Journal of Intelligent Learning Systems and Applications, 2011, 3, 139-154
doi:10.4236/jilsa.2011.33016 Published Online August 2011 (http://www.scirp.org/journal/jilsa)
139
Identifying Causes Helps a Tutoring System to
Better Adapt to Learners during Training Sessions
Usef Faghihi1*, Philippe Fouriner-Viger1, Roger Nkambou1, Pierre Poirier2
1Department of Computer Science, Université du Québec à Montréal, Montréal, Canada; 2Cognitive Science Institute, Université du
Québec à Montréal, Montréal, Canada.
Received August 11th, 2010; revised June 7th, 2011; accepted June 14th, 2011.
ABSTRACT
This paper describes a computational model for the implementation of causal learning in cognit i ve agents. The Con-
scious Emotional Learning Tutoring System (CELTS) is able to provide dynamic fine-tuned assistance to users. The
integration of a Causa l Learning mechanism within CELTS a llows CELTS to first establish, through a mix of da tamin-
ing algorithms, gross user group models. CELTS then uses these models to find the cause of users' mistakes, evaluate
their performance, predict their future behavior, and, through a pedagogical knowledge mechanism, decide which tu-
toring intervention fits be st.
Keywords: Cognitive Agents, Computational Ca usal Modeling and Learning, Emotions
1. Introduction
Conscious Tutoring System (CTS, [1]) to which we
added Emotions (E) and Learning (L) is a general cogni-
tive architecture designed to be put to work as a Tutor for
astronauts learning to manipulate Canadarm2. Cana-
darm2 is a robotic arm installed on the International
Space Station's (ISS). ISS has been designed and imple-
mented to accommodate scientific experiments and life
in the space. Thus, it needs to be supplied constantly
with foods, fuel, inspections, etc. For instance, astronauts
may use Canadarm2 to charge or discharge the received
food from the space shuttles. Thus, manipulating Cana-
darm2 is a difficult task, which requires astronauts to
undergo a serious amount of training. The seven degrees
of freedom of the Canadarm2 is the first difficulty to
overcome, as it considerably increases the number of
possible operations. The second difficulty is sight limita-
tion. It is impossible to have an overall view of the sta-
tion; therefore, theastronaut can only see the arm through
a steady climb camera installed on the station and on the
Canadarm2. Furthermore, the astronaut must choose
among these cameras because there are only three
screens [2].
CELTS’ emotional mechanism simulates the peripher-
alcentral theory of emotions [2]. The peripheral-central
approach takes into account both the short and long route
of information processing and reactions, as in humans.
Both the short and long routes perform in a parallel and
complementary fashion in CELTS’ architecture. The
Emotional Mechanism (EM) learns and at the same time
contributes emotional valences (positive or negative) to
the description of the situation. It also contributes to the
decisions made and the learning achieved by the system.
CELTS’ Episodic Mechanism (EPM) simulates the
multiple-trace theory of memory consolidation [2]. The
multiple-trace theory postulates that every time an event
causes memory reactivation, a new trace for the activated
memory is created in the hippocampus. Memory con-
solidation occurs through the reoccurring loops of epi-
sodic memory traces in the hippocampus and the con-
struction of semantic memory traces in the cortex. Thus,
the cortical neurons continue to rely on the hippocampus
even after encoding. We used sequential pattern mining
algorithms to simulate this behavior of memory consoli-
dation in CELTS. To do so, every informationbroad-
casted in the system during a training session between
CELTS and learners is assigned to a specific time. Thus,
CELTS’ EPM extracts information from registered
events in the system. Given a problem, EPM is capable
of finding the best solution among different solutions
conceived by the expert in its BN [2].
One of CELTS’ most significant limitations in its cur-
rent implementation is its incapacity to find out why a
learner made a mistake-the causes of the mistake. To

Identifying Causes Helps a Tutoring System to Better Adapt to Learners during Training Sessions 140
address this issue, we propose to implement a Causal
Learning mechanism (CLM) in CELTS and combine it
with its existing Emotional Learning mechanism (see [3,
4] for more details).
In humans, the process of inductive reasoning stems in
part from activity in the left prefrontal cortex and the
amygdala; it is a multimodular process [5]. We base our
proposed improvements to CELTS’ architecture on this
same logic. CELTS’ modular and distributed organiza-
tion is ideal for the use of distinct mathematical methods
and algorithms that can be tailored to the specific re-
quirements of the Emotional Learning mechanism and
the newly integrated CLM. In humans, causal memory is
influenced by the information retained by the episodic
memory. Inversely, new experiences are influenced by
our causal memory [6-9].
Furthermore, causal learning is the process through
which we come to infer and memorize an event’s reasons
or causes based on previous beliefs and current experi-
ence that either confirm or invalidate previous beliefs
[10]. In the context of CELTS, we refer to Causal Learn-
ing as the use of inductive reasoning to generalize causal
rules from sets of experiences. CELTS’ observes learn-
ers’ behavior without complete information regarding the
reasons for their behavior. Our prediction is that, through
inductive reasoning, CELTS can infer the proper set of
causal relations from its observations of the learners’
behavior. It must be noted that CELTS in its Episodic
Learning mechanism uses sequential pattern mining al-
gorithms as a means for deductive reasoning. That is,
from the information exchanged between learners and
CELTS, the Episodic learning mechanism infers that if a
user forgets to adjust the camera before any Canadarm2
movement and chooses a bad joint, he or she will make a
collision risk on the ISS.
The goal of CELTS’ Causal Learning Mechanism
(CLM) is two-fold: 1) is to find causal relations between
events during training sessions in order to better assist
users; and 2) to implement partial procedural learning in
CELTS’ Behavior Network (BN)1, which is based on
Maes’ Behavior Network [11]. To implement CELTS’
CLM, we draw inspiration from Maldonado’s work [10]
that defines three hierarchical levels of causal learning: 1)
the lowest level, responsible for the memorization of task
execution; 2) the middle level, responsible for the com-
putation of retrieved information; and 3) the highest level,
responsible for the integration of this evidence with pre-
vious causal knowledge.
In the present paper, we begin with a brief review of
the existing work concerning the implementation of
Causal Learning in cognitive agents (Section 2). In sec-
tion 3, we propose our new architecture combining ele-
ments of the Emotional mechanism and Causal Learning,
focusing especially on the two-fold aspect of the causal
learning mechanism described above. Finally, we present
results from our experiments with CELTS.
2. Causal Learning Models and Their
Implementation in Cognitive Agents
Scientists propose causal Bayes nets (acyclic graphs) as
an alternative approach to establishing causal relation
between events. Different methods are proposed for
finding causal relations between events such as scientific
experiments, statistical relations, temporal order, prior
knowledge, and so forth [12].The key issue for the con-
struction of a causal Bayes net is finding conditional
probability between events. Mathematics is used to de-
scribe conditional and unconditional probabilities be-
tween a graph’s variables. The structure of a causal graph
restricts the conditional and unconditional probabilities
between the graph’s variables. We can find the restric-
tion between variables using the Causal Markov As-
sumption (CMA). The CMA suggests that every node in
an acyclic graph is conditionally independent of its as-
cendants, given the node’s parents (direct causes). For
instance, suppose one observes that each time one forgets
to adjust his car’s side and front mirrors (M), he tends to
have poor control over the wheel (W) and cause colli-
sions (C) with other cars. We can link these variables in
the following way: 1) M → W → C; and 2) W ← M →
C. The first graph shows that the probability of for- get-
ting mirror adjustment is independent of the probability
of making a collision with other cars, conditional on the
occurrence of poor wheel control. The second graph
demonstrates that the probability of poor wheel control is
independent of the probability of making a collision with
other cars and is conditional on forgetting mirror adjust-
ment. The CMA establishes such separation between
nodes to all acyclic graphs’ nodes. Thus, knowing a
graphs’s structure and the values of some of the variables,
we are capable of predicting the conditional probability
of other variables. Causal Bayes nets are also capable of
predicting the consequences of direct external interven-
tions on their nodes. When, for instance, an external in-
tervention occurs on a node (N), it must solely change its
value and not affect other node values in the graph ex-
cept through the node N’s influences. In conclusion, one
can generate a causal structure from sets of effects and
conversely predict sets of effects from a causal structure
[13].
To our knowledge, two research groups have at-
tempted to incorporate Causal Learning mechanisms in
their cognitive architecture. The first is Schoppek with
1CELTS’ Behavior Network (BN) (Figure 4(D)) is a high-level proce-
dural memory, a network of partial plans that analyses the context to
decide what to do, which behavior to set off.
Co pyright © 2011 SciRes. JILSA

