. However , t he most one expect s to see, in
t he near fut ure, is ITSs communi cat i ng wi t h t he st udent pr i mar i l y via graphi cs, as
r esear ch in nat ural language under s t andi ng and comput er speech r ecogni t i on/
gener at i on are still at such earl y and pr i mi t i ve stages. In fact, research i nt o user
i nt erfaces is just commenci ng, as t hey have onl y r ecent l y been acknowl edged as
di st i nct part s of ITSs.
5.2 Ot her archi t ect ures
Ot her ar chi t ect ur es have been suggested, some largely similar, but ot hers r adi cal l y
di f f er ent to t hat depi ct ed in Fig. 3. For exampl e, t here are four component s to
Ander son' s Advanced Comput er Tut or i ng (ACT) ITS ar chi t ect ur e ( Ander son et al.,
1985a,b) (see Fig. 4).
(1) The domai n exper t (ideal st udent model ): this modul e cont ai ns all t he correct
rul es used for sol vi ng pr obl ems in t he domai n.
(2) The bug catalogue: t hi s is an ext ensi ve l i brary of common mi sconcept i ons and
errors for t he domai n.
Intelligent Tutoring Systems: an overvi ew 263
(3) The t eachi ng knowl edge (tutoring modul e).
(4) The user interface.
Thi s archi t ect ure is not onl y a proposal; at least t wo tutors are based on it
i ncl udi ng t he Geomet ry t ut or and the Lisp tutor (Anderson et al., 1985a).
Tutoring knowledge
L D om ain expert L
F
I
U ierinterfoce
L
F
I
Student I
J
v
Student I
Fig. 4. Andersons ITS architecture.
Fig. 5. O' Shea et al.' s architecture.
2B4 H.S. Nwa n a
Hartley & Sleeman, whose 1973 architecture is probably the closest proposal to
the general architecture depicted in Fig. 3, suggest that an ITS ought to have four
distinct knowledge bases:
(1) knowledge of the domain (expert knowledge),
(2) knowledge of the person being taught (student model),
(3) knowledge of the teaching strategies (tutoring knowledge),
(4) knowledge of how to apply the tutoring knowledge to the needs of an indi-
vidual,
Hartley & Sleeman's proposals differ from Anderson's inasmuch as they do not
give the misconceptions in the domain (the bug catalogue) primary importance but
instead introduce the student model as a primary component (Yazdani, 1987).
Furthermore, this proposal subsumes the user interface in a more tutoring-
orientated module which includes meta-rules that guide the tutoring rules. These
differences also reflect the different tutoring philosophies involved in both
architectures; Anderson plays down the importance of the student model and, in its
place, substitutes two knowledge bases of ideal and buggy representations of
knowledge of the domain (see Fig. 4). Immediately a behaviour is exhibited which
indicates an error or bug, the student is nudged to follow the correct path by being
presented with the ideal solution.
O'Shea et al. {1984) present a five-ring model as shown in Fig. 5, This bears some
similarity to the Hartley & Sleeman architecture. However, it also clearly demons-
trates how differences in emphasis on student modelling and teaching lead to an
architecture which is starkly different from Anderson's (see Fig. 4). Its components
include:
(1) student history,
(2) student model,
(3) teaching strategy,
(4} teaching generation,
(5) teaching administration.
In this proposal, the explicit representation of knowledge in the domain (expert
knowledge), and the common misconceptions in the domain (the bug catalogue),
are undermined in favour of teaching skills: hence the introduction of the teaching
generation and teaching administration components.
Figure 6 presents another architecture which introduces the self-improving
{learning) concept. The architecture thus forms the basis of self-improving systems
which attempt to improve their tutoring capabilities over time. Such systems are
still rare and typically consist of two components: an adaptive teaching program
(which may have any of the architectures discussed earlier or some other), and a
self-improving component that makes experimental changes using data collected
during teaching sessions. O'Shea's (1982) Quadratic tutor was one of the earliest to
exploit this idea. This ambitious architecture, which attempts to automate a process
that is even difficult for humans, is an important contribution; most researchers
agree but pay lip service to the fact that ITSs should improve or learn over time as
J
- I
Int el l i gent Tut or i ng Syst ems: an overvi ew
[ $1"uden'l" l
I
AdQptive te(]ching progrom
265
Self-improving program
Fig. 6. Sel f-i mprovi ng ITS architecture.
good h u ma n t ut or s do. Ki mbal l ' s (1982) I nt egr at i on t ut or is gener al l y acknowl edged
to be t he first s el f - i mpr ovi ng t ut or to have been devel oped.
In addi t i on to t he above ar chi t ect ur es, mor e recent s ys t ems concent r at e on ot her
f eat ur es s uch as pl anni ng (e.g. see Peachey & McCal l a, 1986). The SCENT-3
ar chi t ect ur e evol ved by McCal l a' s t eam (McCalla & Greet , 1988) is mu c h mor e
c ompl e x and i nt er est ed r eader s are r ef er r ed to t hi s paper . Suffi ce to say t hat it is
bas ed on a bl ackboar d phi l os ophy; t he bl ackboar d medi at es c ommuni c a t i on a mong
t he nume r ous c ompone nt s of t he ar chi t ect ur e.
5.3 Di scussi on
It is t hus evi dent t hat di fferi ng t ut or i ng phi l os ophi es pl ace e mpha s e s on di f f er ent
aspect s of t he i nst r uct i onal pr ocess (e.g. st udent model l i ng) , and t hi s in t ur n l eads
to di f f er ent ar chi t ect ur es, as in t he case of O' Shea et al. and Ander son. It is
i nconcei vabl e t hat t her e woul d be a cons ens us amongs t r esear cher s on a t ut or i ng
phi l os ophy; rat her, mor e var i at i ons of pr esent phi l os ophi es are emer gi ng. Conse-
quent l y, t hi s has r esul t ed in t he nume r ous ITS ar chi t ect ur es in t he l i t erat ure.
Nat ur al l y, not all t hese ar chi t ect ur es can s uppor t a range of t ut or i ng st rat egi es
wi t hi n a gi ven ITS.
Most of t hese ar chi t ect ur es still r e ma i n onl y pr oposal s. Also, al mos t all t he
ar gument s to t hei r s uppor t or as to whi ch of t hem is t he ' bes t ' have, di s appoi nt i ng-
ly, been t heor et i cal r at her t han pract i cal . Unt i l ITSs bui l t ar ound t hese pr opos al s
are eval uat ed and demons t r at ed to be educat i onal l y wor t hy, none can cl ai m any
s uper i or i t y over anot her . Regret t abl y, a si gni fi cant fract i on of t hese pr opos al s have
not even had s ys t ems devel oped whi ch are based on t hem. Admi t t edl y, a f ew have
(e.g. Ander s on' s ) , but an i nsuf f i ci ent numbe r of ITSs based on t hese ar chi t ect ur es
have been devel oped, eval uat ed and pr oven ' us ef ul ' to war r ant s uch ar gument s. In
fact, due to t he exper i ment al nat ur e of t he area, even t he so- cal l ed gener al
ar chi t ect ur e of Fig. 3 s houl d be vi ewed wi t h s ome degr ee of s cept i ci s m and s houl d
def i ni t el y not be seen as a basi s for all ITSs. When mor e de ve l ope d ITSs be c ome
eval uat ed, as is expect ed to be t he case, it is possi bl e t hat s ome addi t i onal , or
266 H. S . Nwona
per haps al t ernat e, set of bui l di ng bl ocks mi ght emer ge (as is
bl ackboar d ar chi t ect ur e of McCal l a & Greer, 1988).
t he case in t he
6 Re v i e w of s o me c l a s s i c ITSs
In t hi s sect i on, s ome i mpor t ant ITSs are r evi ewed; an over vi ew paper of t hi s sort
woul d be i ncompl et e wi t hout it. It is by no means meant to be exhaust i ve: rat her, it
s houl d augment t he pr evi ous sect i ons of t hi s paper. The syst ems r evi ewed are
ei t her chosen for t hei r hi st or i cal si gni f i cance or because t hey are good exampl es of
s ys t ems whi ch di spl ay s ome of t he i nt el l i gent t ut or i ng pr i nci pl es r epor t ed in t hi s
paper.
