Dynamic Selection of Learning Situations in Virtual Environment

Kevin Carpentier, Domitile Lourdeaux and Indira Mouttapa-Thouvenin

Heudiasyc - UMR CNRS 7253, Universit

e de Technologie de Compi

egne, 60200 Compi

egne, France

Keywords:

Adaptation, Virtual Environment for Training.

Abstract:

In a lot of industrial contexts, workers may encounter novel situations which have never occured in their

training. Yet, such situations must be handeld successfully to prevent high-cost consequences. Such conse-

quences might be human casualties (in high-risk domains), material damages (in manufacturing domains) or

productivity loss (in high performance industry). To address this lack in their training, virtual environments

for training should provide a large spectrum of learning situations. The difﬁculty lies in generating these

situations dynamically according to the learners proﬁle while they have a total freedom of interaction in the

virtual environment. To address this issue, we propose to generate activities by operationnalising the Zone

of Proximal Development in a multidimensional space. The ﬁlling of this space will be updated dynamically

based on user interactions.

1 INTRODUCTION

Nowadays’ working contexts are getting more and

more complex, they are composed of a wide range

of situations. Training is a major issue in indus-

try for different reasons. It prevents accident in

domain where security is critical (high-risk indus-

try, nannies training), it fosters productivity in high-

performance industry (aeronautic assembly, subma-

rine maintenance), it also prevents manufacturing de-

fect where customer satisfaction is a key point. Most

commonly, in a professional environment, operative

attends a short training before getting on the site.

They lack of experience and each new situation is dif-

ﬁcult to handle because it is a whole new one. It is

widely accepted that experience is the mort impor-

tant way to develop professional skills in these do-

mains. By encountering various situations, appren-

tices may consolidate their knowledge and build their

own effective mental representations of the task pro-

cessing. Moreover, it is accepted that situated learn-

ing can offer an efﬁcient learning framework. As such

training is expensive and requires the material to be

requistionned, virtual environment for training have

been proven to be a good solution to provide learn-

ing in complex situations (Amokrane and Lourdeaux,

2009).

By simulating the work context, these environ-

ments deliver a wide range of real situations. How-

ever, providing content is not enough to ensure an ef-

ﬁcient learning. The content must be adapted to the

learner’s proﬁle and historic : what has been learned?

What needs to be learned next? Which errors are most

commonly made? Besides, the content answering

these questions must be provided in an engaging way.

Our goal is to generate pedagogical content adapted

to the learner level and presented through a story in

which the learner will feel involved. The content pro-

posed must enable learners to meet many and varied

kind of situations and keep their motivation at a high

level. To fulﬁll this requirement, we propose to dy-

namically generate relevant learning situations with

regard of the learner’s trace and learning objectives.

A relevant learning situation is a set of states of the

world that will test a subset of skills and knowledge

in a efﬁcient and engaging way. As our works ﬁts in

the situated learning theory, we considered that each

learner builds his own mental representation in disre-

gard of an elicitation of knowledge and skills. Thus,

it makes it difﬁcult to control knowledge acquisition.

Another issue is to ensure that the generated content is

relevant, which means it fulﬁlls both pedagogical and

narrative requirements. This also raised the underly-

ing question about the balance between narrative and

motivational factors and pedagogical needs.

SELDON, standing for ScEnario and Learning sit-

uations adaptation through Dynamic OrchestratioN,

aim to generate and control scenario within a virtual

environment. As part of the SELDON model, we pro-

pose the TAILOR model to generate a canvas which

is a sequence of constraints on the state of the world,

called situations, that should be met or prevented to

101

Carpentier K., Lourdeaux D. and Mouttapa-Thouvenin I..

Dynamic Selection of Learning Situations in Virtual Environment.

DOI: 10.5220/0004247901010110

In Proceedings of the 5th International Conference on Agents and Artiﬁcial Intelligence (ICAART-2013), pages 101-110

ISBN: 978-989-8565-39-6

 2013 SCITEPRESS (Science and Technology Publications, Lda.)

facilitate knowledge learning and skills acquisition.

The canvas is then used by the other part of SELDON,

a scenario planner, DIRECTOR (Barot et al., 2013) to

constrain the simulation. This paper presents our con-

tribution on activities selection based on belief about

learner’s aptitudes and pedagogical needs. In section

2 we present how our contribution positions in rela-

tion to different approaches on adaptive scenarisation.