Identifying Causes Helps a Tutoring System to Better Adapt to Learners during Training Sessions141
the ACT-R architecture [14], who has not included a role
for emotions in his causal learning and retrieval proc-
esses. ACT-R constructs the majority of its information
according to the I/O knowledge base method. It also uses
a sub-symbolic form of knowledge to produce associa-
tions between events. As explained by Schoppek [15]
causal learning in ACT-R occurs through implicit learn-
ing of information. To learn causes in ACT-R, sub-
symbolic knowledge applies its influence through activi-
tion processes that are inaccessible to production rules.
However, the causal model created by Schoppek in
ACT-R “overestimates discrimination between old and
new states” and every assumption for the creation of a
causal model must be detailed by a programmer. The
second is Sun [16], who proposed the CLARION archi-
tecture. In CLARION’s current version, during bottom-
up learning, the propositions (premises and actions) are
already present in top level (explicit) modules before the
learning process starts, and only the links between these
nodes emerge from the implicit level (rules). Thus, there
is no unsupervised causal learning for the new rules cre-
ated in CLARION [17].
As it is mentioned above, various causal learning
models have been proposed, such as Gopnik’s [18]. All
proposed models use a Bayesian approach for the con-
struction of knowledge. Bayesian networks work with
hidden and non-hidden data and learn with little data.
However, Bayesian networks need experts to assign pre-
defined values to variables, and this is often a very diffi-
cult and time-consuming task [19]. In the context of a
tutoring agent like CELTS, this is a serious issue, be-
cause we wish that CELTS could learn and adapt its
knowledge of causes automatically without any human
intervention. Another problem for Bayesian learning,
crucial in the present context, is the risk of combinatory
explosion in the case of large amounts of data. In the
case of our agent, constant interaction with learners cre-
ate the large amount of data stored in CELTS modules.
For this last reason, we believe that a combination of
sequential pattern mining (SPM) algorithms with asso-
ciation rules (AR) is more appropriate to implement a
causal learning mechanism in CELTS.
The other advantage of causal learning using the com-
bination of AR and SPM is that CELTS can then learn in
a real-time incremental manner—that is, the system can
update its information by interacting with various users.
A final reason for choosing the combination of AR and
SPM is that the aforementioned problem explained by
Schoppek, which occurs with ACT-R, cannot occur
when using association rules for causal learning. How-
ever, it must be noted that although data mining algo-
rithms learn faster than Bayesian networks when all data
is available, they have problems with hidden data. Fur-
thermore, like Bayesian learning, there is a need for ex-
perts, since the rules found by data mining algorithms
must be verified by a domain expert [19].
3. Causal Memory and Causal Learning in
CELTS’ Architecture
CELTS architecture relies on the functional “conscious-
ness” [20] mechanism for much of its operations. It also
bears some functional similarities with the physiology of
the nervous system. Its modules communicate with one
another by contributing information to its Working Me-
mory through information codelets2 [21] (see [22] for
more details). Before explaining CELTS’ Causal Learn-
ing, we will describe in the next subsection in detail our
causal model for cognitive architectures.
CELTS’ causal model takes into account the existence
of specific cognitive processes—be they associative or
causal.
CELTS’ Causal Learning takes place during its cogni-
tive cycles.A cognitive cycle starts by a perception and
usually ends with an action. It is conceived as an iterative,
cyclical, active process that allows interactions between
the different components of the architecture
After the information from the environment reaches
CELTS’ Perceptual mechanism (Figure 1), it is sent to
the Working memory (WM). CELTS’ WM (Figure 1) is
monitored by expectation codelets and other types of
codelets (see CELTS’ emotional mechanism for more
details [23]). If expectation codelets observe information
coming in WM confirming that the behaviors expected
result failed, then the failure brings CELTS’ Emotional
and Attention mechanisms back to that information. To
deal with the failure, emotional codelets that monitor
WM first send a portion of emotional valences sufficient
to get CELTS’ Attention mechanism to select informa-
tion about the failed result and bring it back to con-
sciousness. The influence of emotional codelets at this
point remains for the next cognitive cycles, until CELTS
finds a solution or has no remedy for the failure. Since
relevant resources need to be recruited to allow CELTS’
modules to analyze the cause of the failure and to allow
deliberation to take place concerning supplementary
and/or alternative actions, the consciousness mechanism
broadcasts this information to all modules. Among dif-
ferent modules inspecting the broadcasted information
by the consciousness mechanism, the Episodic and
Causal Learning mechanisms are also collaborating to
ind previous sequences of events that occurred before f
2Based on Hofstadter et al.’s idea, a codelet is a very simple agent, “a
small piece of code that is specialized for some comparatively simple
task”. Implementing Baars theory’s simple processors, codelets do
much of the processing in the architecture. In our case, each informa-
tion codelet possesses an activation value and an emotional valence
specific to each cognitive cycle.

Identifying Causes Helps a Tutoring System to Better Adapt to Learners during Training Sessions 142
Figure 1. CELTS’ architecture.
the failure of the action. These sequences of events are
the interactions that took place between CELTS and us-
ers during Canadarm2 manipulation by users in the vir-
tual world (Figures 2(B)-(D)). They are saved to differ-
ent CELTS’ Memories respecting the temporal ordering
of the events that occurred between users and CELTS.
The retrieved sequences of events contain the nodes
(Figures 2(B)-(D)). Each node contains at least an event
and an occurrence time (see CELTS’ Episodic Learning
[23] for more information). For instance, in Figure 2(D),
different interactions may occur between users and
CELTS depending on whether the nodes’ preconditions
in the Behavior Network (BN) become true. Through this
information, CELTS, using sequential pattern mining
algorithms, extracts useful information. For example, if a
user forgets to adjust the camera before he or she moves
Canadarm2 and consequently chooses an incorrect joint,
then the user will make collision risk. Such aforemen-
tioned information is gained through deductive reasoning
in CELTS. Given such information, we were interested
in finding the causes of the problem produced by the
users in the virtual world. To do so, from all past events,
the Causal Learning mechanism (CLM) constantly ex-
tracts association rules (e.g. X → Y) between sets of
events with their confidence and support3 [24]. These
rules indicate the groups of events that are frequently
associated to other groups of events. From these associa-
tion rules, CLM then eliminates the association rules that
do not meet a minimum confidence and support accord-
ing to the temporal ordering of events within a given
time interval. This eliminates a large amount of non-
causal rules from the retrieved sequences of events. After
finding the candidate rule as the cause of the failure,
ELTS’ CLM re-executes it and waits for the user feed- C
3Given a transaction database D, defined as a set of transactions T = {t1
t2,, tn} and a set of items I = {i1, i2,, in}, where t1, t2,, tnI,
the support of an itemset XI for a database is denoted as sup(X) and
is calculated as the number of transactions that contain X. The support
of a rule X  Y is defined as sup




XYT. The confidence of a rule
is defined as conf (X  Y) = sup


XY/sup(X).

Identifying Causes Helps a Tutoring System to Better Adapt to Learners during Training Sessions143
Figure 2. Virtual world simulator of Canadarm2 and CELTS’ BN.