6.1 SCHOLAR
As t he s ys t em whi ch successf ul l y l aunched t he new par adi gm of i nt el l i gent t ut or-
ing, SCHOLAR sur el y deser ves a pl ace in any r evi ew of ITSs. Needl ess to say it was
t he first ITS to be const r uct ed. It was a r evol ut i onar y s ys t em when cons i der ed in its
hi st or i cal cont ext ; mos t of t he t hen exi st i ng ITSs wer e of t he ad-hoc frame-
or i ent at ed (AFO) t ype. SCHOLAR was cr eat ed by Jai me Carbonel l ; t hi s aut omat i cal -
ly ear ned hi m a pl ace in hi st or y as f ounder of ITSs. He used SCHOLAR to l aunch a
new par adi gm whi ch he cal l ed ' i nf or mat i on- s t r uct ur e- or i ent ed' (ISO) CAI as
oppos e d to t he pr e domi na nt AFO- t ype CAI of t he t i me (Carbonell, 1970, 1971).
These t wo appr oaches cor r es pond in ma ny ways to what are now r es pect i vel y
cal l ed t r adi t i onal CAI and ITS. Because SCHOLAR, like most ot her ITS proj ect s,
evol ved, t he ver si on descr i bed in t hi s r evi ew is Car bonel l ' s ori gi nal versi on: be-
si des, it is the i mpor t ant one because of its hi st ori cal si gni f i cance to ITS r esear ch.
SCHOLAR was a pi oneer i ng effort in t he devel opment of comput er t ut ors capabl e
of handl i ng unant i ci pat ed st udent quest i ons and of generat i ng i nst r uct i onal mat e-
rial in var yi ng l evel s of detail, dependi ng on the cont ext of t he di al ogue. It was a
mi xed- i ni t i at i ve ITS: bot h t he s ys t em and t he st udent coul d i ni t i at e conver s at i on by
aski ng quest i ons. Bot h t he pr ogr am' s out put and t he s t udent ' s i nput s wer e Engl i sh
sent ences. It seems appr opr i at e to r evi ew it f ur t her by consi der i ng its f our compo-
nent s, as def i ned in Sect i on 5.1.
The knowl edge in t he exper t knowl edge modul e is t hat of t he geogr aphy of Sout h
Amer i ca, whi ch was r epr es ent ed in a semant i c net wor k whose nodes i nst ant i at ed
geogr aphi cal obj ect s and concept s. St at ement s like ' Tel l me mor e about Brazi l ' j ust
i nvoked a r et r i eval of facts st ored in t he semant i c net wor k. However , t he real power
of t hi s r epr es ent at i on s chema comes by r ecogni zi ng t hat it is possi bl e to ans wer
quest i ons for whi ch answer s are not st ored. Thi s aut omat i cal l y r el i eves t he s ys t em
of t he me mor y pr obl ems encount er ed in ant i ci pat i ng and st ori ng all sol ut i ons by
t r adi t i onal CAI syst ems. For exampl e, it is not necessar y to store in t he s emant i c
net wor k t hat ' Li ma is in Sout h Amer i ca' pr ovi ded t hat t he pr ogr am whi ch i nt er-
pret s t he net wor k can make t he r el evant i nference. In ot her wor ds, t he pr ogr am
mus t know about t he at t r i but es concer ned, e.g. ' l ocat i on' and ' capi t al ' , and in
par t i cul ar , t hat if x is capi t al of v and y is l ocat ed in z t hen x is in z: t hi s is a rul e of
i nference.
Int el l i gent Tutoring Syst ems: an overvi ew 267
A semantic network representation was chosen because Carbonell thought it to be
close to the teacher's conceptualization of knowledge. By implication, the network
also represented the ideal student' s conceptualization (ideal student model): hence,
overlay modelling becomes feasible. So SCHOLAR could associate flags with each
node of the network to indicate whether the student was thought to know the
information represented by that node. More ambitiously, Carbonell proposed to
model student errors by introducing small 'perturbations' to the network; this
proposal was not followed up in SCHOLAR. Its student modelling was then ex-
tremely rudimentary.
SCHOLAR' s tutorial strategies were also fairly primitive, consisting mainly of
local topic selections. The teacher using it was expected to provide an agenda.
Whenever a topic was too general, SCHOLAR generated a subtopic on an essentially
random basis. For example, the teacher might specify the topic of 'South America',
and SCHOLAR would select a subtopic, e.g. 'Peru', and then perhaps a sub-
subtopic, e.g. 'topography of Peru'. This random element led to somewhat discon-
nected discussions lacking the systematic development of ideas characteristic of a
good tutorial, though it was necessary since SCHOLAR' s semantic network had
little information about desirable orders of presentations of topics. Nevertheless,
suitable relevant tags in the network (from the agenda), could provide SCHOLAR
with some reasonable guidance in selecting topics.
SCHOLAR possessed language processing capabilities that were also rather li-
mited. Text was generated by sentence and question templates that was filled up
with information from the network. The parsing of student' s questions followed the
same principle in reverse, while the parsing of student's answers was done by
matching key words from a list dynamically generated from the network for each
question. Consequently, SCHOLAR did not understand wrong answers and so
could not glean diagnostic information from them.
SCHOLAR has not been widely used except in NLSoSCHOLAR, an intelligent
on-line consultant for a text editor (Grignetti et al., 1975). This was partly due to
some fundamental limitations such as difficulty of representing procedural know-
ledge using semantic nets. However, despite all its shortcomings, SCHOLAR intro-
duced many methodological principles that have become central to ITS design, e.g.
separation of tutorial strategies from domain knowledge, more explicit representa-
tion of knowledge, student modelling, etc. Indeed, SCHOLAR' s significance as a
milestone for the entire field cannot be over-emphasized.
6.2 SOPHIE
SOPHIE (a SOPHisticated Instructional Environment) is an ITS which reflects a
major attempt to extend Carbonell's notion of mixed-initiative CAI (introduced in
SCHOLAR) for the purpose of encouraging a wider range of student initiatives
(Brown et al., 1975). It was developed by John Seely Brown, Richard Burton, and
their coleagues at Bolt Beranek and Newman, Inc. This project went through
successive phases spanning more than 5 years; the three stages of development of
SOPHIE (I-III) incorporate the most intensive attempt at building at complete ITS
268 H. S. Nwa n a
so far. SOPHI E is a mi l est one for t he field, and hence it wel l earns its pl ace in t hi s
revi ew.