Then, we introduce the overall process of situation

constraints generation and present a detailled method

of selection of the constraints depending on pedagog-

ical needs in section 3. Section 4 shows an illustration

of this selection through the case of nannies training.

Then we will present the perspectives we foresee to

extend this work and conclude over the whole contri-

bution.

2 RELATED WORKS

Adaptive scenarisation is the process of reacting to

user’s actions to provide content ﬁtted to their need.

In videogames, it might be used to adjust difﬁculty ac-

cording to player’s level without using typical discrete

mode such as ”Easy”,”Hard”, etc. With adaptive fea-

tures, players are always in the ﬂow (Csikszentmiha-

lyi, 1991): the difﬁculty remains high enough to pro-

pose a suitable challenge, yet, players can overcome it

so that they do not get bored or frustrated. Such a con-

cept might be used to adapt difﬁculty in a training ses-

sion so that learners keep a high level of motivation.

The system can propose activities that are always dif-

ﬁcult enough to challenge the learner but always man-

ageable to prevent frustration and loss of motivation.

Our objective is twofold: providing adapted content

(1) and presenting this content in such a manner it

does not cut the user from the ﬂow and, moreover,

motivate him (2).

The adaptation can be made at different levels of

granularity. A ﬁrst approach is to have a global adap-

tation: a whole scenario has been written (Marion,

2010) or generated (Niehaus et al., 2011) and the out-

comes of the events were scripted beforehand. This

approach allows the building of a scaffolding scenario

which present many advantages:

• Pedagogical Coherence: the scenario ensures a

progressive learning through the session, assis-

tance can be given easily at relevant key points;

• Narrative Involvement: it is therefore possible to

unfold the event as a story which will involve the

learner.

A main drawback is the lack of reactivity of the sys-

tem. As the whole session has been planned, the sys-

tem cannot reorient the scenario to adapt to the very

current learner’s state. The only way to cope with it

is to foresee each possible path which can represent a

huge amount of work. An opposite approach is to pro-

vide reactive adaptation by controling the outcomes

of learner’s actions. It enables:

• Dynamic Adaptation: the system triggers out-

comes of learner’s actions and provides assis-

tances depending on pedagogical needs.

For example, in the application V3S (Barot et al., ),

the triggering of a hasardous matter leak is computed

in real-time by HERA (Amokrane and Lourdeaux, ),

an intelligent tutoring system, according to a learner’s

model.

The simulation where the adaptation takes place

can be run with opposite approaches: the controlled

approaches versus the emergent approaches.

The controlled approach aims to provide a very

efﬁcient learning by orchestrating each part of the

simulation: state of the objects, virtual character’s be-

haviours, possibilities of action the learner, etc. It per-

mits:

• Pedagogical Control: each element of the simula-

tion serves the scenario and pedagogical needs.

This approach which is used in the Generic Virtual

Training (Gerbaud et al., 2008) helps building peda-

gogically efﬁcient scenarios but disables the possibil-

ity to encounter unintended - though relevant - situ-

ations. Moreover, such an approach demands an ex-

haustive modeling of the world functionment which

handicaps the evolutivity of the system. The whole

modeling has to be reconsidered to avoid incoherence

each time an author adds new contents. Any attempt

to interfere with the simulation can cause incoher-

ence for the learner: virtual characters become un-

predictible, states of objects changes with no coher-

ent reason. As a result, there is no way to explain

a posteriori the unfolding of events. These explana-

tions are critical for the learner to understand causes

and consequences of events and actions and they can

be provided at the end of the session or reviewed by a

teacher.

By a clever modelling of small behaviors of the

world, emergent approaches allow new situations to

arise (Shawver, 1997). It also enables:

• Freedom of Action: learners are not framed by the

task they are supposed to do, they can experiment

and discover the outcomes of their actions;

• Autonomous Virtual Characters: as they are not

being controlled by a supervisor, virtual charac-

ters maintain their autonomy and their behaviours

remain coherent throughout the simulation run.