Identifying Causes Helps a Tutoring System to Better Adapt to Learners during Training Sessions 144
back. However, if after the execution of the candidate
rule it turns out that the rule did not help the user to solve
the problem, then CELTS’ CLM writes in a failure in the
WM. The failure leads CELTS’ Causal Learning to ex-
amine other related nodes to the current failure with the
highest support and confidence. Each time a new node is
proposed by Causal Learning and executed by BN, an
expectation node brings back to the consciousness
mechanism the confirmation from users to make sure
that the found rule was the cause of the failure. Finally, if
a new cause is found, it will be integrated in CELTS’
Causal Memory. In the end, if o solution can be found,
the Causal Learning mechanism puts the following mes-
sage in WM: “I have no solution for this problem.”
This way of finding causal relations between different
events is in accordance with what is suggested at the onset
of this document such as statistical relations, temporal
order, and prior knowledge.
After having proposed our causal model for CELTS,
we now explain in detail the intervention of the causal
process in CELTS’ cognitive cycles. It is important to
remember that two routes are possible during CELTS’
cognitive cycle—a short route4 (no causal learning oc-
curs in this route) and a long route (various types of
learning occur in this route such as episodic, causal and
procedural). In both cases, the cycle begins with the per-
ceptual mechanism. Hereafter, we briefly summarize
each step in the cycle and in italics describe the influence
of emotions (here called pseudo-amygdala5 or EM and/
or of CLM).
For a visual representation of the process, please refer
to Figure 1.
Step 1: The first stage of the cognitive cycle is to
perceive the environment, that is, to recognize and
interpret the stimulus (see [1] for more information)6.
EM: All incoming information is evaluated by the
Emotional Mechanism when low-level features recog-
nized by the perceptual mechanism are relayed to the
emotional codelets, which in turn feed activation to emo-
tional nodes in the Behavior Network (BN). Strong reac-
tions from the “pseudo-amygda la” may cause an imme-
diate reflex reaction in CELTS [23,25].
Step 2: The percept enters Working Memory
(WM): The percept is brought into WM as a network
of information codelets that covers the many aspects
of the situation (see [1] for more information).
CLM: CLM also inspects and fetches WM relevant in-
formation. Relevant traces from different memories are
automatically retrieved. These will be sequences of events
in the form of a list relevant to th e new information. The
sequences include the c urrent event, its relevant rules and
the residual information from previous cognitive cycles in
WM. The retrieved traces contain codelet links with other
codelets. Each time new information codelets enter WM,
the memory traces are updated depending on the new
links created between these traces and the new informa-
tion codelets. Once information is enriched, CLM sends it
back to the WM.
Step 3: Memories are probed and other uncon-
scious resources contribute: All these resources react
to the last few consciousness broadcasts (internal
processing may take more than one single cognitive
cycle).
Step 4: Coalitions assemble: In the reasoning phase,
coalitions of information are formed or enriched. At-
tention codelets join specific coalitions and help them
compete with other coalitions toward entering “con-
sciousness.”
EM: Emotional codelets observe the WM’s content,
trying to detect and instill energy to codelets believed to
require it and attach a corresponding emotional tag. As
a result, emotions influence which information comes to
consciousness and modulate what will be explicitly
memorized.
Step 5: The selected coalition is broadcasted: The
Attention mechanism spots the most energetic coali-
tion in WM and submits it to the “access conscious-
ness,” which broadcasts it to the whole system. With
this broadcast, any subsystem (appropriate module or
team of codelets) that recognizes the information may
react to it.
CLM: CLM starts by retrieving the past frequently re-
appearing information that best matches the current in-
formation resident in WM, ignoring their temporal part.
This occurs by constantly extracting associated rules
from the broadcasted information and the list of events
previously consolidated. Then, CLM eliminates the rules
that do not meet the temporal ordering of events.
Steps 6 and 7: Unconscious behavioral resources
(action selection) are recruited. Among the modules
that react to broadcasts is the Behavior Network (BN):
BN plans actions and, by an emergent selection proc-
ess, decides upon the most appropriate act to adopt.
The selected Behavior then sends away the behavior
codelets linked to it.
4The short route is a percept-reaction direct process, which takes place
when the information received by the perceptual mechanism is strongly
evaluated by the pseudo-amygdala. The short route is described else-
where see [3,4]. The long route is CELTS’ full cognitive cycle.
5Let us note that in CELTS, a “pseudo-amygdala” is responsible fo
r
emotional reactions [2].
6The following steps, in bold characters, describe CELTS’ full cogni-
tive cycle. They are the same as in [2]. We restate them here; the dif-
ference is in what occurs during those steps from the causal learning
p
erspective, which is in italics.
EM: When CELTS’ BN starts a deliberation, for in-
stance to build a plan, the plan is emotionally evaluated
Co pyright © 2011 SciRes. JILSA

Identifying Causes Helps a Tutoring System to Better Adapt to Learners during Training Sessions145
as it is built, the emotions playing a role in the selection
of the steps. If the looping concerns the evaluation of a
hypothesis, it gives it an emotional evaluation, perhaps
from learned lessons from past exp eriences.
CLM: The extraction of the ru les in step 5, ma y invo ke
a stream of behaviors related to the current event, with
activation passing through the links between them (Fig-
ure 2(D)). At this point CLM waits for CELTS’ Behavior
Network and CELTS’ Episodic Learning Mechanism
solution for the ongoing situation) [23]. Then, CLM puts
its proposition as a solution in CELTS’ WM, if the pro-
positions from the decision making and the episodic
learning mechanisms are not energetic enough to be
chosen by CELTS’ Attention Mechanism.
Step 8: Action execution: Motor codelets stimulate
the appropriate muscles or internal processes.
EM: Emotions influence the execution, for instance in
the speed and the amplitude of the movements.
CLM: The stream of behaviors activated in CELTS’
BN (Step 7) may receive inhibitory energies, from CLM,
for some of their particular behaviors. This means that,
according to CELTS’ experiences, CLM may use a
shortcut (i.e. eliminate some intermediate nodes) be-
tween two nodes in behavior Network (BN) to achieve a
goal (e.g. in Figure (2D) two points v and z). In some
cases, again according to CELTS’ experiences, CLM
may prevent the execution of unnecessary behaviors in
CELTS’ BN during the execution of a stream of behav-
iors.
4. The Causal Learning Process
The following subsections explain in detail the three
phases of the Causal Learning mechanism as it is imple-
mented in CELTS’ architecture.
4.1. The Memory Consolidation Process
The Causal Memory consolidation process, which occurs
in the step 2 of CELTS’ cognitive cycle, takes place
during each CELTS’s cognitive cycle. CELTS’ Causal
Learning mechanism (CLM) extracts frequently occur-
ring events from its past experiences, as they were re-
corded in its different memories [26]. In our context,
CELTS extracts sequences of events during training ses-
sions for Canadarm2 manipulation by astronauts inthe
virtual world [27] (Figure 2(A)).To do so, a trace of
what occurred in the system is recorded in CELTS’ dif-
ferent memories during consciousness broadcasts [23].
For instance, each event X = (ti, Ai) in CELTS repre-
sents what happened during a cognitive cycle. The time-
stamp ti of an event indicates the cognitive cycle number.
The set of items Ai of an event contains an item that
represents the coalition of information codelets (see Step
4 of CELTS’ cognitive cycle) that were broadcasted
during the cognitive cycle.
For example, one partial sequence recorded during our
experimentations was (t = 1, c2), (t = 2, c4). This se-
quence shows that during cognitive cycle 1 the coalition
c2 (indicating that user forgot to adjust camera in the
virtual world) was broadcasted, followed by the broad-
cast of c4 (indicating that user brought about an immi-
nent collision in the virtual world). If this subsequence
appears several times during interactions of learners with
CELTS, the following rule could be discovered: {Forget-
camera adjustment (F), Bad joint (B)  {Collision
risk(C)}.
4.2. Learning by Extracting Rules from What is
Broadcasted in CELTS
The second phase of Causal learning, which occurs in
Step 5 of CELTS’ cognitive cycle, deals with mining
rules from the sequences of events recorded for all of
CELTS’ executions. To do so, the algorithm presented in
takes as input the sequence database which contains se-
quences of coalitions that were broadcasted for each
execution of CELTS, minsup, minconf and User Trace
which are the traces of what occurs between the current
user and CELTS. CELTS’ uses the first three parameters
to discover the set of all causal rules (R1, R2,, Rn) con-
tained in the database (Figure 3 Step 1). It then tries to
inspect rules that match with the interactions between the
current user and CELTS (User trace) in order to dis-
cover probable causes that could explain the user’s be-
havior (Step 2). When CELTS does, one cause is re-
turned.
The algorithm (Figure 3) performs as follows. 1) In
STEP1, it saves in a sequence database the sequences of
nodes (the coalitions) that are broadcasted by CELTS’
Behavior Network (BN) during interactions with users to
solve a problem. Then, in STEP2, the algorithm uses the
Apriori algorithm [23] for mining association rules be-
tween nodes. This uncovers association rules of the form
Ri: NODEi  NODEf, where NODEf and NODEi are
potential causes and effects of the failure. The meaning
of an association rule Ri is that if NODEf appears, we are
likely to also find NODEi in the same sequence. But
NODEi can appear before or after NODEf. For this rea-
son, the algorithm reads the original sequence database
one more time to eliminate rules that do not respect the
temporal ordering of the events.
To do this, we use two user-defined thresholds that a
rule should meet in order to be kept. These thresholds are
called minimal causal support and minimal causal confi-
dence. Let s be the number of sequences in the sequence