The pedagogi c phi l os ophy is di fferent in SOPHIE: it is not so much to i mi t at e a
di al ogue wi t h a huma n t eacher (as S CHOLAR sought to do) as to pr ovi de a react i ve
l earni ng envi r onment in whi ch t he st udent can t ry hi s ideas, have t hem assessed,
and r ecei ve advi ce. Its phi l os ophy is t hus ' l ear ni ng by doi ng' as opposed to ' l earn-
ir~g by bei ng t ol d' as in t he case of SCHOLAR. Br own et al. (1982) suggest t hat
comput er t echnol ogy can be used to make exper i ment at i on bot h ' easi er' and ' safer'
by si mul at i ng envi r onment s t hat capi t al i ze on t he mot i vat i onal val ue of expl or at or y
pr obl em- sol vi ng act i vi t i es. SOPHI E' s si mul at ed area of exper t i se is el ect r oni c
t r oubl eshoot i ng. Si nce t he component s of a si mul at i on can be made faul t y, t roub-
l eshoot i ng means per f or mi ng a series of measur ement s to pr opose and test hypot h-
eses concer ni ng t he l ocat i on and nat ure of t he fault. Thi s not onl y gives t he st udent
t he oppor t uni t y to appl y a t heoret i cal knowl edge of el ect r oni c laws, but also to
acqui r e genera] t r oubl eshoot i ng strategies. In essence, it enabl es t he st udent to have
a one-to~one r el at i onshi p wi t h an ' exper t ' who hel ps create, exper i ment wi t h, and
debug hi s/ her own i deas (Brown et al., 1975). In keepi ng wi t h t he obj ect i ves of t hi s
paper, SOPHI E wi l l f ur t her be r evi ewed by consi der i ng t he four component s as
def i ned in Sect i on 5.1.
The exper t knowl edge modul e of SOPHI E compr i ses a ' st rong' model (simula-
t i on) of el ect r oni c t r oubl eshoot i ng for t he IP-28 regul at ed power suppl y and a
' canned' art i cul at e exper t t r oubl eshoot er whi ch can not onl y sol ve probl ems, but is
also capabl e of expl ai ni ng its t act i cs and hi gh-l evel strategies for at t acki ng t he
probl em. For exampl e, t he exper t can expl ai n ' why' a measur ement was made and
' what ' l ogi cal l y fol l ows from t he meas ur ement obt ai ned.
SOPHI E' s t ut or i ng modul e possesses numer ous heur i st i c strategies for answer i ng
and cr i t i qui ng a st udent or generat i ng al t ernat i ve t heori es to hi s/ her cur r ent hypot h-
eses. SOPHI E I's t ut ori al capabi l i t i es wer e i mpr essi ve as ment i oned in t he pr evi ous
paragraph: however , wi t h t he evol ut i on from SOPHI E I to SOPHI E III, t he strategies
wer e enhanced to become mor e ' human- l i ke' . Thi s was because its i mpl ement or s
obser ved t hat SOPHI E i and II's appr oaches to pr obl em sol vi ng were forei gn to
humans (Brown et al., 1982). Consequent l y, SOPHI E evol ved from a si mul at i on-
based i nf er ence syst em to a mor e power f ul and human- l i ke r easoni ng syst em (using
qual i t at i ve r easoni ng t echni ques). By i mpl i cat i on, SOPHI E' s sophi st i cat ed t ut or i ng
modul e requi res a si mi l arl y sophi st i cat ed st udent model .
By all account s, t he SOPHI E i nt erface demonst r at es a ver y i mpr essi ve nat ural
l anguage capabi l i t y; it uses t he power f ul not i on of semant i c grammars pr oposed by
Ri chard Burt on. It is robust ( handl i ng ' near l y all sent ences gener at ed by users who
have had a few mi nut es exposur e to t he syst em' ), effi ci ent ( ' under st ands a t ypi cal
st at ement in a f r act i on of a second' ), and of some general i t y (' si nce t he not i on of
semant i c grammars has been successf ul l y appl i ed to ot her areas besi des el ect ro-
nics' ): t he i nt erface undeni abl y demonst r at es that t echni ques for pr ocessi ng nat ural
l anguage are suffi ci ent l y devel oped to be usabl e in ITSs (O' Shea & Self, 1983).
Summar i zi ng, SOPHI E' s per f or mance as a ' compl et e' ITS is pr esent l y unsur pas-
sed. SOPHI E was act ual l y commi s s i oned by t he Amer i can Defence Depar t ment and
had l i mi t ed use for on-si t e job t rai ni ng over t he ARPA net wor k/ ARPA I nt er net is a
Int el l i gent Tut ori ng Syst ems: an overvi ew 269
net wor k of several net works; t he most i mpor t ant ARPANET links US r esear ch
cent r es and uni ver si t i es (wi t h 150 nodes)) for 2 years; it is no l onger mai nt ai ned (no
r eason has been gi ven in t he l i t erat ure for this). However , consi der i ng t hat most
ot her pr ot ot ype ITSs never ever get used aft er t hei r devel opment , SOPHIE was qui t e
a success story. Probabl y, SOPHIE' s most i mpor t ant cont r i but i on to ITS, t hough,
was to est abl i sh it as a r espect abl e subarea in the eye of t he AI c ommuni t y (Wenger,
1987).
6.3 GUIDON
GUIDON, an ITS for t eachi ng di agnost i c pr obl em solving, was devel oped by Wil-
l i am Cl ancey and his col l eagues at St anf or d Uni versi t y. The GUIDON proj ect is also
uni que as it r epr esent s t he first at t empt to adapt a pre-exi st i ng exper t syst em i nt o an
i nt el l i gent t ut or. Pr obabl y because it was st rongl y i nf l uenced by SCHOLAR and
SOPHIE, it t ur ned out to be one of t he most concer ned efforts so far at desi gni ng an
ITS. Like t he ot her t wo, it went t hr ough many stages spanni ng mor e t han 5 years
dur i ng whi ch it yi el ded many i mpor t ant fi ndi ngs. All of this coupl ed wi t h t he fact
t hat it is bui l t ar ound t he most wel l - known exper t syst em, MYCIN, earns it a pl ace
as one of t he key ITS proj ect s ever under t aken.
GUIDON' s goal is to t ut or t he knowl edge from t he famous exper t syst em, MYCIN
(Shortliffe, 1976), a medi cal exper t syst em that suggests t r eat ment for bact eri al
i nfect i ons. It at t empt s to t ransfer exper t i se to t he st udent s excl usi vel y t hr ough case
di al ogues wher e a sick pat i ent (the ' case' ) is descri bed to t he st udent in general
terms. The st udent is t hen asked to pl ay t he role of a physi ci an and ask for
i nf or mat i on he/ she t hi nks mi ght be rel evant to t he case. GUIDON compar es t he
st udent ' s quest i ons to t hose whi ch MYCI N woul d have asked and cri t i ques hi m/ her
on this basis; this demonst r at es a di fferent t ut ori ng strategy to t hat of SCHOLAR or
SOPHIE. It is easy to i nfer from t he pr evi ous sent ence that st udent model l i ng is
l argel y of t he over l ay- t ype. GUIDON also separat es its t ut ori al strategies (compri s-
ing 200 rules), whi ch was l argel y i nf l uenced by SOPHIE' s, from its domai n know-
ledge. Nevert hel ess, its nat ural l anguage cgpabi l i t i es are far less sophi st i cat ed t han
SOPHIE' s, but cer t ai nl y i mpr ove on SCHOLAR' s.
