=Paper=
{{Paper
|id=Vol-2333/paper7
|storemode=property
|title=ExPLoRAA: An Intelligent Tutoring System for Active Ageing in (Flexible) Time and Space
|pdfUrl=https://ceur-ws.org/Vol-2333/paper7.pdf
|volume=Vol-2333
|authors=Amedeo Cesta,Gabriella Cortellessa,Riccardo De Benedictis,Francesca Fracasso
|dblpUrl=https://dblp.org/rec/conf/aiia/CestaCBF18
}}
==ExPLoRAA: An Intelligent Tutoring System for Active Ageing in (Flexible) Time and Space==
https://ceur-ws.org/Vol-2333/paper7.pdf
ExPLoRAA: An Intelligent Tutoring System for
Active Ageing in (Flexible) Time and Space?
Amedeo Cesta, Gabriella Cortellessa, Riccardo De Benedictis, and Francesca
Fracasso
CNR - Italian National Research Council, ISTC, Rome, Italy,
{name.surname}@istc.cnr.it
http://www.istc.cnr.it/it/group/pst
Abstract. The “Città Educante” project aims at radically rethinking
the learning experience through advanced ICT technology to enrich and
innovate didactic methods and tools. Among the project results, ExPLo-
RAA is an Intelligent Tutoring System, specifically tailored for senior
citizens, which, by integrating artificial intelligence and state-of-the-art
ICT techniques, is able to support older adults during visits to cultural
locations in a city. In particular, ExPLoRAA integrates both the users’
psycho-physiological aspects as well as geo-localization information and
temporal constraints in the attempt to personalize the learning stimuli
during the visit while favouring the concept of active ageing for the older
people. After a generic introduction to the “Città Educante” project, this
paper presents both ExPLoRAA as a whole and some of the underlying
solutions. Specifically, the paper shows some of the choices that have
been made to solve problems related to temporal flexibility, supporting
the dynamic adaptation of stimuli over time while ensuring the possibility
for the users to further adapt a visit according to their current feelings.
The paper describes both the choices made in the current system proto-
type and its embodiment in a concrete scenario which, by implementing
a game similar to “treasure hunt”, aims at fostering the physical and
cognitive activity of the participating older people.
Keywords: Intelligent Tutoring System · Active Ageing · Temporal
Flexibility.
1 Introduction
The “Città Educante” project1 (the name means “city that educates” in Ital-
ian) aims at proposing new educational approaches, enriching and innovating
methods and tools, overcoming the classical systems and the traditional role of
educators. An Intelligent Tutoring System (ITS) [15, 18] is a computer system
that, thanks to personalized stimuli to the learners, enables learning in an ef-
fective and meaningful manner. Such systems aim at radically rethinking the
?
Authors work is partially funded by MIUR under Cluster Program 2012. Send cor-
respondence to amedeo.cesta@istc.cnr.it.
1
http://www.cittaeducante.it
learning environment through the application of the most advanced ICT tech-
nology and represent, hence, a valid contribution to the project.
Common ITSs, typically, aim at replicating the benefits of one-to-one per-
sonalized tutoring in contexts where students would have access to one-to-many
instruction from a single teacher (e.g., classroom lectures), or no teacher at
all (e.g., on-line homework) [17]. Such systems are typically related to classical
learning environment, hence neglecting the possibility to act in time (e.g., life-
long) and space (e.g., at school, in an outdoor environment, during leisure, etc.),
thus overcoming the classical systems and the traditional “lessons”.
Through the “Città Educante” project, the theme of learning is framed in re-
lation to the response to social challenges linked to the renewal of the educational
system, to be achieved by means of the implementation of new learning/teaching
models and/or the optimization of the existing ones on the various areas of life
and knowledge, as well as new systems/evaluation processes, in which the tech-
nology (platforms and web) becomes an enabling factor.
Specifically, the authors’ goal in the project has been the one of thinking
an incarnation of the “Città Educante” for the continuous education of older
people. In particular, considering the more recent experiences in the interaction
of elderlies with complex machines [6], keeping in mind previous experience in
training elders [4] and crisis managers [2], we have built a new learning environ-
ment, called ExPLoRAA (for ExPeriential LeaRning for Active Aging), aimed
at improving the active aging and the participation in the social life of the elders
living at home, in the community, and at work.
2 Televita: the Target Organization
While pursuing the objective of the project, which aims at reformulating the
learning environments through the creation of platforms, services and ICT ap-
plications, we got in touch with several volunteering organizations addressing,
specifically, elderlies’ needs. Among the different organizations, in particular,
Televita2 is a volunteering association whose main objective is to maintain the
elderlies active and motivated, leveraging upon individual aptitudes and/or com-
petencies [5]. Most of the Televita’s volunteers, in fact, are, themselves, elders
who want to keep active by offering their abilities and competences to the orga-
nization. Although Televita’s main activity consists in providing tele-assistance
services (tele-friendship) and a 24h active helpline devoted to lonely elders who
need support, it also manages several laboratories that involve elders both as
attendees and “teachers”. Examples include a computer lab, a tailoring lab, a
cooking lab and an Italian language teaching for foreign people. Furthermore,
the association organizes cultural events as concerts, museum visiting, theater,
etc.