Identifying Causes Helps a Tutoring System to Better Adapt to Learners during Training Sessions 146
Figure 3. Causal Learning algorithm.
database. The causal support and confidence of a rule are
defined respectively as sup(NODEi  NODEf)/s and sup
(NODEi  NODEf)/sup(NODEi), where sup(X  Y)
denotes the number of sequences such that NODEi ap-
pears before () NODEf and sup(NODEi) represents the
number of sequences containing NODEi (see [28] for
more details).
After eliminating the association rules that do not meet
the minimum causal support and confidence thresholds,
the set of rules that is kept is the set of all causal rules. A
causal rule NODEi  NODEf is interpreted as: if NODEi
occurs, then NODEf is likely to occur thereafter. In that
case we will call NODEi the cause of the failure and
NODEf the effect. Note that this definition can be ex-
tended for rules where the left part and the right part can
contain more than one node. We here only provide ex-
amples with rules between two nodes. In Step 2 of the
algorithm, CLM tries to select the more likely cause for a
failure. To do so, this algorithm first sets a variable
named MaxCE to zero. It then computes the causal esti-
mation (CE) for each rule that matches with User Trace
by multiplying its support and confidence.
Causal estimation (CE) of Ri
= (support of Ri × confidence of Ri)
This calculation is done for each rule to determine
which node is the most likely to be the cause (the left
part of the rule having the highest CE). The CE of a rule
represents the causal estimation of a rule according to all
the information that broadcasts in the system. For each
such rule r, if the CE is higher than MaxCE, MaxCEis
set to that new value and a variable candidate coalition is
set to the right part of r. If the CE is lower than MaxCE,
MaxCE remains intact. Finally, when the algorithm fin-
ishes iterating over the set of rules, the algorithm returns
to CELTS’ working memory the node (coalition) Candi-
dateCoalitions contained in the right part of the rule
which has the highest CE value.
Using this method for each node of the retrieved se-
quence, CELTS’ CLM finds the most probable causes of
the problem produced by the user while manipulating
Canadarm2 in the virtual world. This node (coalition)
will be broadcasted next by CELTS’ consciousness
mechanism to the user for further confirmation (see the
next subsection for detail).
4.3. Construction of CELTS’ Causal Memory
The creation of CELTS’ Causal Memory (CM) occurs in
Steps 7 and 8 of the cognitive cycle. The main elements
of Causal Memory are the rules of the form X  Y. Like
CELTS’ Behavior Network (BN), the rules’ left and right
parts are nodes which are the coalitions broadcasted dur-
ing CELTS’ interactions with users. Each rule has a
support and a confidence (used to calculate the CE, as
described in the previous subsection).
Each new node (such as NODEp) includes a context,
an action, a result, and one or more causes. The context
in this newly created node describes an ongoing event.
The left part of the rule is filled by the node that caused
the failure. The right part of the rule is considered as the
effect. In what follows, we explain in detail how causal
memory is formed. The algorithm is presented in Figure
4. It takes as parameters the sequence database, the
Co pyright © 2011 SciRes. JILSA

Identifying Causes Helps a Tutoring System to Better Adapt to Learners during Training Sessions147
Figure 4. Causal Memory Construction Algorithm.
maximum causal estimated node (MaxCE) calculated in
the previous section, and the NODEf which is brought
about by the user’s error. Given the node NODEf that
caused the error after execution by CELTS’ BN, CLM
creates (see Figure 4. Step 1) an empty rule (R) in
CELTS’ Casual Memory (CM) and copies the informa-
tion in NODEf into the right part of the rule. During a
user’s manipulation of Canadarm2, CLM finds, from the
current sequence of executed nodes, the node NODEp
executed prior to NODEf which caused the user’s error.
It then attaches an expectation codelet to node NODEp,
puts it into the WM to be executed by BN and waits for
the user’s confirmation to find the cause of the problem.
If the cause of the failure is NODEp, CLM copies the
action of node NODEp into the cause of the node NODEf.
CLM then copies the information of NODEp into the left
part of the created rule R in CELTS’ Causal Memory and
makes a direct link between NODEf and NODEp.
If, however, it turns out that the node NODEp in the
previous step is not the cause of the error, then CLM
(Figure 4. Step 2) searches for the node NODEn with the
next highest CE value (MaxCE, explained in the pre-
vious subsection). It then, attaches an expectation codelet
to it, puts it into the WM to be executed by BN and waits
for the user confirmation. If the cause of the error is
NODEn, CLM copies its action to the NODEf’s cause
and all information into the left part of the created rule R
in CELTS’ CM. Finally to save the traces of what was
done to find the cause, 1) CLM creates a sequence of
empty nodes similar to what is retrieved as the sequences
of executed nodes from CELTS’ different memories, 2)
assigns NODEn to its first node and NODEf to the last
node and 3) copies to the sequence created in CM all
intermediate nodes between NODEn and NODEf, and
then creates links between them. The nodes NODEn and
NODEf in this sequence are tagged as the cause and ef-
fect of the problem that caused the error.
However, if, in the execution of the node NODEn in
the previous step, the resulting information brought back
by the expectation codelet to WM does not meet the ex-
pected results, CLM then (Figure 4. Step 3) repeatedly
searches for the node of the sequence from NODEn-1 to
NODE1 with the highest CE value but less than the
NODEn’s CE value and pursues the same previous proc-
esses as explained in steps one and two to find the cause
of the error. This process will continue for the remaining
nodes retrieved from CELTS’ LTM if each attempt fails.
If CELTS cannot find any cause, the message “I cannot
find the cause of the problem” is shown.
4.4. Using Mined Patterns to Improve CELTS’
Behavior
The third part of CELTS’ Causal Learning occurs in Step