ICAART2013-InternationalConferenceonAgentsandArtificialIntelligence

102

The issue with emergent approaches is the lack of

pedagogical control. The simulation runs itself ac-

cording to initial parameters and there is no way to

orchestrate the events to adapt the simulation to the

current learner’s state. Each of these approaches has

attractive features but none of them fulﬁlls our re-

quirements as explained below.

3 PROPOSITION

3.1 Approach

Our work aims to provide a relevant adaptation at dif-

ferent level of granularity. At the lowest level, adapta-

tion should modulate the consequences of the actions

of the learner, this means an even set of events might

have different outcomes depending of the expertise of

the learner. Then, the adaptation must work on a mid-

dle level basis by producing complex sequences of

events leading to a speciﬁc learning in a session. Fi-

nally, skills development requires a learner to follow

a path of different learning situations during different

learning sessions, the adaptation should also provide

information about the path to follow between differ-

ent sessions. For this purpose, we will try to adopt a

balanced approach which is both global and local.

Besides, learners should have a total freedom and

the system must react as it would in reality to help

them to develop skills from their mistakes. Tech-

nical, organisational and human systems are getting

more and more complex in working context. An ex-

haustive explicitation of each possible scenarios be-

forehand would result in a combinatorial explosion.

To adress the growing comlexity of such systems, we

chose to model them through an emergent approach.

However, as our purpose is to provide an efﬁcient situ-

ated learning, we must ensure that relevant assistances

are provided to learners as they would be provided in

a working context. Moreover, we need to orchestrate

dynamically the course of the training to adapt to cur-

rent learner state.This can only be achieved by con-

trolling the ﬂow of events to some extent. We need to

adopt an emergent approach to model the world but

we want to provide pedagogical control over it.

To be able to have both global and local adaption

with pedagogical control over an emergent simula-

tion, we propose to orient dynamically the simulation

towards speciﬁc situations which are consistant with

the current state of the world, without breaking the

coherence of neither object states nor the behaviour

of virtual characters. This is the purpose of SELDON,

standing for ScEnario and Learning situation adapta-

tion through Dynamic OrchestratioN), which is a part

of the HUMANS platform described below.

3.2 HUMANS Framework

The HUman Models-based Artiﬁcial eNvironments

Software platform is dedicated to the simulation of

virtual environments within complex domains where

human factors are critical. HUMANS platform allows

high cognitive virtual characters and learners to coex-

ist in a simulation.

HUMANS uses three models which were designed

to be informed by domain experts (ergonomists, di-

dacticians,etc.):

• DOMAIN (ﬁgure 1) describes the world in a static

way, the object, physical or abstract, that exists in

the world and the relations between them through

an ontology. It also includes a dynamic descrip-

tion: possible actions, the behaviours these ac-

tions trigger and events that might occur through

rules;

Figure 1: Part of an ontological representation of DO-

MAIN.

• ACTIVITY (ﬁgure 2) uses a hierarchical represen-

tation of the task to describe the activity as ob-

served on a real site and not as depicted in proce-

dures and protocols. The basic tasks are the ac-

tions referenced in DOMAIN;

Figure 2: Hierarchical representation of ACTIVITY.

• CAUSALITY (ﬁgure 3) expresses pertinent causal

chains occuring in the environment through a di-

rect acyclic graph. It might describe causal chains

of risks (when informed through a risk analysis)

or errors induction (which can be generated using

an error model generation);

DynamicSelectionofLearningSituationsinVirtualEnvironment

103

Figure 3: Graph representation of CAUSALITY.

These models manipulate common entities and

each unit (Entity in DOMAIN, Task in ACTIVITY and

Event in CAUSALITY) can be tagged to specify some-

thing to which a unit is related (skills, risks, perfor-

mance criteria.etc.). TAILOR is the ﬁrst of two parts

constituting SELDON. It produces constraints for the

second part, DIRECTOR whose role is to apply this

constraints.

3.3 General Overview

As shown in Figure 4, the TAILOR model of con-

straints generation is divided in three parts: diagnosis,

pedagogical selection, narrative framing.

The ﬁrst part updates a dynamic model based on

the Zone Of Proximal Developpement to establish a

diagnosis of learner’s capacities.

Second part computes this model to determine

a set of situation constraints that fulﬁll pedagogical

needs along with metrics on these situations. They

describe if situations shoud be avoided or should be

met. Situation constraints describe states of the world

which should bring learners to discover/develop/use

speciﬁc skills and knowledge. One of these situation

contraints deﬁnes a goal situation toward the simula-

tion should be leading. This situation is not the end of

the scenario but merely one of its key points.