Among the offered services, we focused in particular on two specific activities
that were both in line with the “Città Educante” concept and that are outlined
in Figure 1:
2
http://www.televita.org
Fig. 1. Examples of Televita activities.
– The AttivaMente laboratory: aims at keeping elders mentally active, so as
to limit the cognitive decline associated with the advancement of age, by
proposing them cognitive stimuli. Such stimuli, mostly consisting in general
culture quiz and/or crosswords, are proposed to the elders in a context sim-
ilar to a school lesson. Specifically, by relying on some previous knowledge
of the involved persons, as well as on their interactions during the lesson, a
teacher, a volunteer himself, yet with more experience than others, controls
the course’s progress and slightly adapt it to the specific context’s needs.
Since the stimuli are predetermined, however, such adaptations tend to be
limited to the possibilities of the case. The use of artificial intelligence tech-
niques, in this case, could support the personalized delivery of stimuli by
taking into account previous knowledge of the participants as well as their
interactions with the AI system increasing, compared to the classical case,
the personalization capabilities.
– The Art History lessons and cultural events: analogously to the AttivaMente
case, before attending to cultural events, some of the participants, accord-
ing to their abilities and competencies, are asked to find information about
some specific aspects of the event (e.g., a particular work of art within a
museum, relevant historical happenings related to a visited site, etc.). Such
information are then shared, in a “sort of” lesson with other participants in
a classroom context and also outside during the event, enriching the overall
knowledge of the group while encouraging the interaction among the mem-
bers. Similarly to the AttivaMente case, such information may result to be
limited and/or not customized to the specific members of the event. AI tech-
niques, in this case, might offer the opportunity to enrich the cultural events
experience by providing further personalized stimuli to the event partici-
pants, as well as interaction requests to test the level of engagement and
actively stimulate them.
Combining the above activities represents an opportunity for the elders to keep
themselves active while learning in time and space. Additionally, the use of AI
can support the development of more effective and engaging learning experience.
3 The ExPLoRAA Intelligent Tutoring System
The idea of developing the ExPLoRAA system is born as a consequence of a
field experience. Specifically, by taking inspiration from a previous work [3], in
which students were trained for managing crisis, the approach used within Ex-
PLoRAA is based on the idea of dynamically composing lessons through the use
of automated planning [11]. In particular, starting from a static representation
containing an high-level lesson track, initially stored in a database, the lesson is
planned and dynamically adapted and personalized to the involved users. The
idea of using the technology related to automated planning comes from the need
to create a sufficiently extensive didactic experience to reproduce a large num-
ber of different situations which are, at the same time, characterized by a high
variability of stimuli, aimed at increasing the involvement level of users. Auto-
mated planning, indeed, favors the generation of different lessons that would be
too complicated to obtain with a simple pre-compilation of stories. The timeline-
based approach to automated planning [14], in particular, represents the unifying
element of the various modules by ensuring the dynamic adaptability of plans
by promoting experiential learning.
Fig. 2. The ExPLoRAA general idea.
From a high-level point of view, the main modules of the system are de-
scribed in Figure 2. In particular, it is possible to distinguish between two kinds
of involved users: the students, i.e., a group of people, potentially, of any age, in-
terested in using the learning services offered by the ExPLoRAA environment,
and the teachers, i.e., users with special privileges who have the opportunity to
observe students, monitor the progress of the lessons and of the overall learning
environment. The above users interact with the ExPLoRAA system which is
composed of three functional blocks, intended as architectural subsystems, imple-
menting the corresponding high-level functionalities: (i) the user modeling, whose
goal is to create and maintain a user model and provide guidance for improving
the learning process; (ii) the lesson modeling, whose role consists in combining
the information from the previous subsystem and to create the customized lesson
as well as to control its evolution; (iii) the lesson presentation, whose purpose
is to represent the lesson through effective graphical interfaces. Additionally, it
is worth highlighting that the proposed system provides users, whether students
or teachers, the opportunity to adapt the learning environment’s evolution in
real time by interpreting their decisions. In fact, the architecture is based on a
sense-plan-act paradigm implementing, in a continuous loop, the three primi-
tives (a) sense, in which information is collected from questions and sensors, (b)
plan where a lesson blueprint is created, using the information available, and (c)
act, in which the action, chosen by the planning process, is actually executed.
It is worth highlighting that, by exploiting mobile technology, the lesson pre-
sentation module interacts remotely with the system allowing different learning
modalities: (a) on-site learning, closer to the classical teaching, in which the
technology is used as a support to the teaching in a classroom, with the aim
to create richer lessons, and (b) distributed learning, in which the technology
aims to support lessons outside the classroom during a practical experience.
More specifically, in the on-site training modality, the system can be used by
a group of students at the same time. This mode represents an extension to
the classical learning method in which a teacher teaches to a group of students.
In this case, however, compared to the classical approach, the teaching is en-
hanced by the introduction, within the lesson, of the ExPLoRAA technology.