Identifying Causes Helps a Tutoring System to Better Adapt to Learners during Training Sessions 148
7 and Step 8 of CELTS’ cognitive cycle. It consists of
improving CELTS’ behavior by making it reuse found
rules to predict why users are making mistakes, deter-
mine how to best help them, and in some specific cases,
to reconstruct the Causal Memory (CM). Finding causes
will directly influence the actions that will be taken by
CELTS’ Behavior Network (BN). CELTS’ behavior will
improve due to the fact that the more it interacts with
users and they confirm the correctness of the found
causes for their mistakes or not, the more the estimated
CE values for the nodes in the rules get reinforced or
weakened. After some interactions between CELTS and
the users, CLM may find for instance a chain of interre-
lated nodes. For instance, in Figure 2(D), the node V is
usually in relation with node Y and node Y is in relation
with node Z, according to user confirmations and the
minimum support and confidence defined by the domain
expert. For instance, CLM learned after several interac-
tions with users that 60 % of the time “user chose the
wrong joints → user makes Canadarm2 pass too close to
the ISS”. This means that after a while CELTS’ CLM is
capable of jumping from a start point in the BN to a goal
and eliminates unnecessary nodes between them.
However, jumping from one point to a goal point in
the BN is not always a good decision as CELTS is a tutor
and some intermediate nodes are very important hints to
users. To solve this problem, in the first step, we tagged
the important nodes in the BN as not to be eliminated.
Thus, some experiments go from one point to the other
(for instance in Figure 2. Dnodes V → Z), CELTS’
CLM makes an obligatory passage through intermediate
nodes such as node Y and eliminates only unnecessary
nodes between them. In the second step, to automatically
eliminate unnecessary nodes that have not been pre-
tagged by a human expert, we used the aforementioned
algorithms (previous subsection Figure 4. Step 2 and
Step 3) for finding causes when the users make an error
while interacting with CELTS. This means that to
achieve a goal from a start point in the BN, according to
CELTS’ experiences with users, CLM must decide to
preserve important nodes and only eliminate those that
are unnecessary in the BN (e.g. Figure 2(D) two points v
and z).
Finally, it is worth noting that CELTS’ BN is an
acyclic graph. The Causal Markov Assumption (CMA)
postulates that for any variable X, X is conditionally in-
dependent of all other variables in an acyclic causal
graph (except for its own direct and indirect effects)
based on its own direct causes. Accordingly, the refined
BN produced by CLM could be considered a primitive
proposition for the construction of a causal Bayesian
network.
For instance, like the cars’ side and front mirrors ex-
ample given above, after several interactions with users
rules are extracted by the algorithms: 1) Forgetting cam-
era adjustment (F) → Choosing Bad joint (B) → colli-
sion risk (C); 2) Choosing bad joint (B) ← Forgetting
camera adjustment (F) → collision risk (C). If we as-
sume that CMA holds, both structures in our example
entail exactly the same conditional and unconditional
inde- pendent relationships: In both, F, B and C are de-
pendent and F and C are independent conditional on B
[18]. Moreover, an important difference between the BN
and Bayesian Networks is that no node in the BN pos-
sesses a table of conditional probabilities like nodes of
Bayesian networks do. Instead, the information about
probabilities is stored in each causal rule as the causal
confidence and causal support values which can be in-
terpreted as an estimate of the probability P(Y|X)
[29-31].
5. Testing Causal Learning in the New
CELTS
To validate CELTS' Causal Learning mechanism (CLM),
we integrated it into Canadarm2 simulator, our simulator
designed to train astronauts to manipulate Canadarm2
(Figure 2(A)). Users were invited to perform arm ma-
nipulations using the simulator. In these experiments,
users had to move Canadarm2 from one configuration to
another in the virtual world while avoiding collisions
between Canadarm2 and the space station. This is a com-
plex task, as Canadarm2 has seven joints and the user
must 1) choose the best three cameras (from a set of
about twelve cameras on the space station) for viewing
the environment (since no camera offers a global view of
the environment), 2) not move Canadarm2 too close to
the ISS, 3) choose the right joint for the arm movements,
and 4) adjust parameters of cameras properly. These ex-
periments sought to validate CELTS’ ability to find the
causes of mistakes made by users. During these experi-
ments, we observed that CELTS was able to find the
causes and propose appropriate hints to help users. Some
experiments are described next.
Users’ Learning Situations
A user learns by practicing Canadarm2 manipulations
while receiving hints created initially by an expert and
given to the user by CELTS. We performed more than
300 CELTS executions of Canadarm2 in the virtual
world including good moves and dangerous moves, such
as collisions. During each execution, CELTS chose a
tutoring scenario depending on the situation. Our ex-
periments showed that CELTS is often capable of find-
ing the right causes of problems created by users in dif-
ferent situations. In what follows, two different experi-
ments are detailed.
Experiment 1: Approximate Problem
Co pyright © 2011 SciRes. JILSA

Identifying Causes Helps a Tutoring System to Better Adapt to Learners during Training Sessions149
When manipulating Canadarm2, it is important for the
users to know the exact distance between Canadarm2
and ISS at all times. This prevents future collisions or
collision risks on the ISS. Figure 2(D) shows the sce-
nario created by an expert in the CELTS’ Behavior Net-
work (BN). This scenario is an intervention by CELTS to
help the user while manipulating Canadaram2 to avoid
collisions between Canadaram2 and ISS. The user
weakly estimated the distance between Canadaram2and
ISS because 1) the user chose to move the wrong joint; 2)
the user was tired; 3) the user did not remember his
course; 4) the user has never passed through this zone.
As we can see in the Figure 2(D), this is a long sce-
nario and each time, to find the cause of mistakes made
by the user, CELTS may be required to interact for a
long period of time (e.g. asking questions, giving hints,
and demonstrating some examples) to find the causes
and provide appropriate feedback. The scenario starts
when CELTS detects that a user has chosen the wrong
joint and is moving Canadarm2 too close to the ISS.
CELTS first prompts the following message: Have you
ever passed through this zone? 1) If the answer given by
the user is yes, CELTS asks the user to verify the name
of the joint that she has selected. If the user fails to an-
swer correctly, CELTS proposes a hint in the form of a
demonstration or it stops Canaradm2 manipulation. In
this case, the user needs to revise the course before start-
ing Canaradm2 manipulation again; 2) if the user’s an-
swer is no, CELTS asks her to estimate the distance be-
tween Canadarm2 and ISS. If the user fails to answer
correctly then the next hint from CELTS asks the user if
she is tired or forgot the course about this zone or if he or
she needs some help; if the user answers correctly, it
means that the user is an expert user and that the situa-
tion is not dangerous.
Interacting with various users and according to the us-
ers’ answers, CELTS found the following rule 1) 60 % of
the time “user chose the wrong joints → user makes Ca-
naradm2 pass too close to the ISS”; 2) in 35% of the time
“user has never passed through this zone → user ma-
nipulates near to the ISS”, Figure 2(D); 3) in 5% of the
time “user is an expert → user makes Canaradm2 pass
too close to the ISS”.
It must be noted that the percentage value attributed to
the extracted rules varies depending on the users’ an-
swers to CELTS’ questions.
Experiment 2: Camera Adjustment Problem
As explained above, forgetting to adjust the camera prior
to moving Canadarm2 increases collision risk (as de-
picted in Figure 2(A)). During our experiments, we
noted that users frequently forgot this step, and moreover,
users frequently did not realize that they had neglected
this step. This increases the risk of collisions (as depicted
in Figure 2(A)) in the virtual world. We thus decided to
implement this situation as a medium-threat situation in
CELTS’ BN (e.g. Figure 2(D)).
When a user forgot to perform camera adjustments,
CELTS had to make a decision; it could either 1) give a
direct solution such as “You must stop the arm immedi-
ately” or 2) give a brief hint such as “I think this move-
ment may cause some problems. Am I wrong or right?”
or 3) give a proposition such as “Stop moving Cana-
darm2 and revise your lessons”. Through interactions
with different users, CELTS recorded sequences of
events, each of them carrying emotional valences (see
[2] for more details).
From the interactions that occurred between CELTS
and users to solve camera adjustment, CLM drew the
following conclusions: 1) 60% of the time, “the user is
tired → the user performs a camera adjustment error”; 2)
30% of the time, “the user has forgotten this lesson →
the user performs a camera adjustment error” and 3) 10%
of the time, “the user lacks motivation → the user is in-
active”.
After some trials, CELTS’ CLM is capable of induc-
ing (by jumping from one point to another point in the
BN, Figure 2(D)) the source of the users’ mistakes and
proposing a solution for them in the virtual world. How-
ever, given that CELTS is a tutor and must interact with
the user, jumping from the start point to the end of the
scenario (Figure 2(D), V → Z) causes the elimination of
some important steps in the BN. To prevent this, as men-
tioned before, we tagged the important nodes in the BN
as not to be eliminated. Thus, after some experiments, to
go from V → Z, CLM obligatorily passed through inter-
mediate nodes such as node Y in Figure 2(D) We call
this process as CELTS’ partial Procedural Learning (Step
8 of CELTS’ cognitive cycle).
Experiment 3: Complex Situation
To evaluate the extent of CELTS’ capabilities when
equipped with CLM, we decided to examine a very com-
plex path in the virtual world. We considered an exercise
between two ISS modules, JEMEF01 (labeled and is
referred as A) and MPLM02 (labeled by red cube and is
referred to as END) in the virtual world (as shown in
Figure 5(A)) in which users’ mistakes while moving
Canadarm2 from configuration A to END are very likely.
As shown in Figure 5(A), Canadarm2 is very close to
configuration A. Thus, the exercise starts near to the
module A and finishes at module END. In the first step
of this experiment, the user has handled the collision risk
problem with the configuration A. In the second step, the
user faces at least four paths, from configuration A to
END (Figure 5 (A)). Importantly, the expert system has