The component whi ch has evol ved consi der abl y has been t he exper t modul e. The
ori gi nal ver si on (GUIDON 1) was i mpl ement ed by ' reversi ng' MYCIN' s 450 rul es
(Clancey, 1984). Thi s i mpl ement at i on was i neffect ual l argel y because medi cal
di agnosi s is not made ' cookbook' st yl e - - i.e. medi cal pract i t i oners do not di agnose
di seases by usi ng perfect recal l on hundr eds of medi cal facts and rul es IClancey,
1982, 1983, 1987). He r eal i zed t hat MYCI N' s rules r epr esent ' compi l ed' knowl edge
devoi d of t he l ow-l evel det ai l and r el at i on necessary for l earni ng and t ut ori ng.
GUIDON 1 's fai l ure was l argel y due to t he fact that it woul d have had to ' decompi l e'
and augment t hese rul es wi t h data and di agnost i c hypot heses t hat t he medi cal
pr act i t i oner uses i mpl i ci t l y. Thus, MYCIN' s rules wer e r econf i gur ed to separat e t he
st rat egi c knowl edge from t he domai n facts and rules, resul t i ng in NEOMYCI N
(Cl ancey & Let si nger, 1981). Thi s in t ur n became t he new basis ar ound whi ch
GUIDON 2 was built, and wi t h some i mpr oved t eachi ng strategies, GUIDON 2 has
had great er success as a pr ot ot ype t ut or i ng syst em t han GUIDON 1 (Clancey, 1987).
270 H. S. Nwana
The GUIDON project provi ded a fasci nat i ng i nqui ry into the epi st emol ogi cal
quest i ons rel at ed to i nt el l i gent t ut ori ng as well as produci ng many i mpor t ant
fi ndi ngs about desi gni ng ITSs. For example, it cl earl y demonst r at ed that an expert
syst em is not a sound basis for t ut ori ng (Elsom-Cook, 1987). More i mpor t ant l y
however, GUIDON pr oduced spect acul ar demonst rat i ons of the fi el d' s abi l i t y to
bring to light f undament al AI research issues (Wenger, 1987).
6.4 WEST
The WEST coach is a program devel oped, again, by Ri chard Burt on and John Seely
Brown to hel p st udent s pl ay a game on the PLATO syst em (PLATO was one of the
largest ever CAI projects ever undert aken). It is i ndeed an i nt el l i gent t ut or but in
vi ew of the i mpl ement at i on of this program in an i nformal l earni ng envi r onment ,
the t erm ' coach' , ori gi nat ed by Gol dst ei n (1982), appeared more congeni al t han
' t ut or' (Burton & Brown, 1977, 1982). WEST was a spin-off from the SOPHIE project;
hence, it is still in keepi ng wi t h the concept of a reactive l earni ng envi r onment
central to SOPHIE, but requires much si mpl er skills. WEST is also the first ever
comput er coach and it demonst rat es how a different emphasi s on di fferent compo-
nent s of the ITS (in this case the t ut ori ng modul e) can produce a radi cal l y di fferent
ITS, so radi cal t hat the t erm ' coach' is preferred to ' tutor' .
WEST si mul at es a board game requi ri ng players to travel in a series of moves. The
number of spaces for each move are det ermi ned by digits on three ' spi nners'
suppl i ed from a r andom- number generator. Players can combi ne these three digits
by usi ng any l egi t i mat e mat hemat i cal operat i on i ncl udi ng exponent i at i on, or by
usi ng negat i ve numbers, parent heses, etc. The game also has such features as short
cuts to the goal; opport uni t i es to ' bump' opponent s, forcing t hem to ret urn to the
begi nni ng, and spaces safe from bumpi ng. Al t hough two st udent s can pl ay against
each other, t hey t ypi cal l y pl ay against the machi ne. At each move, the st udent ' s
skill in wri t i ng algebraic equat i ons is compared to the expert' s sol ut i on for the same
skill. If the t wo sol ut i ons differ, the coach (tutoring component ) can i nt ervene and
provi de the st udent wi t h hel pful hi nt s as to how to i mprove hi s/ her game or make
better moves.
WEST' s expert knowl edge modul e comprises the si mul at ed board game and an
art i cul at e expert whi ch can moni t or and eval uat e the st udent ' s moves. The st udent
model l i ng t echni que is largely a si mpl er version of overlay model l i ng cal l ed dif-
ferential model l i ng. This is because, apart from out ri ght ari t hmet i cal errors, the
st udent ' s moves are never wrong; t hey are just poor. What is i mpor t ant is the
di fference in compar i son bet ween the expert' s move and the st udent ' s: hence the
word ' di fferent i al ' . WEST' s interface is si mpl e as its i nput s are mai nl y ari t hmet i cal
expressi ons i nvol vi ng integers (e.g. 1 + 2 2) or just pl ai n integers.
However, t he component t hat makes WEST radi cal l y di fferent from the previ ous
t hree syst ems di scussed is its t ut ori ng component . Its mai n strategy is to encourage
skill acqui si t i on and general probl em-sol vi ng abilities by engaging the st udent in
some game-like activity. In effect, the i mmedi at e aim is to have fun; skill acquisi-
t i on and l earni ng is an i ndi rect consequence.
Intelligent Tutoring Systems: an overvi ew 271
WEST has act ual l y been used in el ement ar y school cl assrooms. In a cont r ol l ed
exper i ment , a coached group exhi bi t ed ' a consi der abl y greater vari et y of pat t erns'
in t he expr essi ons t hey f or med and t hey even ' enj oyed pl ayi ng t he game consi der -
abl y mor e t han t he uncoached gr oup' (Burt on & Brown, 1982). These resul t s are
qui t e encouragi ng: t hey demonst r at e t hat t he coach succeeded in fost eri ng l ear ni ng
wi t hout any appar ent adverse effect on t he f un of t he game. Unf or t unat el y,
coachi ng cur r ent l y appear s onl y to be appl i cabl e to trivial domai ns; it is har d to see
how it coul d be used to t each, say, el ect r oni c t r oubl eshoot i ng or some non- obvi ous
f r act i on addi t i on probl em.
However , WEST' s i nf l uence on ITSs has been significant, and it is still a r ef er ence
for r esear cher s t oday. Regrettably, it seems it is no longer used. Wi t h so many
wor t hl ess comput er games avai l abl e on t he market, it is real l y unf or t unat e t hat a
pr ogr am like WEST, whi ch has act ual l y been demonst r at ed to be f unct i onal , shoul d
still r emai n a l aborat ory pr ot ot ype (Wenger, 1987).
7 The r a n g e of pr o t o t y pe ITSs
Thi s sect i on, as in Ross (1987), lists exampl es of ITSs that have been devel oped.
The list of ITSs and envi r onment s shown in Tabl e I bel ow is r easonabl y r epr esent a-
t i ve and pr ovi des an appr eci at i on of t he vi brant nat ur e of this new and i nt erest i ng
fi el d and also reveal s t he scope of al r eady const r uct ed systems. As also echoed by
Ross (1987) most of t hem do a lot less t han t he domai n i ndi cat i on suggests and are
also most l y exper i ment al as har dl y any have been t est ed on mor e t han a ver y few
peopl e. The list also reveal s that t he domai ns chosen by ITS researchers are mai nl y
few. Nevert hel ess, wi t h t he exper i ence bei ng cur r ent l y accr ued by researchers,
mor e compl ex domai ns are expect ed to be t ackl ed in fut ure research.