Each lesson is instantiated by the teacher by defining specific learning objectives
[12] that we call goals. The system processes the lesson and presents, at proper
time, the information to the students through the presentation tools. Students
interact directly with the system, providing their answers to certain circum-
stances proposed by the system, and transmitting data from sensors available
on the adopted devices (e.g., physiological parameters) enriching the users’ mod-
els. Conversely, in the distributed training case, the lesson does not happen in
a single physical room and is distributed among the students who are remotely
connected to the system. The lesson is still instantiated by the teacher defining
the specific learning objectives but may have variable, potentially infinite, dura-
tion. This kind of approach, compared to the previous one, is more innovative.
Students interact directly with the system while on the move within the city,
providing their answers to certain proactive stimuli proposed by the system as
well as constantly transmitting data from the sensors available on the chosen
devices (e.g., geographic location, physiological parameters, etc.). Sensor data,
in particular, enriches the users’ models which, in turn, adapt the lesson to the
students resulting in a highly personalized learning experience.
4 Timeline-based planning to support dynamic lessons
Since most of the components of the ExPLoRAA system strongly depend on
temporal aspects, we have chosen to rely on a specific automated planning tech-
nique, called timeline-based, which allows to explicitly reason on time. Timeline-
based planning, indeed, allows to reason about events in time and, hence, rep-
resents a valid tool for meeting our pedagogical needs. Planning a lesson, in
particular, requires dispatching information at proper time. Additionally, react-
ing to users’ interactions requires plan adaptation capabilities which can more
hardly be achieved through other automatic planning techniques. Furthermore,
the dynamic adaptation of the user profiles, which can take place on the different
features that represent the user’s model, can also be achieved through timelines.
As already said, timeline-based planning constitutes a technology that easily
allows us to solve our problems. In order to better contextualize the choices that
we have made and to better explain the different components of the system,
however, it is worth introducing some basic formalism about constraint networks,
on which timeline-based planning search strongly relies, and about the main
concepts related to timeline-based planning. Specifically, the main ingredients of
constraint networks are variables and constraints.
Definition 1. A variable is an object that has a name and is able to take dif-
ferent values.
A variable (whose name is) x must be given a value from a set that is called
the domain of x and is denoted by dom (x). The domain of a variable x may
evolve in time but is always included in a set called initial domain. Depending on
the nature of these domains, variables can be distinguished between continuous,
having an infinite initial domain usually defined in terms of real intervals, and
discrete, whose initial domain contains a finite number of values.
Definition 2. A constraint is a restriction on combinations of values that can
be taken simultaneously by a set of variables.
A constraint c is defined over a set of variables which constitute the scope
of c and are denoted by scp (c). Finally, a structure composed of variables and
constraints is called a constraint network.
Definition 3. A constraint network N is composed of a finite set of variables,
denoted by vars (N ), and a finite set of constraints, denoted by cons (N ), such
that ∀c ∈ cons (N ) , scp (c) ⊆ vars (N ).
Since constraint networks are fundamentals for timeline-based planning, it
is worth introducing some further concepts without going into too much formal
details. Specifically, an assignment of values to some or all the variables is called
an evaluation. Furthermore, an evaluation is said to be consistent if it does not
violate any constraint. An evaluation is said to be complete if it includes all the
variables. Finally, given a constraint network, the problem of finding a consistent
and complete evaluation is called Constraint Satisfaction Problem (CSP) (refer
to [9, 13] for a comprehensive introduction to CSPs).
As regards timeline-based planning, the main data structure is the timeline
which, in generic terms, is a function of time over a finite domain. Values on the
timelines are extracted from a set of temporally scoped predicates (i.e., predicates
endowed with extra arguments belonging to the Time domain T, either real or
discrete), with their parameters, called tokens. Formally,
Definition 4. A token is an expression of the form:
n (x0 , . . . , xk ) @ [s, e, τ ]
where n is a predicate name, x0 , . . . , xk are the predicate’s parameters (i.e., con-
stants, numeric variables or object variables), s and e are temporal parameters
belonging to T such that s ≤ e and τ is a parameter (i.e., a constant or an object
variable) representing the timeline on which the token apply.
The overall idea pursued in ExPLoRAA consists in using such tokens for
representing the planned stimuli. Compared to the general formalization above,
however, we can afford some simplifications. Specifically, since the stimuli have no
duration, the s and e variables of each token would always be equal. We address
this by removing one of the two variables, e.g., the e variable. Additionally, since
all the tokens will apply on a single “lesson” timeline, the τ variable would be a
constant. For example, an expression st0 () @[10 : 00] would represent a stimulus
st0 which is planned to happen at time 10:00.
It is worth noticing that the tokens’ parameters, including the temporal ones,
are constituted, in general, by the variables of the constraint network as in-
troduced in the Definition 1. In order to reduce the allowed values for such
parameters, and thus decreasing the system’s allowed behaviors, it is possible,
indeed, to impose constraints, as introduced in the Definition 2, among them
(and/or between the parameters and other possible variables). Such constraints
include temporal constraints, usually expressed by means of interval relations [1],
binding constraints between object variables as well as linear constraints among
numerical, including the temporal one, variables.
The set of tokens and constraints is used to describe the main data structure
that will be used to represent the nodes of the timeline-based search space: the
token network.