Identifying Causes Helps a Tutoring System to Better Adapt to Learners during Training Sessions 150
Figure 2. Simulation of the International Space Station
(ISS).
conceived only three scenarios in the BN regarding only
three paths with their corresponding obstacles to be
avoided: P1 (AECDH), P2 (AEBCDH) and P3 (AEFGHD)
(Figure 5(A)).
Whichever paths are chosen by the users, obstacles A,
E, B, C, D, H, G, and F have to be avoided in the virtual
world to prevent any collision. Therefore, the nodes in
the BN corresponding to those obstacles in the virtual
world are marked as “Not to be eliminated”, by the do-
main expert.
The domain expert marks Configurations A, C and D
as very important for the paths P1 and P2. Thus, in con-
figuration C, in order to go through them without causing
any collision risk, the user must first rotate camera8 60
degrees horizontally (Figure 5) and then choose the spe-
cific joint EP and then joint SP (Figure 5). In configure-
tion D, the user must first adjust camera6 in order to
have a good view of obstacles G and H before perform-
ing any movements. In path P3, the user must respect the
following steps to prevent collision risk while manipu-
lating Canarm2 from the configuration A to the END.
First, in configuration E, camera2 must be turned 30 de-
grees, and the manipulation must then be continued using
joint SR. Then, in configuration F, the joint SY must be
selected and rotated 90 degrees to prevent any collision
with ISS. In configuration G, the obstacle H must be
avoided by rotating Canadarm2 60 degrees.
It must be noted that in “Part One” of this experiment,
some specific nodes in CELTS’ BN (Figure 5) are
marked as “not to be eliminated,” since our purpose here
is to examine CELTS’ capacity to find the best scenario
among different solutions given by the expert. And in
“Part two” of this experiment, nodes in CELTS’ BN can
be eliminated. Thus, CELTS must find: 1) the best Sce-
nario; 2) the cause of users’ mistakes and eliminate un-
necessary nodes between points L and T in the BN (Fig-
ure 5).
Thus, after a number of interactions with different us-
ers, we expect CELTS to propose the most self-satisfying
paths from configuration K to L and eliminate unneces-
sary nodes between points L to T. The experiment is di-
vided into two parts:
Part One
When the user (Figure 5(A)) begins a manipulation and
makes a mistake, the precondition of BN nodes activates
and waits for the relevant information to fire corre-
sponding nodes and demonstrate a message to the user.
For instance, the BN node K activates when Canadarm2
approaches configuration A in the virtual world. To help
users handle the collision risk problem with configura-
tion A, the domain expert conceived two paths in
CELTS’ BN (from points K to L in Figure 5(B)) that
correspond to this situation in the virtual world. After
interacting with users at point L, at the end of scenario1
and scenario2, CELTS asks an evaluation question to be
sure that the hints or questions given to the users were
useful and that users are aware of the collision risk in the
virtual world.
It must be noted that due to the imminent collision risk,
users’ incorrect answers to CELTS’ inquiries will acti-
vate the short route and trigger direct emotional interven-
tions as explained in CELTS’ cognitive cycles (see [2]
for more details).
As explained above, during the collision risk, CELTS
has here two choices to help users handle the situation. It
can give a direct solution to the users (scenario2, Figure
5(B)) or start by providing hints to help them handle the
situation by themselves (scenario1, Figure 5(B)).
After many executions, CELTS extracted correspond-
ing frequent event sequences for the first part of this ex-
periment (Figure 5(B) from node K to L in the BN), with
a minimum support (minsup) higher than 0.45. Using the
information extracted from this experiment, CELTS
proposed scenario1 to help users prevent collision risk in
the virtual world (Figure 5(B)), because it contains a
positive emotional valence as opposed to scenario2.
Part Two
In the second part of the experiment, after CELTS
Co pyright © 2011 SciRes. JILSA

Identifying Causes Helps a Tutoring System to Better Adapt to Learners during Training Sessions151
learned to choose the best scenario to help users prevent
collision risk with configuration A (Figure 5(A)), users
were asked to continue their manipulation and move
Canadarm2 to the configuration END. CLM learned how
to help users when they choose paths P1, P2, P3 to move
Canadarm2 from configuration A to END based on the
domain experts’ hints and questions in the BN during the
300 random executions mentioned at the onset of this
section.
Here are the details. The extracted information from
the second part of our experiment is 1) 50% of the time,
“the user is tired → the user forgot to adjust camera8”; 2)
40% of the time, “the user is tired → the user performs a
bad manipulation of Canadarm2”; 3) the remaineder10 %
rules found by CLM are that “the user is tired → user
must revise the course”, “the user is tired → user did not
make a collision risk”, and “the user is tired → user
wants to continue Canadarm2 manipulation”. Note, how-
ever, that the third rule found by CLM is not always true.
Other extracted information demonstrated that 1) 50%
of the time, “the user forgot to adjust cameras → the user
had a bad view of ISS’ configurations and Canadarm2; 2)
20% of the time, the “the user had a bad view of ISS”
configurations and Canadarm2 → the user caused a col-
lision risk near obstacles C, D, E, and F; 3) 20% of the
time “the user forgot to adjust the cameras → the user
manipulates very near to obstacles G and H”; 4) the re-
mainder 10% rules found by CLM are that “the user for-
got to adjust cameras → the user must review the lesson”,
“the user forgot to adjust cameras → the user adjusted
camera8”, and “the user was not tired → the user forgot
to answer questions”.
The extracted rules in this experiment demonstrate that
if a user forgets to adjust the cameras in the virtual world,
he or she will have a bad view of the virtual world and
this will increase collision risk.
The extracted rules could be interpreted such that the
probability of the user forgetting to adjust the cameras is
independent of the probability of a collision with ISS’
configurations, provided that the user has poor visibility
in the virtual world. The extracted rules could also be
interpreted such that the probability of having a poor
view of ISS’ configurations is independent of the prob-
ability of causing collisions in the virtual world provided
that the user has forgotten to adjust the camera.
The percentage values CELTS attributed to the various
possible causes are true most of the time, although they
must be verified by a domain expert before use. These
experiments demonstrated that CELTS is capable of
choosing the best scenario for a given situation, selecting
that which has received the highest positive emotional
valence during its interactions with the users. It is fur-
thermore capable of eliminating unnecessary nodes in the
BN.
The text has referred up to now to paths 1 through 3 in
the explanation of how to go from configuration A to
END procedures. However, there exists a path P4 which
could be considered as a shortcut.
The relevant obstacles to be avoided for this path are:
A, E, and D. Ideally, CELTS would eventually ask users
if they have some information about the obstacles they
will encounter. However, CELTS cannot ask these ques-
tions when users choose path P4 prior to starting Cana-
darm2 manipulation, since the domain expert has not
conceived relevant scenarios for this path (P4) in
CELTS’ BN. In this case, CELTS’ CLM automatically
connects to the CanadarmTutor database [27]. The data-
base contains different paths that users such as experts
and novices have previously performed to move Cana-
darm2 on the ISS. Searching all the information about
paths, CELTS’ CLM, has the capacity of giving primi-
tive hints to users when they encounter obstacles E, D
and H in path P4.
One of our future goals would be to equip CELTS
with the capacity of asking users about obstacles they
might encounter in this path, before the manipulation
starts.
6. CELTS’ Performance after the
Implementation of Causal Learning
The rule mining algorithm explained in this paper was
tested in different projects. We first explain how the
proposed algorithm improved CELTS performance. Sec-
ond we briefly explain the use of the proposed algorithm
in another e-learning project. Third, we discuss the per-
formance of the proposed algorithm with real data such
as biological database.
1) We performed a second experiment with CELTS’
causal learning mechanism, but this time to observe how
our rule algorithm behaves when the number of recorded
sequences increases. The experiment was done on a 3.6
GHz Pentium 4 computer running Windows XP, and
consisted of performing more than 250 CELTS execu-
tions for various situations (e.g., scenario 1 and scenario
2 in Figure 5(B)). In this situation, CELTS conducts a
dialogue with the user that includes from four to 20
messages or questions depending on what the user an-
swers and the choices CELTS makes. During each trial,
we randomly answered the questions asked by CELTS,
and took various measures during CELTS' learning
phase. Each recorded sequence contained approximately
30 broadcasts. Figure 6 presents the results of the ex-
periment. For all graphs, the X axis represents the execu-
tions from 1 to 250. The Y axis denotes execution times
in graph A, and rule counts in graph B-D. The first graph