8 S o me v i e wp o i n t s
It is appr opr i at e to concl ude this paper wi t h some vi ewpoi nt s on some of t he
cont ent i ous i ssues in t he i nt el l i gent t ut or i ng domai n; it also has its own on-goi ng
debates. Two per t i nent ones are:
(1) is i nt el l i gent t ut or i ng just old wi ne in a new bottle, or is it a new vi nt age
( Ok- choon et al., 1987)?
(2) is i nt el l i gent t ut or i ng r eal l y possi bl e (Ridgway, 1988)?
Before pr ovi di ng r esponses to t hese quest i ons, it seems fair to r emar k t hat t hey
ari se in t he first pl ace because of publ i c opi ni on about t he AI ent er pr i se in general.
Bobr ow et al. (1986) cor r ect l y poi nt out t hat publ i c opi ni on about AI is schi zophr e-
nic, rangi ng f r om ' it wi l l never wor k' to ' it mi ght cost me my job' . Thi s range of
opi ni on refl ect s t he col l ect i ve conf usi on about AI. It appears t hat t he AI f r at er ni t y
onl y has itself to bl ame. Mi sl eadi ng publ i ci t y, most l y in order to at t ract grants, and
mi suse of f l amboyant expr essi ons like artificial i nt el l i gence, exper t syst ems, i nt el l i -
gent t ut or i ng syst ems, etc., have hel ped cont r i but e to unreal i st i c expect at i ons of t he
state-of-the-art. Consequent l y, quest i ons such as t hose above are bound to be asked.
272 H. S. N w a n a
Table 1. A reasonably comprehensive list of ITSs and environments
ITS Domain Reference
ACE/PSM
ATDS E
ARI THMEKI T
AL GEBRAL AND
BIP-I/BIP-II
BLOCKS Tutor
BRIDGE
BUGGY
DEB UGG Y
EDSMB
EUROHELP
EXCHECK
FGA
FI TS
FLOW Tutor
GEOMETRY Tutor
GERMAN Tutor
GUI DON I/II
I NTEGRATI ON Tutor
LISP Tutor
LMS
MA CS Y MA Advi s or
MAL T
MENO- Tut or
METEOROLOGY ITS
NEOMYCI N
PIXIE
PBOUST
QUADRATI C Tutor
QUEST
SCENT- 3 Advi s or
SCHOLAR
SI ERRA
SOPHI E I/II/RI
SPADE
SPI RI T
S TEAMER
TALUS
THEVENI N
TUTOR
WES T
WHY
WUS OR
NMR spectra interpretation
Basic subtraction
Basic subtraction
Algebraic proofs
Basic programming
Troubleshooting in a BLOCKS world
Programming
Basic subtraction
Basic subtraction
Basic multiplication
UNIX mail
Basic logic
Basic French grammar
Basic fractions addition
FLOW computer language
Geometry proofs
Basic German
Basic medical diagnosis
Basic integral calculus
Lisp programming
Basic algebra
Use of MA CS Y MA
Basic machine language programming
Basic Pascal programming
Basic meteorology
Medical diagnosis
Basic algebra
Pascal programming
Quadratic equations
Basic electrics
Li sp programming
South American geographical facts
Learning basic arithmetic procedures
Basic electronic troubleshooting
Basic LOGO programming
Probability theory
Marine steam propulsion plant
Basic Li sp programming
Basic electrical circuits
British highway code
Basic arithmetic skills
Basic meteorology
Maze game skills
Sleeman (1975)
Attisha & Yazdani (1983)
Brown (1983)
Brown (1985)
Barret al. (1976)
Brown & Burton (1978b)
Bonar (1985)
Brown & Burton (1978a)
Burton (1982)
Attisha & Yazdani (1984)
Breuker (1987)
Blaine (1982)
Barchan et al. (1986)
Nwana (1990)
Genter (1977)
Anderson et ol. (1985a)
Weischedel et al. (1978)
Clancey (1987)
Kimball (1982)
Anderson & Reiser (1985)
Sleeman & Smith (1981)
Genesereth (1982)
Koffman & Blount, 1975
Woolf & McDonald (1984)
Brown et al. (1973)
Clancey & Letsinger (1981)
Sleeman (1987)
Soloway & Johnson (1984)
O'Shea (1982)
White & Frederiksen (1985)
McCalla et al. (1988)
Carbonell (1970)
Vanlehn (1987)
Brown et al. (1982)
Goldstein & Miller (1976)
Barzilay (1985)
Hollan et al. (1984)
Murray (1987)
Joobbani & Talukdar (1985)
Davies et al. (1985)
Brown & Burton (1978b)
Collins & Stevens (1982)
Goldstein (1982)
In the first question, the old wine was basically the AFO-type CAI whi ch has been
observed to have taken a very naive view of the instructional process. It has since
been realized that providing a truly ' intelligent' tutor is a non-trivial task requiring
experts from several disciplines. ITSs combine at least some of the following: AI,
psychological models of the student and expert, and educational theory. There is
thus a substantial change in the quality of wine and, therefore, a new vintage.
Intelligent Tut or i ng Systems: an overvi ew 273
The second quest i on i nvest i gat es whet her i nt el l i gent t ut ori ng is real l y possi bl e.
Many answer s to such quest i ons in t he past have t ur ned out to be far t oo opt i mi st i c.
For exampl e, Suppes ' (1966) pr edi ct i on t hat ' i n a few mor e years mi l l i ons of school
chi l dr en wiI1 have access to what Phi l l i p of Macedon' s son had as royal prerogat i ve:
t he per sonal servi ces of a t ut or as wel l - i nf or med as Ari st ot l e. ' Clearly, besi des t he
obvi ous fi nanci al const r ai nt s f or bi ddi ng this and wi t hout much del i ber at i on on
what Suppes meant by ' few' , this pr edi ct i on is still far from bei ng achi eved a
quar t er of a cent ur y on. Hence, t he r esponse to this second quest i on requi res mor e
caut i on. The answer seems to depend on how t he quest i oner vi ews t he use of t he
wor d ' i nt el l i gent ' in AI terms. If he/ she ri ght l y vi ews it as st ri ct l y speaki ng a
mi snomer , at least for now, t hen i nt el l i gent t ut ori ng becomes ver y feasible, at least
i n var i ous l i mi t ed domai ns. On t he ot her hand, if t he quest i oner is mor e literal, as
many AI scept i cs in t he l i t erat ure seem to be (again pr obabl y due to t he mi sl eadi ng
publ i ci t y), t hen of cour se i nt el l i gent t ut ori ng woul d appear i mpossi bl e, and so
woul d AI in general. Nevert hel ess, t her e is a suspi ci on that no mat t er how ' i nt el -
l i gent l y' an ITS may event ual l y per f or m (e.g. even if one were demonst r abl y to
per f or m bet er t han most huma n tutors), t hese scept i cs woul d still not be satisfied,
mai nl y because ITSs are comput er - based. It t herefore appears t hat researchers will
event ual l y have to t ur n to t he famous Tur i ng test of i nt el l i gence for an answer when
ITSs come of age. Thi s is because Tur i ng' s test ci r cumvent s t he pr obl em of lack of
consensus on t he def i ni t i on of t he wor d ' i nt el l i gent ' : it regards i nt el l i gence as
undef i nabl e but ' i nt el l i gent behavi our ' as recogni zabl e.