Definition 5. A token network is a tuple π = (T , C), where:
– T = {t0 , . . . , tk } is a set of tokens.
– C is a set of constraints, required to be consistent, on the variables of the
tokens in T .
Additionally, tokens can be partitioned into two groups: facts and goals.
While facts are, by definition, inherently true, goals have to be achieved. Specif-
ically, causality, in the timeline-based approach, is defined by means of a set o
rules indicating how to achieve goals. Formally,
Definition 6. A rule is an expression of the form
n (x0 , . . . , xk ) @ [s, e, τ ] ← r
where:
– n (x0 , . . . , xk ) @ [s, e, τ ] is the head of the rule, i.e., an expression in which
n is a predicate name, x0 , . . . , xk are constants, numeric variables or object
variables, s and e are temporal variables belonging to T such that s ≤ e and
τ is an object variable representing the timeline on which the token apply.
– r is the body of the rule (or the requirement), i.e., either a slave token, a
constraint among tokens (possibly including the x0 . . . xk variables), a con-
junction of requirements or a disjunction of requirements.
Rules define causal relations that must be complied to in order for a given goal to
be achieved. For each goal having the form of the head of a rule, the body of the
rule must also be present in the token network. As an example, the expression
{st0 () @[s] ← {st1 () @[s1 ] ∧ 10 ≤ s − s1 ≤ 20}} represents a rule asserting that,
for each stimulus st0 there must exist, from 10 to 20 seconds before, a stimulus
st1 .
We have now all the ingredients to define a timeline-based planning problem.
In particular, the definition can rely on the above concept of requirement.
Definition 7. A timeline-based planning problem is a triple P = (T , R, r),
where:
– T is a set of timelines.
– R is a set of rules.
– r is a requirement, i.e., either a (fact or goal) token, a constraint among
tokens, a conjunction of requirements or a disjunction of requirements.
The role of a timeline-based solver consists, basically, in applying the proper
rules to achieve all the goals of the problem while maintaining the constraint
network consistent. It is worth highlighting that, in general, the application
of the rules might result in the introduction of further goals into the token
network. Such goals, also called sub-goals, require to be achieved as well. The
process ends up when, for all the goals of the token network, either the body
of its corresponding rule is present in the token network or it is recognized as
semantically equivalent to another token (in this case we talk about unification
of the tokens). Notice that, despite the simplicity of the above solving procedure,
the combination of disjunctions and constraints in the rules make the resolution
process, in generally, extremely challenging from a computational point of view.
For this reason, indeed, heuristics are often used to make the resolution process
more efficient (see, for example, [7, 8]).
5 Modeling Students and Lessons through Timelines
As already mentioned, the ExPLoRAA system is composed of different func-
tional blocks. In particular, the user modeling module aims at creating and dy-
namically maintaining an updated model of the users which is used as a starting
point for the personalization of the lessons. By pursuing the overall objective of
enhancing the learning experience, indeed, it is necessary to keep a user model
up-to-date in order to consider how their emotional, psychological, physiological
and geographical parameters can influence the learning process. Specifically, the
student modeling has three main objectives:
– Model and monitor relevant factors through which the lesson can be cus-
tomized;
– Develop a model that can represent the user’s profile;
– Provide a high level guidance for customizing learning objectives.
The set of considered relevant factors include, among the other things, the
health status, the fatigue, the personal interests, the level of engagement and
the current performance assessment. The use of Bluetooth bracelets (e.g., the
Empatica E43 ), for example, allows the extraction of physiological values such as
peripheral skin temperature, skin conductance, heart rate and heart variability.
Additionally, it is possible to leverage on geo-localization services to get a good
estimate of the users’ position in time. The initial evaluation of these variables,
used as a baseline to initialize the didactic experience and as a reference point
for subsequent measurements, can be done through the use of standardized ques-
tionnaires or physiological measurements performed off-line before the lesson. It
is worth to notice, however, that the profile of a student can also be updated
exploiting the interactions of the users with the system asking them, for exam-
ple, to answer to sporadic questions. As an example, the users’ engagement is
measured through a five levels Likert-type scale which is administered to the
users at regular intervals. Finally, particular emphasis is given to the students’
performance which is monitored and observed through the administration of
questions and the interpretation of the provided answers.
By processing the above information, the system generates a user model that
is constantly updated to perceive and represent significant changes in the emo-
tional state (note that parameters can generally change over time). In addition,
students’ performance is analyzed and processed in order to gather further us-
able information useful for a finer customization of the lesson. Furthermore, the
teacher can access this information in order to supervise and control the cus-
tomization. For this purpose, this component can provide guidance on how to
customize the lesson. Personalization of a training course can therefore be done
automatically, but it can also be suggested to a teacher who independently de-
cides whether to adapt the training course (i.e., according to a mixed-initiative
style).