In a third part, key points are then framed into

common narrative patterns to generate a story and

modulate the dramatic tension. The canvas is the suc-

cession of situations constraints build upon time. The

description of the metrics and of the narrative framing

is beyond the scope of this paper.

We present in the following subsection a model

for selecting a goal situation according to a uncertain

learner’s model.

3.4 Description of the Pedagogical

Process

3.4.1 Input Data

The dynamic mecanism of selection of activities un-

derlying TAILOR lies on both inputs from the learner

and the teacher.

Figure 4: General Overview.

• Learner’s Inputs: each session of training the

learner follows is recorded through a trace based

on HERA model (Amokrane and Lourdeaux, ).

As the model is based on activity analysis, traces

identify previous situations encountered, errors,

causes and consequences, risk produced. Traces

are also enriched with activity traces of virtual

characters as well as the tracking of causal chains

within the environment;

• Teacher’s Inputs: the teacher can inﬂuence the

simulation beforehand at different levels: he can

select situations which should be encounter dur-

ing the session in the CAUSALITY model, task

to be performed in ACTIVITY model and perfor-

mance criteria to favour.

3.4.2 Diagnosis

On this purpose, the ﬁrst matter is to establish and

maintain a diagnosis of learner’s current level of

ICAART2013-InternationalConferenceonAgentsandArtificialIntelligence

104

knowledge. As seen in (Brusilovsky and Mill

an,

2007), most systems use an elicitation of knowledge,

of the inﬂuences they have between each other and of

the events which are clues of learning.

Actually, we ﬁt our work in the paradigm of situ-

ated learning, it would be paradoxal to build a model

of skills and knowledge acquisition. Moreover, we

want to produce a progressive learning individualized

to each learner. Vygotsky proposes the model of Zone

Proximal of Developpement (ZPD) (Vygotsky, 1978)

in which a student can develop skills inside its con-

fort zone and enlarge it by the help of the teacher. We

think the ZPD can be used in a more general learning

context than education and that the teacher, which is

responsible of the scaffolding, might be played by vir-

tual scaffolding and by an intelligent scenarisation of

events. To operationnalize this approach, we choose

to deal with a belief the system has in the capabili-

ties of a learner to handle a certain type of situation

depicted by constraints. They might include a set of

skills, performance criteria, tasks, errors, etc., which

are informed by pedagogical experts of the domain

within the models presented in 3.2. These constraints

”mold” a space where each situation is reported with

a belief on whether the learner handled it successfully

or unsucessfully. This space is constructed so that

two points in space are representative of semantically

close situations in the simulation. Beliefs are prop-

agated around each point to estimate a belief on the

capabilities of a learner to handle a situation match-

ing another set of constraints semantically close. The

Transferable Belief Model and the conjunctive rule of

combination (CRC) described in (Smets and Kennes,

1994) are used to represent and udpate these beliefs.

For each point of this space, describing a type of

situation, we have four values:

• h - Belief on the hability to handle this situation,

• d - Belief on the dishability to handle this situa-

tion,

• i - Ignorance, either hability or disability

• c - Conﬂict between belief of hability and disha-

bility

With h + d + i + c = 1

TAILOR parses traces produced by the trace-

based system called MONITOR that exists within

the HUMANS framework. Based on ACTIVITY and

CAUSALITY models, MONITOR aims to record every

action agents makes whether they are real learners or

virtual characters. These actions are linked to task

and high-level tasks in the activity hierarchy and are

associated to a potentiality to trigger an error, a risk

or affecting a performance criteria. Each trace is

used as a source of information to update the beliefs

about a type of situation. New values are compute

according to the application of the conjunctive rule of

combination as shown in 1,2,3,4.

new

= h

cur

∗ h

source

+ i

cur

∗ h

source

+ i

source

∗ h

cur

(1)

new

= d

cur

∗d

source

cur

r ∗d

source

∗d

cur

(2)

new

= i

cur

∗ i

source

(3)

new

= 1 − h

new

− d

new

− i

new

(4)

Where h

cur

, d

cur

, i

cur

are the current values,

source

, d

source

, i

source

are values provided by the trace

and h

new

, d

new

, i

new

, c

new

are the updated values.