However , one pr edi ct i on looks secure: that despi t e t hese cont r over si es, ITS
r esear ch will grow. Thi s is because, apart from t hei r pract i cal needs, t he area
appear s to pr ovi de an excel l ent t est -bed for t heori es from AI sci ent i st s, educat i onal
t heori st s and cogni t i ve psychol ogi st s (for i nst ance, it has been not ed t hat t he
Carnegi e Mel l on psychol ogi st John Ander s on came into t he area to test hi s psycho-
logical theories). Fur t her mor e, ITS r esear cher s t hemsel ves woul d be ver y i nt er est ed
in i ncor por at i ng any pr omi si ng r esear ch resul t s from vari ous AI/Cognitive Sci ence
subdomai ns, e.g. qual i t at i ve reasoni ng, pl anni ng, nat ural language under st andi ng,
h u ma n - c o mp u t e r i nt er act i on, etc., i nt o t hei r ITSs, and so t hey shoul d. In concl u-
sion, it appear s cer t ai n t hat t he best of ITS research is yet to come.
9 Co nc l us i o ns
Thi s paper has i nt r oduced t he AI s ubdomai n of i nt el l i gent t ut ori ng syst ems and t he
mot i vat i on of ITSs. It l ooked at its hi st ory, mot i vat i on as well as r evi ewed some
not abl e ITSs. The paper also reveal s how di fferent emphases on di fferent compo-
nent s of t he ITS (because of di fferent phi l osophi es adopt ed by vari ous researchers)
resul t in cont r ast i ng syst ems; it pr ovi des exampl es of di fferent ar chi t ect ur es to
s uppor t this. Thi s also refl ect s t he di fferent di sci pl i nes whi ch ITS spans. The paper
also lists a r easonabl y compr ehens i ve set of pr ot ot ype ITSs whi ch have been
devel oped to date. The paper draws to a concl usi on wi t h t he aut hor ' s vi ewpoi nt s on
t wo debat abl e i ssues in t he ITS domai n.
For f ur t her readi ng, see Sl eeman & Br own (1982a), O' Shea & Self (1983), Wenger
(1987) or Nwana (1989).
274 H. S . Nwa n a
Acknowl edgment s
T h e a u t h o r s i n c e r e l y a c k n o wl e d g e s t h e i n v a l u a b l e c o mme n t s of h i s s u p e r v i s o r , Dr
P e t e r C o x h e a d a n d of h i s e x a mi n e r , Dr P e t e r Ros s .
References
Anderson, J. R~ (1987) Pr oduct i on systems, learning and tutoring. In Production System Models of
Learning and Development (eds D. Klahr, P. Langley & R. Neches). MIT Press, London, pp.
437- 458.
Anderson, J. R., Boyle, D. G. & Reiser, B. ]. (1985a) Intelligent t ut ori ng systems. Science, 228,
456- 462.
Anderson, J. R., Boyle, D. F. & Yost, G. (1985b) The geometry tutor. In Proceedings of the 9th
International Joint Conference on Artificial Intelligence, Los Angeles, CA, pp. 1- 7.
Anderson, J. R. & Reiser, B. J. (1985) The lisp tutor. Byte, 10(4).
Attisha, M. G. & Yazdani, M. (1983) A mi cr o- comput er based tutor for t eachi ng ari t hmet i c skills.
Instructional Science, 12, 333- 342.
Attisha, M. G. & Yazdani, M. (1984) An expert system for di agnosi ng chi l dr en' s mul t i pl i cat i on
errors. Instructional Science, 13, 79- 92.
Barchan, J., Woodmansee, B. J. & Yazdani, M. (1986) A prlog-based tool for French grammar
analysis. Instructional Science, 14.
Barr, A., Beard, M. & At ki nson, R. C. (1976) The comput er as a tutorial laboratory: the Stanford BIP
Project. International lournal of Man-Machine Studies, 8, 567- 596.
Barr, A. & Feigenbaum, E. A. (1982) The Handbook of Artificial Intelligence, Vol, 2. Kaufmann, Los
Altos.
Barzilay, A. (1985) SPIRIT: a flexible t ut ori ng style in an i nt el l i gent tutoring system. In Artificial
Intelligence Applications: The Engineering of Knowledge-Based Systems (ed. R. C. Weisbin).
IEE Comput er Society, North Holland.
Blaine, L. H. (1982) EXCHECK. Handbook of Artificial Intelligence, Vol. 2 (eds A. Barr & E. A.
Feigenbaum). Addi son-Wesl ey, Reading, MA.
Bloom, B. S. 11984) The 2 Sigma Problem: the search for met hods of group i nst ruct i on as effective
as one-t o-one tutoring. Educational Researcher 13, 4- 16.
Bobrow, D. G., Mittal, S. & Steffik, M. (1986) Expert systems: perils and promise. Communications
of the ACM, 29, 880- 893.
Bonar, J. (1985) Understanding the Bugs of Novice Programmers. PhD Thesis, Dept of Comput er
and Informat i on Sci ence, Uni versi t y of Pittsburgh, Pittsburgh, PA.
Bonnet. A. (1985) Artificial Intelligence: Promise and Performance. Prent i ce Hall, London.
Breuker, J. (1987) Coachi ng in hel p systems. In Artificial Intelligence and Human Learning:
Intelligent Computer-aided Instruction (ed. J. A. Seft). Chapman & Hall, London.
Brown, J. S. (1985) Process versus product: a perspect i ve on tools for communal and i nformal
el ect roni c learning, lournal of Educational Computing Research 1, 179- 201.
Brown, J. S. & Burton, R. E. (1978a) Diagnostic model s for procedural bugs in basic mat hemat i cal
skills. Cognitive Science, 2, 155- 192.
Brown, J. S. & Burton, R. R. (1978b) A paradi gmat i c exampl e of an artificially i nt el l i gent instruc-
t i onal system. International Journal of Man-Machine Studies, 10, 323- 339.
Brown, J. S., Burton, R. R. & Bell, A. G. (1975) SOPHIE: a step towards a react i ve l earni ng
envi r onment . International/ournal of Man-Machine Studies, 7, 675- 696.
Brown, J. S., Burton, R. R. & de Kleer, J. (1982) Pedagogical, natural language, and knowl edge
engi neeri ng t echni ques in SOPHIE I, II and II1. In Intelligent Tutoring Systems (eds D. H.
Sl eeman & J. S. Brown). Academi c Press, London, pp. 227- 282.
Brown, J. S., Burton, R. R. & Zydel, F. (1973] A model - dr i ven quest i on-answeri ng system for
mi xed-i ni t i at i ve CAI. IEEE Transactions on Systems, Man, and Cybernetics, 3, 248- 257.
I nt el l i gent T u t o r i n g S y s t e ms : a n o v e r v i e w 275
Burns, H. L. & Capps, C. G. (1988) Foundat i ons of i nt el l i gent t ut ori ng systems: an i nt roduct i on. In
Foundations of Intelligent Tut ori ng Systems (eds M. C. Poison & J. J. Richardson). Lawrence
Erlbaum, London~ pp. 1- 19.
Burton, R. (1982) Di agnosi ng bugs in a si mpl e procedural skill. In Intelligent Tut ori ng Systems (eds
D. H. Sl eeman & J. S. Brown). Academi c Press, London, pp. 157-183.
Burton, R. & Brown, J. S. (1977) A tutoring and st udent model l i ng paradigm for gami ng envi ron-
ments. SIGCSE Bulletin, 8, 236- 246.
Burton, R. & Brown, J. S. (1982) An i nvest i gat i on of comput er coachi ng for informal learning
activities. In Intelligent Tut ori ng Systems (eds D. H. Sl eeman & ]. S. Brown). Academi c Press,
London, pp. 79- 98.