5.1 Modeling the Lessons
The modeling of the lessons is the key feature of the ExPLoRAA system since
it creates and manages the network of stimuli that guides the entire learning
session. Nodes on this network are tokens and are intended to represent tempo-
rally annotated stimuli (e.g., videos, text messages, questions, etc.) to be sent,
3
https://www.empatica.com/research/e4
at appropriate time, to the users while edges represent causal and temporal rela-
tions among such stimuli introduced either in the planning problem definition or
through the application of the rules. Additionally, tokens are endowed with ad-
ditional information including a set of covered topics (e.g., “art”, “architecture”,
“religion”, etc.) and some content dependent on the nature of the stimulus (e.g.,
a text for textual stimulus or a URL for a video stimulus). Specifically, when a
user enrolls to a lesson, defines which of the topics covered by the lesson are of
interest to her/him. When sending stimuli to the users, a filtering procedure will
guarantee the delivery of only those stimuli for which the user has declared an
interest.
It is worth noting that although the above network is initialized in order
to represent an abstract blueprint of a lesson, it is afterwards customized and
dynamically adapted to the profile of the involved user. Personalization, indeed,
takes place both in terms of users’ interest in some topics, as explained earlier,
as well as in terms of the type (and the number) of tokens in the token network.
Specifically, adaptations to the network are made thanks to the application of a
set of rules (introduced in Definition 6) associated to each lesson, which define
how to “react” to the users’ profile, to their updates and to their actions (e.g.,
moving to a specific location or answering to a question). Such rules, in particu-
lar, are intended to create the “conditions”, in terms of events and their relations
within the network, for other events to be present. An example of rule can be “in
order to stimulate the cognitive activity of the group, either propose a simple
crosswords and the group’s performance is low, or propose a complex crosswords
and the group’s performance is high”. Notice that by taking advantage of the
possibility of defining disjunctions within the rules and being able to combine
such rules sequentially, it is possible to obtain a great wealth of possible lessons’
evolutions. Finally, since some of these rules may contain conditions which con-
cern the user model, not all of them are applicable (e.g., in the above example,
in case the group’s current performance is low, only the simple crosswords is
proposed), resulting in an overall network which is always compatible with the
current users’ profiles.
Broadly speaking, the teacher loads the chosen lesson from a database re-
sulting in the construction of an initial token network corresponding to a set of
stimuli, positioned over time, which will be communicated to users like videos,
text messages, questions, etc. The network is, through the application of the
rules, afterwards customized to the users participating in the lessons. By execut-
ing the lesson, then, tokens, representing stimuli and interaction requests, are
dispatched, at proper time, to the users interested to them. It is worth noticing
that in order to foster interaction and collaboration among them, the distributed
information may be partial, requiring users the need to send messages to other
students so that they can build an overview and respond appropriately to the
challenges posed by the system, thus fostering cooperation. Whenever the pro-
file of a user changes because, for example, her/his level of fatigue increases,
specific rules are applied resulting in an update of the network so as to bring it
back to be “consistent” with the current status of the users. Similarly, updates
might happen as a consequence of users’ actions, resulting in a network which
is always representative of the initial lesson while being dynamically adapted to
the specific context.
Fig. 3. An example of ExPLoRAA dynamic lesson adaptation in case either action
a0 or action a1 is performed by a user.
As an example, Figure 3 shows, at its top, an initial network containing three
stimuli (i.e., st0 , st1 and st3 ) and an interaction request (i.e., ir2 ) representing,
for example, a question. Each event has its own execution time and its covered
topics (e.g., st0 will be dispatched at 08:20). In the figure, the arrows represent
the causal relations among the tokens that emerge from the application of the
rules. In other words, the st1 token is in the network “because of” the st0 goal.
Additionally, the st1 token is constrained to be from 15 to 40 seconds before the
st0 goal. Notice that in order to simplify the explanation and make the speech
clearer we have omitted the information regarding the other temporal constraints
which, however, can be considered embedded within the arrows.
At 08:05, the interaction request ir2 requires the interested users (i.e., those
whose interests match the topics of the stimulus associated to the ir2 token) to
take some action (i.e., either a0 or a1 ). Suppose, as an example, that the action
a0 is chosen, the network is adapted to the one on the left by adding the new
token st4 at 08:30. Again, the st4 token is added within the network “because
of” the a0 action. Conversely, in case action a1 is chosen, the network is adapted
to the one on the right by adding the new tokens st5 , at time 08:35, and st6 at
08:38. As already mentioned, it is worth noticing that this type of adaptation
occurs in a similar way as a consequence of the actions carried out by the users
as well as for the dynamic changes that occur to their profile, ensuring a discrete
availability of flexibility and adaptation to the particular conditions that may
arise in the different lessons.
6 Supporting Lesson Adaptation
An interesting feature of the proposed system regards the possibility for adapt-
ing the lessons thanks to the supported temporal flexibility. In particular, the
constraint network introduced in Definition 3 manages temporal variables and
constraints through a Simple Temporal Network (STN) (see [10]). Specifically,
each constraint between two temporal variables xi and xj is limited to taking the
form l ≤ xj − xi ≤ u where l and u are two constants representing, respectively,
the lower and the upper bound on the amount of time that lasts between the
two time points xi and xj . The application of the rules during the resolution
procedure requires the dynamic introduction of further variables, for each of the
slave tokens, and additional constraints into the constraint network. For this
reason we use incremental algorithms (see, for example, [16]) which are able to
efficiently manage the propagation of such constraints. These algorithms are able
to maintain the temporal flexibility of the involved variables, producing, for each
variable x, a couple of bounds [lx , ux ] representing its lower and upper allowed
values. Dechter et al., in [10], demonstrate that assigning to each variable x its
earliest possible time lx results in a consistent temporal network4 .