The association between beliefs and the multidi-

mensionnal space described above draws our ZPD we

call zpd-space.

3.5 Pedagogical Selection of Activities

As the learner progresses throughout activities and

sessions, the space is ﬁlled with points and associated

beliefs are updated.

TAILOR will then select a set of points in this

space to generate a new situation. The difﬁculty lies

in determining which points will produce an efﬁcient

learning. The selection is made based on the 4 values

aforementionned using pedagogical rules.

• Points where belief has a high ignorance-value are

not likely to be interesting;

• Points where belief has a high hability-value are

not interesting to produce new learning, but they

can be used in the beginning of the session to

make the learner at ease;

• Points where belief has a high dishability-value

are interesting, because they are the proof of an

error, a violation and more generally a miscon-

ception. Depending on speciﬁc pedagogical rules,

the situation will be avoided or, on the contrary,

a learning situation will be generate to break the

misconception through an assistance;

• Points where belief has a high conﬂict-value are

interesting. Mathematically, it means the differ-

ent sources of information are contradictory. In

our case, it means the learner is able to handle a

situation in a speciﬁc context but a misconception

prevent him from using the same skills in another

context.

A set of pedagogical rules helps selecting relevant

points according to these values and pedagogical ob-

jectives. After this ﬁltering, TAILOR compute DO-

MAIN, ACTIVITY and CAUSALITY model to deter-

mine which events and which activities responds to

these constraints.

DynamicSelectionofLearningSituationsinVirtualEnvironment

105

Table 2: ZPD-space initialisation in nannies training simulator.

In this example we consider there is no conﬂict, and that the

learner has successfully handled situation S

. The darker the

color the surer the system knows that the learner will be able

to handle a situation generated from this point. Blank areas

express the lack ok knowledge about the learners habilities.

The ZPD is where the belief on the hability of the learner is

beyond a threshold (0.4 in this case).

In this example we consider there is no conﬂict, and that the

learner has successfully handled many situations.

Table 1: Situations and desirability.

Situation constraints Desirability

Sit

∈ [−1, 1]

... ...

Sit

∈ [−1, 1]

Sit

Goal

3.5.1 Output Data

At each iteration, TAILOR generates a set of situation

constraints associated to a desirability which repre-

sent how desired this state of the world is (see Ta-

ble 1). A negative value describe a situation that

should be avoid. One situation is tagged as the goal

situation. A situation is depicted by a subset of triple

describing a speciﬁc state of the world using the for-

malism of DOMAIN.

4 EXAMPLE: NANNIES

TRAINING SIMULATOR

We applied the model generation to a scenario of nan-

nies training. The french organism for learning pro-

vided a large amount of didactic information for a

previous work on the project SimADVF (ﬁgure 5).

The learner is the nanny and has to take care of two

children: Marion, a six month-old baby and Jean, a

ﬁve year-old boy. Each activity has various indica-

tors about performance criteria. For instance, chang-

ing the nappies of a baby requires Vigilance, Inter-

vention planiﬁcation and Protocol compliance. Let us

consider the case of a novice learner which has very

little use of the training simulation. The skills space

is drawn by performance criteria which values range

from 0 to 5. There are eleven performance criteria

which means a discrete space of 6

values e.g. more

than 360 million types of situation. Fortunately, all

combinations do not deﬁne a valid situation i.e. some

points are pointless and the space is lacunar. For the

sake of the demonstration and to facilitate the visibil-

ity, we shall consider only two performance criteria:

Self Control and Protocol Compliance. The pedagog-

ical objective is arbitrarily set to Self Control. Two ex-

amples are presented in Table 1 to Table 5 for a novice

learner which has used the system once and for a more

advanced learner.

5 PERSPECTIVES AND FUTURE

WORKS

Selection of activities is the ﬁrst part of our work. To

ICAART2013-InternationalConferenceonAgentsandArtificialIntelligence

106

Table 3: Situations generation.

In regard of the pedagogical objectives, which here emphasize

on Self Control, the system selects points to privilege this per-

formance criteria. The point S

(SC = 3, PC = 1) is selected.