Carbonell, J. R. (1970) AI in CAI: an artificial i nt el l i gence approach to comput er-assi st ed instruc-
tion. IEEE Transactions on Man-Machine Systems, II, 190- 202.
Carbonel l , J. R. (1971) Artificial i nt el l i gence and large i nt eract i ve man comput er systems. Proceed*
ings of the Joint National Conference on Major Systems, Anahei n, CA, pp. 167- 173.
Clancey, W. J. (1982) Tut ori ng rules for gui di ng a case met hod dialogue. In Intelligent Tut ori ng
Systems (eds D. H. Sl eeman & J. S. Brown). Academi c Press, London, pp. 201- 225.
Clancey, W. J. (1983) GUIDON. Journal of Computer-Based Instruction, 10, 8- 15.
Clancey, W. J. (1984) Met hodol ogy for bui l di ng an intelligent t ut ori ng system. In Methods and
Tactics in Cognitive Science (eds W. Kintsch, J. R. Miller & P. G. Polson). Lawrence Erlbaum,
London.
Clancey, W. J. (1987) Knowledge-Based Tutoring. MIT Press, London.
Clancey, W. J. & Letsinger, R. (1981) NEOMYCIN: reconfi guri ng a rule-based expert system for
appl i cat i on to teaching. Proceedings of the 7th International Joint Conference on Artificial
Intelligence, Vancouver, Canada, pp. 829- 835.
Collins, A. & St evens, A. L. (1982) Goals and strategies for i nqui ry teachers. In Advances in
Instructional Psychology, Vol 2 (ed R. Glaser). Lawrence Erlbaum, Hillsdale, NJ.
Crowder, N. A. (1959) Aut omat i c tutoring by means of intrinsic programming. In Automatic
Teaching: The State of the Art, Wiley, New York, pp. 109- 116.
Davies, N. G., Dickens, S. L. & Ford, L. (1985) Tutor - - a prot ot ype ICAI system. In Research and
Development in Expert Systems (ed. M. A. Bramer). Cambridge Uni versi t y Press, Cambridge.
Elsom-Cook, M. (1987) Int el l i gent Comput er-ai ded i nst ruct i on research at t he Open Uni versi t y.
Technical Report No: 63. Comput er-Assi st ed Learning Research Group, The Open Uni versi t y,
Mi l t on Keynes.
Geneseret h, M. R. (1982) The role of plans in intelligent t eachi ng systems. In Intelligent Tut ori ng
Systems (eds D. H. Sl eeman & J. S. Brown). Academi c Press, London, pp. 137- 155.
Genter, D. R. (1977) The FLOW tutor: a schema-based tutorial system. Proceedings of the Fifth
International Joint Conference on Artificial Intelligence, Cambridge, MA 787- 790.
Gi l more, D. & Self, J. A. (1988) The appl i cat i on of machi ne learning to intelligent t ut ori ng systems.
In Artificial Intelligence and Human Learning: Intelligent Computer-Aided Instruction (ed.
J. A. Self). Chapman & Hall, London, pp. 179- 196.
Gol dst ei n, I. P. (1982) The genetic graph: a represent at i on for the evol ut i on of procedural know-
ledge. In Intelligent Tut ori ng Systems (eds D. H. Sl eeman & J. S. Brown). Academi c Press,
London, pp. 51- 77.
Gol dst ei n, I. P. & Miller, M. L. (1976) AI-based personal learning envi ronment s: di rect i ons far long
term research. AI lab Memo 384. Massachussetts Institute of Technol ogy, Cambridge, MA.
GrignettL M., Hausman, C. L. & Gould, L. (1975) An i nt el l i gent on-l i ne assistant and tutor:
NLS-SCHOLAR. Proceedings of the National Computer Conference, 775- 781.
Hartley, J. R. & Sl eeman, D. H. (1973) Towards more i nt el l i gent t eachi ng systems. International
Journal of Man-Machine Studies, 5, 215- 236.
Hasemann, K. (1981) On di ffi cul t i es wi t h fractions. Educational Studies in Mathematics, 12,
71- 287.
Hawkes, W. L, Sharon, J. D., Kandel, A. & Taps Project Staff (1986) Fuzzy expert systems for an
i nt el l i gent comput er-based tutor. Technical Report No: 86-5. Learning Syst ems Institute,
Centre for Educat i onal Technol ogy, Florida State University.
276 H. S . Nwa n a
Hol l an, [. D_ Hut chi ns , E. L. & Wei t zman, L. (1984) STEAMER: an i nt er act i ve i nspect abl e si mul a-
t i on- based t r ai ni ng syst em. A1 Magazine, 5, 1 5 - 2 7
Joobbani , R. & Tal ukdar , S. N. (1985) An exper t syst em for unde r s t a ndi ng expr essi ons for el ect ri c
ci r cui t s anal ysi s. Proceedings of the Ninth International Joi nt Conference on Artificial Intelli-
gence, Los Angel es, pp. 2 3 - 2 5 .
Ki mbal l , R. A. (1982) A sel f - i mpr ovi ng t ut or for symbol i c i nt egrat i on. In Intelligent Tutoring
Systems (eds D. H. Sl eeman & J. S. Brown). Academi c Press, London, pp. 283- 307.
Koffman, E. B. & Bl ount , S. E. (1975) Art i fi ci al i nt el l i gence and aut omat i c pr ogr ammi ng i n CA1.
Artificial Intelligence, 6, 215- 234.
Mandl , H. & Lesgold, A. (eds) (1988) Learning Issues for Intelligent Tutoring Systems. Spri nger-
Verlag~ London.
McCalla, G. I., Greer, J. E. & SCENT Team (1988) I nt el l i gent advi si ng i n pr obl em sol vi ng domai ns:
t he SCENT-3 ar chi t ect ur e. In Proceedings of the International Conference on Intelligent Tut or-
ing Syst ems, Mont r eal , Canada, pp. 124- 131.
Mur r ay, W. R. (1987) Aut omat i c program debuggi ng for i nt el l i gent t ut or i ng syst ems. Computation-
al Intelligence, 3(1/.
Nwana, H, S. I1989) An iterative-style approach to constructing intelligent tutoring syst ems i n
mathematics. PhD Thesi s, Ast on Uni ver si t y, Bi r mi ngham.
Nwana, H. S. C1990) The anat omy of FITS: a mat hemat i c t ut or. Intelligent Tut or i ng Media, 1(2).
Ok- choon, P., Ray, S. P, & Sei del , R, J. (19871 I nt el l i gent CAI: ol d wi ne i n new bot t l es or a new
vi nt age? In Artificial Intelligence and Instruction: Instruction and Methods, Addi son- Wesl ey,
Readi ng, MA, pp. 11- 43.
O' Shea, T. (19821 A sel f - i mpr ovi ng quadr at i c t ut or. In Intelligent Tut or i ng Systems (eds D. H.
Sl eeman & J. S. Brown). Academi c Press, London, pp. 283- 307.
O' Shea, T,, Bornat , R., Du Boul ay. B,, Ei senst adL M. & Page~ 1. (1984) Tool s for creat i ng i nt el l i gent
comput er tutors. In Artificial and Huma n Intelligence (eds ?. El i t hor n & R. Beneiji). El sevi er,
Nort h Hol l and, pp. 181- 199.
O' Shea, R. & Self, ]. (1983) Learning and Teaching with Computers. Har vest er Press, Sussex.