While in a standard lesson execution we would use these values to establish
the timing for dispatching the stimuli, problems arise when the teachers want
to control their position in time. Suppose, for example, we have an STN with
two variables x0 and x1 , representing the time points of two tokens associated
to two stimuli st0 and st1 . Additionally, the 10 ≤ x1 − x0 ≤ 20 constraint
imposes a temporal distance between them which goes from 10 to 20 time units
(although the system uses milliseconds, we maintain in the dissertation a generic
measurement unit for sake of simplicity). The initial bounds of the x0 variable
is [10, 30] while the initial bounds of the x1 variable is [20, 50]. Since the earliest
possible times represent a solution for such a constraint network, st0 and st1 are
going to be dispatched, respectively, at time 10 and at time 20. Suppose, for any
didactic reason, the teacher wants to delay the execution time of the stimulus
st0 of 15 time units, a temporal constraint would reduce the bounds of the x0
variable to [25, 30]. The propagation of the constraints would tighten the bound
of the x1 variable as well to, possibly (notice that other constraints might also
be involved), [35, 50], hence right-shifting the dispatching time of st1 at 35. At a
second step, however, the teacher recognizes she/he has moved the stimulus st1
too far and decides it is better to move it a little back. The lower bound for the
st1 stimulus, however, is now 35 and cannot be moved back through temporal
constraints without introducing an inconsistency in the constraint network.
We can get around this problem by introducing, for each variable x, together
with its lower (i.e., lx ) and upper (i.e., ux ) bounds resulting from the propagation
of the temporal constraints, a third value dx representing its dispatching time,
whose allowed values are always maintained between the bounds of the variable
(i.e., the relation lx ≤ dx ≤ ux always holds). Whenever the user decides to
4
It is worth noticing that the same result holds also when choosing for each variable
their latest possible times.
change the position in time of a stimulus, however, only the dispatching times
are updated, without touching the bounds of any variable.
In order to implement such a procedure we need, first of all, for each tem-
poral variable x, a list of watching constraints watchs [x] (i.e., for each variable,
the list of constraint watching for the variable’s updates). Specifically, when a
new constraint c is introduced into the constraint network, c is added to the
watching lists watchs [xi ] of each variable xi ∈ scp (c) belonging to the scope
of c. Additionally, we need a propagation queue propq responsible for maintain-
ing the list of variables whose dispatching time has been updated. Whenever
the dispatching time of a variable is updated, the variable is enqueued into the
propagation queue.
Algorithm 1 The overall propagation procedure
procedure propagate
while propq 6= ∅ do
x ← propq .dequeue
for all c ∈ watches [x] do c.propagate(x)
The propagation procedure is described in the Algorithm 1. Specifically, while
the propagation queue is not empty, a variable x is dequeued from the queue
and a propagation procedure is called for each constraint c in the watching list
associated to the variable x (i.e., for all the constraints having the variable x
in their scope). Additionally, for each temporal constraint c having the form
l ≤ xj − xi ≤ u the propagation procedure is described in the Algorithm 2.
Specifically, whenever the dispatching time of the xi (xj ) variable is updated,
the propagation procedure checks whether the xj (xi ) variable’s dispatching time
needs to be updated as well. In case it needs, the xj (xi ) variable’s dispatching
time is updated and the xj (xi ) variable is enqueued in the propq propagation
queue.
The last aspect to be considered regards the initial introduction of the vari-
ables within the propagation queue. There are basically two reasons which might
lead to this: either (a) a bound of a variable x is updated, within the STN prop-
agation procedure, in a way such that the expression lx ≤ dx ≤ ux is violated or
(b) a user manually updates the dispatching time of a variable x associated to a
stimulus. In the first case, described in the Algorithm 3, the dispatching time is
updated so as to restore the validity of the lx ≤ dx ≤ ux relation and the vari-
able is enqueued in propq . The second case, described in the Algorithm 4, more
straightforwardly, directly updates the dispatching time and enqueues the vari-
ble in propq . Notice that, after calling both the STN propagation procedure and
the set dspatching time procedure, the propagate procedure (Algorithm 1)
must be called in order to update the dispatching times of the related variables.
Finally, it is worth noticing that none of the above procedures is supposed to
fail. The consistency checking of the constraint network, indeed, is still delegated
to the STN propagation procedure.