By computing the ACTIVITY model, TAILOR determines tasks

requiring a (SC = 3, PC = 1) set of skills. The task ”Dealing

with a capricious child” is an activity that fulﬁlls this require-

ment. Its precondition is the occurence of child in state ”an-

gry”. TAILOR does not control virtual character inner state but

the constraint will be used by DIRECTOR.

In regard of the pedagogical objectives, which here emphasize

on Self Control, the system selects points to privilege this per-

formance criteria. The point S

(SC = 5, PC = 5) is selected,

the system has a belief of 0.4 for the learner to handle this type

of situation. By computing the ACTIVITY model, TAILOR de-

termine the tasks requiring a (SC = 5, PC = 5) set of skills.

The task ”Caring an injured children in emergency situation”

is an activity that fulﬁlls this requirement. Its precondition

is the occurence of a major accident on a children. TAILOR

parses CAUSALITY to determine potential sources of an acci-

dent. There are two possible causes:

• a burn from hot tapwater which may occure when the tap-

water is turned on hot;

• a small fall down the staircase when the gate is open.

By parsing the DOMAIN model, TAILOR determines that each

of these situations can occured in current session: there is a

tapwater and a staircase closed by a gate. As both situation are

possible in current state of the word and in order to constraint

the least possible the simulation, the output is reduced to ”the

occurence of an accident”. Metric considerations, that are not

discussed in this paper, control the gravity of the event. Here,

the gravity is relatively low. In CAUSALITY, an injury on a

baby is tagged as a grave event.

Situations

Goal

Desirability:1

States:

(?child has-state ?angry)

(?angry has-level high)

Situations

Goal

Desirability:1

States:

(?child has-accident ?acc)

(?acc has-gravity major)

Situation1

Desirability:-1

States:

(?child has-accident ?acc)

(?acc has-gravity mortal)

DynamicSelectionofLearningSituationsinVirtualEnvironment

107

Table 4: ZPD-space udpate and new activities selection.

The learner reacted well to the previously generated situation

, the ZPD-space is updated. The point S

(SC = 4, PC = 3)

is now selected, the system has a belief of 0.4 for the learner

to handle this type of situation. By computing the ACTIVITY

model, TAILOR determines tasks requiring a (SC = 4, PC = 3)

set of performance criteria. The task ”Caring an injured chil-

dren” is an activity that fulﬁlls this requirements. Its precon-

dition is the occurence of a minor accident on a children. TAI-

LOR parses CAUSALITY to determine potential sources of a

minor accident. A cause of a minor accident is a child stum-

bling on an toy. It can occur when child is in state ”excited”

and the room is messy. By parsing the DOMAIN model, TAI-

LOR determines that the situation can occured in current ses-

sion. The output is the child to be excited and a room to be

messy.

In the trace providing by the learner tracking module a deletion

error was detected: the learner did not performed a subtask of

the high-level task ”Caring a children in emergency situation”.

Indeed, the learner should have call emergency services or, at

least the parents, to warn them and to ensure no further healing

were to be done. The situation was not handled successfully,

so the ZPD-space is updated with a belief on the inhability of

the learner to handle a situation such as (SC = 5, PC = 5). To

correct this behavior, TAILOR computes CAUSALITY to deter-

mine which event are likely to learn the user the good prac-

tises. It occured when parents are home, in a worried state to

rebuke nannies for their misbehavior.

Situations

Goal

Desirability:1

States:

(?child has-state excited)

(?room is-a room_object)

(?room has-state messy)

Situation1

Desirability:-1

States:

(?child has-accident ?acc)

(?acc has-gravity major)

Situations

Goal

Desirability:1

States:

(?parent at home)

(?parent has-state ?worried)

(?worried has-level ?high)

provide adapted content is essential but to involve the

learner, we need to use motivational factors. Modulat-

ing the dramatic tension is a possible solution. To cre-

ate tension, a story must be built upon the events and

the world depicted within the simulation. The aim is

to provide a purpose to the training session by show-

ing the virtual characters as story characters who can

be helpers or opponents, the events as plot points that

will increase or decrease the tension. We plan to use

narrative pattern, as described by (Campbell, 2008),

(Propp, 1968) or (Greimas, 1966) to extends current

pedagogicial situations. The element described by the

situations will be ﬁtted with element from a pattern

such as location, helpers, opponents, goal,etc. For

ICAART2013-InternationalConferenceonAgentsandArtificialIntelligence

108

Table 5: End of training session.