Identifying Causes Helps a Tutoring System to Better Adapt to Learners during Training Sessions 152
Figure 3. Data mining algorithms’ performance.
(A) shows the time for mining rules which was generally
short (less than 10 s) and after some executions remained
low and stabilized at around 4 rules during the last exe-
cutions. In our context, this performance was very satis-
fying. However, the performance of the rule mining al-
gorithm could still be improved as we have not yet fully
optimized all of its processes and data structures. In par-
ticular, in future works we will consider modifying the
algorithm to perform incremental mining of rules. The
second graph (B) shows the number of causal rules found
after each CELTS execution. This would improve per-
formance, as it would not be necessary to recalculate
from scratch the set of patterns for each new added se-
quence. The third graph (C) shows the average number
of behaviors executed (nodes in the BN) for each CELTS
execution without causal learning. It ranges from 4 to 8
behavior broadcasts. The fourth graph (D) depicts, after
the implementation of causal learning, the number of
rules used by CELTS at each execution. Each executed
rule means that CELTS skip some unnecessary interme-
diate steps in the BN. The average number of executed
rules for each interaction ranged from 0 to 4 rules. This
means that CELTS generally used fewer nodes to per-
form the same task after the implementation of causal
learning.
2) A second use of the causal rule mining algorithm
explained in this paper is in an intelligent agent [32]. The
project is aimed at discovering patterns while observing
humans performing specific procedural tasks (see [32]
for more details). Once, the agent learned to perform the
task it can reuse rules and other kinds of patterns found
to perform the task by itself or teach the task. The
mechanism implemented in this agent is different from
the causal learning mechanism in CELTS, in that it is
designed to learn a task instead of finding causes, and it
is not based on cognitive theories.
3) At this point we discuss very briefly the contribu-
tions of our algorithm to the field of data mining. The
first issue is time complexity. Without going into techni-
cal details, the most costly part of the rule mining algo-
rithm used in this paper is the the Apriori algorithm. The
Apriori algorithm time complexity was shown by Hegland
[33] to be Ο(d2,n) where d is the number of different
items and n is the number of transactions in the database.
The elimination of association rules that do not respect
the temporal ordering to obtain causal rules is performed
in linear time with respects to the number of sequences
that contain the antecedent of each rule, the size of each
sequence, and the total number of rules [28]. However, it
must be noted that, regarding performance, we have also
recently designed an alternative algorithm that is faster
for discovering causal rules and is not based on Apriori
[32]. Integrating it into CELTS would further improve its
performance.
To test the performance of our algorithm for discover-
ing causal rules from the data mining point of view, we
have also performed several performance studies that
have been published in a conference paper on data min-
ing [28]. These studies were carried out on several large
real-life datasets such as click-streams data from web-
sites and biological sequences of proteins. In these stud-
ies, the data mining procedure has shown good perform-
ance for datasets having up to 70,000 sequences even
under very low confidence and support thresholds (for
example, the algorithm terminated in less than 250 sec-
onds for minSup = 0.05 and minSeqConf = 0.3 with
70,000 sequences on one dataset), which demonstrates its
efficiency for much larger amounts of data than what is
recorded in CELTS.
7. Comparison between Different
Architectures’ Learning Capabilities
Now we compare CELTS’ learning capabilities with
three popular architectures: LIDA [34], ACT-R [35] and
CLARION [16] (Table 1).
While LIDA is not equipped with Causal Learning,
CLARION is equipped with supervised Causal Learning.
However, at this point, there is no computational model
for causal learning proposed in CLARION. CELTS’
Causal Learning Mechanism occurs in an unsupervised
fashion and through a type of reinforcement learning, for
it partially depends on the temporal occurrence of the
Co pyright © 2011 SciRes. JILSA

Identifying Causes Helps a Tutoring System to Better Adapt to Learners during Training Sessions153
Table 1. Comparison between LIDA, ACT-R, CLARION
and CELTS ( = the architecture is not equipped with this
specific learning; X = the learning mechanism is imple-
mented).
LIDA
(Franklin,
2006)
ACT-R
(Anderson,
2004)
CLARION
(Sun, 2006)CELTS (2010)
Explicit Per-
ceptual
Learning X  X 
Episodic
Learning X X
 X
Explicit Pro-
cedural
Learning X X X X
Implicit
Procedural
Learning  X X X
Emotional
Learning
help other
types of
learning
   X
Bottom-up
Supervised
Learning X  X X
Supervised
Causal
Learning   X X
Unsupervised
Causal
Learning  X  X
events and the users’ confirmation. As is the case with
LIDA’s architecture, CELTS’ bottom-up learning is im-
plemented for all types of learning such as learning of
Emotional, Episodic, Procedural and Causal learning.
CELTS is not equipped with Attention Learning. One of
our interests is to find a way to integrate it into the archi-
tecture.
8. Conclusions
In this paper, using a combination of sequential pattern
mining and association rule algorithms, we explained
how to integrate causal learning in an Emotional Learn-
ing Tutoring System (CELTS). In CELTS, procedural
learning is dependent on the causal learning which is
further dependent upon the episodic learning. All learn-
ing is influenced by emotions.
As in the case of humans, the episodic and causal
memories in CELTS mutually influence each other dur-
ing interactions with learners. For instance, if the causes
found by CELTS turn out to be false, it influences the
support of the causal rules which in turn influences epi-
sodic memory―the increase or decrease of the events
supports.
To our knowledge, researchers in artificial intelligence
have, up to now, used Bayesian methods to study causal
reasoning and causal learning models for cognitive
agents. However, the Bayesian approach is not applica-
ble when the agent faces large amounts of data. Another
important issue with Bayesian Networks is that they gen-
erally require domain experts to specify conditional
probabilities by hand, which is often a difficult and time-
consuming task. For CELTS, we chose to use data min-
ing algorithms instead of Bayesian Networks because we
wanted to create a completely automatic approach that
could learn causal knowledge incrementally. The com-
bination of techniques used (sequential pattern mining
and association rule mining) is original for proposing a
causal learning model for cognitive agents.
However, the causal learning algorithms used in this
study are not incremental. Therefore, for each CELTS
execution, the algorithms must read the whole database.
Another limitation in our work is that given the observed
data and the confidence and support calculated by
CELTS’ CLM, the question remains as to how one could
produce the probability distribution as it exists in Bayes-
ian Networks.
9. Acknowledgements
The authors thank the Fonds Québécois de la Recher-
chesur la Nature et les Technologies for funding this re-
search.
10. References
[1] D. Dubois, P. Poirier and R. Nkambou, “What Does
Consciousness Bring to CTS?” Springer, Berlin, 2007, pp.
803-806.
[2] U. Faghihi, P. Poirier, P. Fournier-Viger and R. Nkam-
bou, “Human-Like Learning in a Conscious Agent,”
Journal of Experimental & Theoretical Artificial Intelli-
gence, in Press.
[3] U. Faghihi, P. Poirier, D. Dubois and M. Gaha, “A New
Emotional Architecture for Cognitive Tutoring Agents,”
Proceedings FLAIRS Conference, Coconut Grove, 2008,
pp. 445-446.
[4] U. Faghihi, P. Poirier, D. Dubois, M. Gaha and R.
Nkambou, “How Emotional Mechanism Learn and Helps
Other Types of Learning in a Cognitive Agent,” IEEE/
WIC/ACM International Conference on Web Intelligence
and Intelligent Agent Technology (WI-IAT 2008), IEEE
Computer Society Press, Washington, 2008.
[5] D. Purves, E. Brannon, R. Cabeza, S. A. Huettel, K. La-
Bar, M. Platt and M. Woldorff, “Principles of Cognitive
Neuroscience,” Sinauer Associates, Sunderland, 2008.
[6] C. B. Martin and M. Deutscher, “Remembering,” Phi-
losophical Review, Vol. 75, No. 2, 1966, pp. 161-196.
doi:10.2307/2183082
[7] S. Shoemaker, “Persons and Their Pasts,” American Phi-
losophical Quarterly, Vol. 7, No. 4, 1970, pp. 269-285.
[8] J. Perner, “Memory and Theory of Mind,” In: E. Tulving