Algorithm 2 The propagation procedure for the temporal constraint
procedure propagate(x)
if x == xi then
if dxj < dxi + l then
dxj ← dxi + l
propq ← propq ∪ {xj }
else if dxj > dxi + u then
dxj ← dxi + u
propq ← propq ∪ {xj }
else if x == xj then
if dxi < dxj − u then
dxi ← dxj − u
propq ← propq ∪ {xi }
else if dxi > dxj − l then
dxi ← dxj − l
propq ← propq ∪ {xi }
Algorithm 3 The bound update procedure called by the STN propagation
procedure
procedure set bound(x, l, u)
lx ← l, ux ← u
if dx < lx then
d x ← lx
propq ← propq ∪ {x}
else if dx > ux then
dx ← dx
propq ← propq ∪ {x}
Algorithm 4 The dispatching time update procedure
procedure set dispatching time(x, d)
if dx 6= d then
dx ← d
propq ← propq ∪ {x}
7 The ExPLoRAA Prototype and Preliminary Evaluation
By relying on the concept sketched in Figure 2, we have realized a first proto-
type of the ExPLoRAA ITS. In particular, several instances of the Desktop
and Mobile applications, depicted in Figure 4, allow students and teachers the
access to the ExPLoRAA system through the remote interaction with a central-
ized back-end. Specifically, teachers have the opportunity to create and manage
lessons while monitoring the students following them. Students, on the other
hand, enroll to the available lessons specifying their interests and receive cus-
tomized stimuli according to their psychophysiological state assessed through a
combination of sensors and targeted questions.
Fig. 4. The ExPLoRAA Desktop and Mobile Graphical User Interfaces.
We have created a lesson model example inspired by the “Madonnelle Stradaiole”
of Rome. These are present in large numbers on the walls or corners of historic
buildings. The idea consists in sending stimuli to users so as to guide them to
visit such shrines, customizing the path to their psychophysiological state. By
exploiting georeferencing, the ExPLoRAA system asks users to take pictures
of the shrines to be used, afterwards, to build games which, played together,
reduce social isolation and stimulate cognitive activity. The first step for the im-
plementation of a scenario of this kind consists in collecting information about
the shrines like, for example, a brief history about each of them and their GPS
coordinates. Additionally, a customized questionnaire, administered to users, is
intended to extract an initial profile to be used as a baseline. Starting by the
initial profile, a path, compatible with the profile, is selected and, step by step
and at proper time, suggested to the user (e.g., “go to pos1”, “the shrine at
pos1 has been built in 1796”, “take a picture at the shrine in pos1”, “go to
pos4”, etc.). By taking into account physiological data, the system can adapt
the route switching, for example, once the ExPLoRAA system realizes that the
user is not too tired, to a longer one, fostering physical activity and a prolonged
interaction with the other involved users. Finally, once back from the trip, the
ExPLoRAA system builds a “memory” game, challenging the participants to
discover, in the fewest possible steps, pairs of the taken pictures hidden under
the tiles, hence stimulating cognitive activity.
Feedback gathering has taken place at the Televita facilities. Specifically, a
focus group has been held in which volunteers of the association, as well as rep-
resentatives of other associations that deal with assistance to the elderly and/or
psychological distress present on the territory of the III Municipality of Rome,
have participated. Specifically, 6 people dealing with Televita, 3 representatives
of “Società San Vincenzo de Paoli”, and 1 representative of the volunteering or-
ganization “Oltre Le Barriere”. The meeting (see Figure 5) has been structured
around the presentation of the developed prototype that proved to be functional
for eliciting comments and suggestions from the participants. We wanted to ad-
dress the discussion in the direction of feedback collection for what concerns the
improvement of the interface, the modalities of providing the service, the ways to
enhance the engagement, as well as the possibility for the participants to provide
suggestions and free comments.
Fig. 5. Focus group participants.
In general, the feedback obtained on the platform and its services was decid-
edly positive and the participants found the play aspect of the project particu-
larly interesting. In fact, the game was read not only as a recreative tool, but
also as a learning and social cohesion one, a crucial aspect to promote well-being
in the elderly. Furthermore, the value of the possibility offered to citizens who
can participate in the co-management of the public good through the municipal
volunteering has emerged and, thanks to ExPLoRAA, they have the opportu-
nity to take on some aspects in a territorial area thus favoring a certain sense of
belonging. This reflects the need to motivate older people to feel active, rather
than a burden for society or for the family. Participation in activities motivates
people and makes them feel alive. Determinant is also trying to cope with one
of the biggest problems of older people: the suffering of loneliness. It emerged,
indeed, that the proposed formula could represent an effective tool, especially
considering the social character of the activities that are essentially based on the
aggregation and comparison between different people in a game context. The
possibility of personalizing the experience was judged to be very useful, both to
encourage the motivation and involvement of people during the activity as well
as to provide access to a wider range of users, taking into account possible phys-
ical limitations. In this sense, the personalization of services is not defined only
by psicological aspects, such as the level of engagement, attention, emotional
state, but also by more purely physical factors, such as fatigue. According to
the participants, these characteristics make the proposal even more effective and
desirable. Finally, a very interesting aspect that emerged from the interaction
with the volunteers concerns the possibility of using the ExPLoRAA system
to favor inter-generational exchange. In fact, much emphasis has been given to
the possibility of exploiting the system to make elderly and young people inter-
act and to promote the comparison. In fact, a system focused on aspects such
as sharing and group activities favors the exchange, while taking into account
individual preferences and needs, thanks to the possibility of personalizing the
experience.