Papert , S. (1980) Mi ndst or ms: Children, Computers and Powerful Ideas. Basic Books, New York.
Peachey, D. R, & McCalla~ G. I. (1986) Usi ng pl a nni ng t echni ques i n i nt el l i gent t ut or i ng syst ems.
International ]ournal of Man-Machine Studies, 24, 77- 98.
Rich, E. (1979) User model l i ng vi a st ereot ypes. Cognitive Science, 3, 329- 354.
Ri dgway, J. (1988) Of cour se ICAI is i mpos s i bl e, . . worse t hough, it mi ght be sedi t i ous. In Artificial
Intelligence and Huma n Learning: Intelligent computer-aided instruction (ed. J. A. Self).
Cha pma n & Hall. London, pp. 28- 48.
Ross~ P. (1987) I nt el l i gent t ut or i ng syst ems, lournal of Computer Assisted Learning, 3, 194- 203.
Ross, P., Jones, 1. & Mi l l i ngt on, P. I1987) User model l i ng i n i nt el l i gent t eachi ng and t ut ori ng. In
Tr ends in Computer Assisted Instruction (eds R. Lewis & E. D. Tagg). Bl ackwel l , London, pp.
32- 44.
Se l l ], A. 11974) St udent model s i n comput er - ai ded i nst r uct i on. International Journal of Man-
Machine Studies, 6, 261- 276.
Self, ]. A. (1987a) The appl i cat i on of ma c hi ne l ear ni ng to s t udent model l i ng. In Artificial Intelli-
gence and Education 1: Learning Environments & Tut or i ng Syst ems (eds R. Lawl er & M. Yaz-
dani ). Abl ex, Norwood, pp. 267- 280.
Self, J, A. (1987b) Real i sm i n s t udent model l i ng. Alvey-IKBS Research Workshop Tut or i ng Sys-
tems. Uni ver si t y of Exeter.
Self, J. A. (ed.) [1988a) Artificial Intelligence a nd Huma n Learning: Intelligent computer-aided
instruction. Cha pma n & Hall, London.
Self, J. A. I1988b) St udent model s: what use are t hey? In Artificial Intelligence Tools i n Education
(eds P. Ercoli & R. Lewis/. Nor t h Hol l and, Amst er dam, pp. 73- 86.
Short l i ffe, E. H. (1976) Computer Based Medical Consultations: MYCIN. El sevi er, New York.
Ski nner , B, F. {1954) The sci ence of l ear ni ng and t he art of t eachi ng. Harvard Educational Review,
24, 86- 97,
I nt el l i gent T u t o r i n g Sy s t e ms : an o v e r v i e w 2 7 7
Ski nner, B, F. (1958) Teachi ng Machines. Science, 128, 969- 977.
Sl eeman, D. H. (1975) A probl em-sol vi ng moni t or for a deduct i ve reasoning task. International
Journal of Man-Machine Studies, 7, 183- 211.
Sl eeman, D. H. (1983) Int el l i gent t ut ori ng systems: a revi ew. Proceedings of EdCompCon '83
meeting. IEEE Comput er Society, pp. 95- 101.
Sl eeman, D. H. (1985) UMFE: a user model l i ng front-end subsystem. International Journal of
Man-Machine Studies, 23, 71- 88.
Sl eeman, D. H. (1987) PIXIE: a shell for devel opi ng i nt el l i gent tutoring systems. In Artificial
Intelligence and Education (eds R. Lawler & M. Yazdani). Ablex, Norwood, pp. 239- 265.
Sl eeman, D. H. & Brown, J. S. (eds) (1982a) Intelligent Tutoring Systems. Academi c Press, London.
Sl eeman, D. H. & Brown, J. S. (1982b) Int roduct i on: i nt el l i gent tutoring systems. In Intelligent
Tutoring Systems [eds D. H. Sl eeman & J. S. Brown). Academi c Press, London, pp. 1- 11.
Sl eeman, D. H. & Smi t h, M. 1. (1981) Model l i ng st udent s' probl em solving. Artificial Intelligence,
16, 171- 188.
Sol oway, E. & Johnson, W. (1984) Remembrance of bl unders past: a ret rospect i ve on the devel op-
ment of PROUST. Proceedings of the Sixth Cognitive Science Society Conference, Boulder,
CO, p. 57.
Streitz, N. A. (1988) Mental model s and metaphors: i mpl i cat i ons for the design of adapt i ve
user-system interfaces. In Learning Issues for Intelligent Tutoring Systems (eds H. Ma ndl &
A. Lesgold). Springer-Verlag, London, pp. 164-186.
Suppes, P. (19661 The uses of comput ers in educat i on. Scientific American, 25, 206- 221.
Suppes, P. (1967) Some t heoret i cal model s for mat hemat i cs learning, lournal of Research and
Development in Education 1, 5- 22.
Tobias, S. (1985) Comput er assisted instruction. In Adapting Instruction to Individual Differences
(eds M, C. Wang & H, J. Waldberg), McCutchan, Berkeley, CA, pp, 139- 159,
Uhr, L. (1969) Teachi ng machi ne programs that generate probl ems as a funct i on of i nt eract i on wi t h
students. Proceedings of the 24th National Conference, pp. 125-134.
Vanlehn, K. (1987) Learning one subprocedure per lesson. Artificial Intelligence 31, 1- 40.
Wachsmut h, I. (1988) Model l i ng the knowl edge base of mat hemat i cal learners: si t uat i on-speci fi c
and si t uat i on-nonspeci fi c knowl edge. In Learning Issues for Intelligent Tutoring Systems (eds
H. Mandl & A. Lesgold). Springer-Verlag, London, pp. 63- 79.
Wei schedel , R. M., Voge, W. M. & [ames, M. (19781 An artificial i nt el l i gence approach to language
i nst ruct i on. Artificial Intellige~me 10, 225- 240.
Wenger, E. (1987) Artificial Intelligence and Tutoring Systems. Morgan Kaufmann, Los Altos, CA.
Wexler, J. D. 11970) Information networks in generative computer.assisted instruction. IEEE Trans-
actions on Man-Machine Systems 11, 181- 190.
White, B. Y. & Frederi ksen, J. R. (1985) QUEST: qual i t at i ve underst andi ng of electrical system
t roubl eshoot i ng. ACM SIGART Newsletter, 93, 34- 37.
Woods, P. & Hartley, J. R. (19711 Some learning model s for arithmetic tasks and their use in
comput er-based learning. British Journal of Educational Psychology, 41. 35- 48.
Woolf, B. P. & McDonald, D. D. (1984) Cont ext -dependent transitions in t ut ori ng discourse.
Proceedings of the National Conference on Artificial Intelligence. Austin, Texas, pp. 355- 361.
Yazdani, M. (1983) Int roduct i on: artificial i nt el l i gence and educat i on. In New Horizons in Educa-
tional Computing led. M. Yazdani). Wiley, New York.
Yazdani, M. (1986) Intelligent t ut ori ng systems survey. Artificial Intelligence Review, 1, 43- 52.
Yazdani, M. (1987) Int el l i gence tutoring systems: an overvi ew. In Artificial Intelligence and
Education (eds R. Lawler & M. Yazdani). Ablex, Norwood, pp. 182-201,
Zissos, A. Y. & Witten, I. H. (1985) User model l i ng for a comput er coach: a case study. International
Journal of Man-Machine Studies, 23, 729- 750.