Identifying Causes Helps a Tutoring System to Better Adapt to Learners during Training Sessions 154
and F. I. M. Craik, Eds., The Oxford Handbook of Mem-
ory, Oxford University Press, Oxford, 2000, pp. 297-312.
[9] S. Bernecker, “The Metaphysics of Memory,” Springer,
Berlin, 2008. doi:10.1007/978-1-4020-8220-7
[10] A. Maldonado, A. Catena, J. C. Perales and A. Cándido,
“Cognitive Biases in Human Causal Learning,” 2007.
[11] P. Maes, “How to Do the Right Thing,” Connection Sci-
ence, Vol. 1, No. 3, 1989, pp. 291-323.
doi:10.1080/09540098908915643
[12] D. A. Lagnado, M. R. Waldmann, Y. Hagmayer and S. A.
Sloman, “Beyond Covariation: Cues to Causal Structure,”
In: A. S. Gopnik and L. Schultz, Eds., Causal Learning:
Psychology, Philosophy, and Computation, Oxford Uni-
versity Press, Oxford, 2007, pp. 154-172.
[13] A. S. Gopnik and L. Schulz, “Causal Learning: Psychol-
ogy, Philosophy and Computation,” Oxford University
Press, Oxford, 2007.
[14] J. R. Anderson, “Rules of the Mind: Mahwah,” Law-
rence Erlbaum Associates, New York, 1993.
[15] W. Schoppek, “Stochastic Independence between Recog-
nition and Completion of Spatial Patterns as a Function of
Causal Interpretation,” Proceedings of the 24th Annual
Conference of the Cognitive Science Society, Lawrence
Earlbaum Associates, Mahwah, 2002, pp. 804-809.
[16] R. Sun, “The CLARION Cognitive Architecture: Extend-
ing Cognitive Modeling to Social Simulation Cognition
and Multi-Agent Interaction,” Cambridge University
Press, New York, 2006.
[17] S. Hélie, “Modélisation de L’apprentissage Ascendant
des Connaissances Explicites dans une Architecture Cog-
nitive Hybride,” PHD, DIC, UQAM, Montréal, 2007.
[18] A. Gopnik, C. Glymour, D. M. Sobel, L. E. Schulz, T.
Kushnir and D. Danks, “A Theory of Causal Learning in
Children: Causal Maps and Bayes Nets,” Psychological
Review, Vol. 111, No. 1, 2004, pp. 3-32.
doi:10.1037/0033-295X.111.1.3
[19] M. Braun, W. Rosenstiel and K.-D. Schubert, “Com-
parison of Bayesian Networks and Data Mining for Cov-
erage Directed Verification Category Simulation-Based
Verification,” High-Level Design Validation and Test
Workshop, Eighth IEEE International, 2003, pp. 91-95.
[20] S. Franklin and F. G. Patterson Jr, “The LIDA Architec-
ture: Adding New Modes of Learning to an Intelligent,
Autonomous, Software Agent,” Integrated Design and
Process Technology, 2006.
[21] D. Hofstadter, R and M. Mitchell, “The Copycat Project:
A Model of Mental Fluidity and Analogy-Making,” In: K.
J. Holyoak and J. A. Barnden, Eds., Advances in Connec-
tionist and Neural Computation Theory, Ablex, Norwood,
1994.
[22] D. Dubois, “Réalisation d’un Agent Doté D'une Con-
science Artificielle: Application à un Système Tutorel
Intelligent,” Université du Québec à Montréal, Montréal,
2007.
[23] U. Faghihi, P. Fournier-Viger, R. Nkambou, P. Poirier
and A. Mayers, “How Emotional Mechanism Helps Epi-
sodic Learning in a Cognitive Agent,” Proceedings of the
2009 IEEE Symposium on Intelligent Agents, Nashville,
2009, pp. 23-30. doi:10.1109/IA.2009.4927496
[24] R. Agrawal, T. Imielminski and A. Swami, “Mining As-
sociation Rules between Sets of Items in Large Data-
bases,” SIGMOD Conference , 1993, pp. 207-216.
[25] L. R. Squire and E. R. Kandel, “Memory: From Mind to
Molecules,” W. H. Freeman, San Francisco, 2000.
[26] V. Goel and R. J. Dolan, “Differential Involvement of
Left Prefrontal Cortex in Inductive and Deductive Rea-
soning,” Cognition, Vol. 93, No. 3, 2004, pp. 109-121.
doi:10.1016/j.cognition.2004.03.001
[27] R. Nkambou, K. Belghith and F. Kabanza, “An Ap-
proach to Intelligent Training on a Robotic Simulator
Using an Innovative Path-Planner,” Proceeding of the 8th
International Conference on Intelligent Tutoring Systems
(ITS), LNCS, 2006, pp. 645-654.
[28] P. Fournier-Viger, U. Faghihi, R. Nkambou and E. M.
Nguifo, “CMRULES: An Efficient Algorithm for Mining
Sequential Rules Common to Several Sequences,” FLAI-
RS Conference, 2010.
[29] J. Hipp, U. Güntzer and G. Nakhaeizadeh, “Data Mining
of Association Rules and the Process of Knowledge Dis-
covery in Databases,” Industrial Conference on Data
Mining, 2002, pp. 15-36.
[30] J. Deogun and L. Jiang, “Prediction Mining—An Ap-
proach to Mining Association Rules for Prediction,” Pro-
ceedings of RSFDGRC 2005 Conference, Springer-Ver-
lag, Berlin, 2005, pp. 98-108.
[31] L. Li and J. S. Deogun, “Discovering Partial Periodic
Sequential Association Rules with Time Lag in Multiple
Sequences for Prediction,” Proceedings of ISMIS 2005,
Springer, Berlin, LNCS 3488, 2005, pp. 332-341.
[32] P. Fournier-Viger, “Knowledge Discovery in Problem-
Solving Learning Activities,” Ph.D. Thesis, University of
Quebec in Montreal, 2010.
[33] M. Hegland, “The Apriori Algorithm—A Tutorial.
Mathematics and Computation,” Imaging Science and
Information Processing, Vol. 11, 2007, pp. 209-262.
doi:10.1142/9789812709066_0006
[34] S. Franklin, “A Cognitive Theory of Everything: The
LIDA Technology as an Artificial General Intelligence,”
Artificial General Intelligence Research Institute (AGIRI)
2006.
[35] J. R. Anderson, D. Bothell, M. D. Byrne, S. Douglass, C.
lebiere and Y. Qin, “An Integrated Theory of the Mind,”
Psychological Review, Vol. 111, No. 4, 2004, pp. 1036-
1060.
doi:10.1037/0033-295X.111.4.1036