8 Conclusions and Future Works
By pursuing their didactic needs, teachers want flexibility in adapting the re-
sulting network. Furthermore, teachers are human beings and, as such, might
change their idea. This paper introduces the ExPLoRAA system as an AI-
based learning environment specialized so as to support active aging. We have
introduced the general idea of supporting older people in maintaining themselves
mentally and physically active being helped by some intelligent technology both
during a class and during excursions. The ICT intelligent core makes use of a
specific kind of artificial intelligence technology called automated planning which
is responsible for representing the key ingredients of the ExPLoRAA system,
creating a baseline “lesson” and dynamically adapting it to actions and profile
updates so as to integrate both personalized stimuli and requests to the users
over time. Additionally, the proposed system guarantees an efficient management
of the temporal constraints and, thanks to the introduction of the proposed algo-
rithm, the possibility for the teachers to efficiently change their minds regarding
decisions made on temporal aspects. Although the application is currently un-
der development, the constant contact with Televita’s volunteers is allowing the
transition from a lab prototype to an incrementally more robust version of the
system to be tested in realistic scenarios.
References
1. Allen, J.F.: Maintaining Knowledge about Temporal Intervals. Commun. ACM
26(11), 832–843 (1983)
2. Bacon, L., Cesta, A., Coraci, L., Cortellessa, G., De Benedictis, R., Grilli, S., Polut-
nik, J., Strickland, K.: Training crisis managers with PANDORA. In: Proceedings
of the 20th European Conference on Artificial Intelligence. pp. 1001–1002. IOS
Press (2012)
3. Cesta, A., Cortellessa, G., Benedictis, R.D.: Training for Crisis Decision Making
- An Approach Based on Plan Adaptation. Knowledge-Based Systems 58, 98–112
(2014)
4. Cesta, A., Cortellessa, G., Benedictis, R.D.: Using Training with Older People for
Active Ageing in Time and Space. In: AI*AAL@AI*IA (2017)
5. Cesta, A., Cortellessa, G., De Benedictis, R., Fracasso, F.: A tool for managing
elderly volunteering activities in small organizations. In: Esposito, F., Basili, R.,
Ferilli, S., Lisi, F.A. (eds.) AI*IA 2017: Advances in Artificial Intelligence (2017)
6. Cortellessa, G., Fracasso, F., Sorrentino, A., Orlandini, A., Bernardi, G., Coraci,
L., De Benedictis, R., Cesta, A.: ROBIN, a Telepresence Robot to Support Older
Users Monitoring and Social Inclusion: Development and Evaluation. Telemedicine
and e-Health (2017)
7. De Benedictis, R., Cesta, A.: Integrating Logic and Constraint Reasoning in a
Timeline-based Planner. In: AI*IA 2015 - XIVth International Conference of the
Italian Association for Artificial Intelligence (2015)
8. De Benedictis, R., Cesta, A.: New Heuristics for Timeline-Based Planning. In: Pro-
ceedings of the 6th Italian Workshop on Planning and Scheduling A workshop of
the XIV International Conference of the Italian Association for Artificial Intelli-
gence (AI*IA 2015). pp. 33–48 (2015)
9. Dechter, R.: Constraint Processing. Elsevier Morgan Kaufmann (2003)
10. Dechter, R., Meiri, I., Pearl, J.: Temporal Constraint Networks. Artificial Intelli-
gence 49(1-3), 61–95 (May 1991), http://dx.doi.org/10.1016/0004-3702(91)90006-6
11. Ghallab, M., Nau, D., Traverso, P.: Automated Planning: Theory and Practice.
Morgan Kaufmann Publishers Inc. (2004)
12. Gronlund, N.E., Brookhart, S.M.: Gronlund’s Writing Instructional Objectives.
Pearson/Merrill Prentice Hall (2009)
13. Lecoutre, C.: Constraint Networks: Techniques and Algorithms. Wiley-IEEE Press
(2009)
14. Muscettola, N.: HSTS: Integrating Planning and Scheduling. In: Zweben, M. and
Fox, M.S. (ed.) Intelligent Scheduling. Morgan Kauffmann (1994)
15. Ohlsson, S.: Some principles of intelligent tutoring. Instructional Science 14(3),
293–326 (May 1986), https://doi.org/10.1007/BF00051825
16. Planken, L.R., de Weerdt, M.M., Yorke-Smith, N.: Incrementally Solving STNs
by Enforcing Partial Path Consistency. In: Hoffmann, J., Kautz, H., Brafman,
R., Geffner, H. (eds.) Proceedings of the Twentieth International Conference on
Automated Planning and Scheduling (ICAPS-10). pp. 129–136. AAAI Press, Menlo
Park, CA, USA (May 2010), http://repository.tudelft.nl/view/ir/uuid:3ca1a60e-
8de4-4207-93aa-27e1442ecc18/
17. VanLehn, K.: The Relative Effectiveness of Human Tutoring, Intelligent Tutoring
Systems, and Other Tutoring Systems. Educational Psychologist 46(4), 197–221
(2011)
18. Wenger, E.: Artificial Intelligence and Tutoring Systems: Computational and Cog-
nitive Approaches to the Communication of Knowledge. Morgan Kaufmann Pub-
lishers Inc., San Francisco, CA, USA (1987)