The learner reacted well to the previously generated situation

, the ZPD-space is updated. The information provided by the

ZPD-space will be used to initialise further sessions of train-

ing.

The learner made a mistake by not calling emergency services.

The early arrival of the parents provided a diegetic assistance.

The same type of situation is likely be proposed in a future ses-

sion to test whether the learner has learned the good practises

or not.

Figure 5: A screenshot of SimADVF - Nannies training

simulator.

one pedagogical situation, many narrative conﬁgura-

tions are possible. We will use a measure of the nar-

rative utility based on earlier events in the simulation.

The utility will maximize if it furthers the develop-

ment of the story depicted in past events without dis-

rupting the whole coherence.

6 CONCLUSIONS

We proposed in this paper a model to dynamically

generate scenarios in a virtual environment regarding

learner’s capacities. The process uses a phase of di-

agnosis which compute traces at the initialisation and

in real time. It operationnalizes the theory zone of

proximal development through a multidimensionnal

space of belief, updated at each task performed by the

learner within the virtual environment. Then the sys-

tem computes current world state to determine which

situation can take place to answer ZPD and pedagogi-

cal objectives requirements. We build a ﬁrst prototype

within the HUMANS platform working on the exam-

ple of nanny training. Our future works will focus in

the narratives consideration by framing the successive

situations in a narrative pattern to relate a story.

ACKNOWLEDGEMENTS

This work is part of the ANR project: NIKITA (Nat-

ural Interaction, Knowledge and Immersive systems

for Training in Aeronautic). Partners are: Heudiasyc,

Paris Descartes University, CEA-LIST, Emissive and

EADS. We want to thank M. Andribet and G. Michel

from the french association for adult training (AFPA)

who provided enough data to ﬁll our models.

REFERENCES

Amokrane, K. and Lourdeaux, D. Pedagogical system in

virtual environment for high-risk sites. In Proceedings

of ICAART 2010.

Amokrane, K. and Lourdeaux, D. (2009). Virtual reality

contribution to training and risk prevention. In Pro-

ceedings of ICAI 2009.

DynamicSelectionofLearningSituationsinVirtualEnvironment

109

Barot, C., Burkhardt, J.-M., Lourdeaux, D., and Lenne, D.

V3S, a virtual environment for risk management train-

ing. In Proceedings of JVRC11.

Barot, C., Lourdeaux, D., and Lenne, D. (2013). Dynamic

scenario adaptation balancing control, coherence and

emergence. In Proceedings of ICAART 2013.

Brusilovsky, P. and Mill

an, E. (2007). User models for

adaptive hypermedia and adaptive educational sys-

tems. Lecture Notes in Computer Science.

Campbell, J. (2008). The Hero With a Thousand Faces.

New World Library.

Csikszentmihalyi, M. (1991). Flow: The psychology of op-

timal experience. Harper Perennial.

Gerbaud, S., Mollet, N., Ganier, F., Arnaldi, B., and Tis-

seau, J. (2008). GVT: a platform to create virtual en-

vironments for procedural training. In IEEE Virtual

Reality.

Greimas, A. J. (1966). S

emantique structurale : recherche

et m

ethode. Larousse edition.

Marion, N. (2010). Mod

elisation de sc

enarios

edagogiques pour les environnements de r

ealit

virtuelle d’apprentissage humain. PhD thesis,

Universit

e de Bretagne Occidentale.

Niehaus, J. M., Li, B., and Riedl, M. O. (2011). Automated

scenario adaptation in support of intelligent tutoring

systems. In Proceedings of the 24th FLAIRS.

Propp, V. I. (1968). Morphology of the Folktale. University

of Texas Press.

Shawver, D. (1997). Virtual actors and avatars in a ﬂexi-

ble user-determined-scenario environment. In Virtual

Reality Annual International Symposium, 1997.

Smets, P. and Kennes, R. (1994). The transferable belief

model. Artiﬁcial Intelligence, 66(2):191 – 234.

Vygotsky, L. S. (1978). Mind in Society. Harvard Univer-

sity Press, Cambridge, MA.

ICAART2013-InternationalConferenceonAgentsandArtificialIntelligence

110