Reasoning on Uncertainty in Smart Environments

Alencar Machado

1,2

, Vinícius Maran

, Iara Augustin

, João Carlos Lima

Leandro Krug Wives

and José Palazzo Moreira de Oliveira

Colégio Politécnico, Universidade Federal de Santa Maria (UFSM), Santa Maria, Brazil

Programa de Pós-Graduação em Informática, Universidade Federal de Santa Maria (UFSM), Santa Maria, Brazil

Coordenadoria Acadêmica, Universidade Federal de Santa Maria (UFSM), Santa Maria, Brazil

PPGC Instituto de Informática, Universidade Federal do Rio Grande do Sul (UFRGS), Porto Alegre, Brazil

Keywords: Semantic Web, Uncertainty, Smart Environments, Probabilistic Ontologies.

Abstract: Currently, There Is a Convergence of Systems for Smart Environments and Uncertainty Reasoning. Context

Models Are Being Proposed to Support the Detection of Situations in These Environments. However,

Reasoning to Detect Situations Taking into Account Uncertainty Presents a Great Challenge. This Paper

Describes a Contextual Model based on Semantic Web Technologies That Can Deal with Uncertainty. This

Framework May Be Used to Detect Unwanted Situations with a Certain Grade of Contextual Uncertainty.

the Model Was Evaluated in a Scenario Reasoning over Uncertain Data to Predict Unwanted or Perhaps

Dangerous Situations.

1 INTRODUCTION

Homes are becoming Intelligent Environments, able

to assist people who live into it. These systems are

planned to act according to the user profile, and with

complex physical environment, where objects are

added, updated or changed of location. The user

profile also changes over time. For instance, users

may suffer from different diseases during their lives,

and these can affect the interaction with the objects

of their residence. Cognitive problems, such as

forgetfulness, can sometimes put the user into an

unwanted or perhaps dangerous situation; for

instance, forgetting the stove on after preparing a

meal.

Systems for Ambient Assisted Living need to

interpret the context in which the user lives to be

able to act in advance. Using ontologies to represent

the contextual model is the most complete and

expressive way to support reasoning about the user

context for intelligent systems (Strang and Linnhoff-

Popien, 2004) (Bettini et al., 2010). The ontologies

have some constructors that give support the

reasoning about the model domain. Some studies

show contextual models for different domains

(Sixsmith et al., 2009) (Tazari et al., 2010). These

efforts search to model the user's context and

support systems for decision making on situations of

interest. Some works try to support systems in

reasoning about uncertainty (Coronato, 2012)

(Coronato and De Pietro 2013) (Rasch et al., 2011)

(Forkan et al., 2015). These works present

contextual models using ontologies that are

implemented in Ontology Web Language –

Description Logic (OWL-DL).

The main limitation for reasoning about

uncertainty in OWL-DL is the conceptual

foundation of DL, a subset of First Order Logic

(FOL). FOL defines sentences, axioms, that are

always true logical statements about the domain to

which they are representing (Costa et al., 2011)

(Laskey, 2008). However, for systems that seek to

detect situations and make decisions in real

environments, where data may have an error rate,

low but present, it becomes imperative to support

detection of situations and decision-making on

uncertain data. Also the information obtained from

the sensors must be interpreted and this

interpretation normally is surrounded with

uncertainty.

To manage these challenges, some research

efforts try to manipulate information modelled in

ontologies and submit this information to prediction

algorithms using fuzzy logic, neural networks, or

Bayesian networks (Coronato and De Pietro 2013).

These strategies have purely statistical results as the

240

Machado, A., Maran, V., Augustin, I., Lima, J., Wives, L. and Oliveira, J.

Reasoning on Uncertainty in Smart Environments.

In Proceedings of the 18th International Conference on Enterprise Information Systems (ICEIS 2016) - Volume 2, pages 240-250

ISBN: 978-989-758-187-8

prediction is made without the semantic information

contained in the ontological representation.

Consequently, it is necessary that the ontological

contextual models represent the context uncertainty.

Also, it is necessary to have a native support for

reasoning about uncertainty within the context

model. This approach gives more possibility to the

detection of situations and decision making made

with incomplete information.

This paper presents a context model using for

reasoning over uncertainty in smart environments.

More specifically, we employed a Multi-Entity

Bayesian Network (MEBN) for the developing of

semantic model to support the prediction of

unwanted situations in smart environments. The

model is evaluated through empirical experiments

using a case study where an unwanted situation is

simulated. In the case study is presented the use of

the context model which supports reasoning about

uncertainty. It is shown as the model supports the

automatic generation of Bayesian Networks in a

specific situation depending on the available

evidence within the ontology.

This paper is organized as follows. Section 2

discusses the background and related work. Section

3 presents the model for developing the context of

AAL with reasoning with uncertainty in smart

environments. Section 4 describes the developed

case study. Finally, Section 5 discusses our

conclusions and future work.

2 BACKGROUND AND RELATED

WORK

The user interaction with devices in a real-world

environment is uncertain by nature. This

communication is influenced by several factors, such

as the knowledge that the user has about using a

device, the cognitive ability, and the humour of the

user, among others.

There are big challenges for the design of AAL

systems. Currently, the context aware area is

consolidating systems for smart environments.

Context can be understood as “the environment in

which the system operates” (Ye et al., 2011). Dey

and Abowd (1999) characterize context as the

situation of an entity in an environment. The concept

of situation is utilized to characterize the state of the

user environment.

Ye, Stevenson and Dobson (2011) define

situation as the abstraction of the events occurring in

the real world that are derived from the context and

hypotheses about how the observed context relates

to factors of interest. Systems that manipulate the

context detecting situations of interest through

events generated by user’s interaction must handle

uncertainty to assist correctly users in their living

environment.

2.1 Reasoning on Uncertainty

Reasoning for detection of situation and decision

making with uncertain data comprises some

shortcomings related to modeling with regard to the

modeler conception of the world. The modeling

process may have misconceptions and possibly will

be addressed through consistency checks of the

ontologies. Hence, this fact must be taken into

consideration and can be treated through

probabilistic reasoning.

Probabilistic reasoning might support the

processing of uncertainty. It is possible to make

predictions of future situations taking uncertainty

into account. Many works that focuses on the

prediction phase has been published and they present

algorithms to identify the future with an acceptable

accuracy (Paganello and Giuli, 2011).

Ontologies provide a range of features that

search for represent the environmental context in a

broad and expressive form. Context modeling using

ontologies is currently done through the use of the

Ontology Web Language (OWL). Among its

variations (Lite, DL and Full), the most widely used

by the possibility of decidability is the Description

Logics (DL). The interpretation of a theory

determines the definition of each constant, predicate

and symbolic function in relation to the area. Each

symbolic constant denotes a specific entity; each

predicate denotes a group containing entities that the

predicate holds, and each symbolic function is a

function defined in the domain. The logical

sequences of a set of axioms consist of sentences

that are always true in all interpretations, also called

true sentences. Due to sentences be always

interpreted as logical assertions, DL is not suitable

for areas where there are uncertainties in relations

among the concepts (Laskey, 2008).

One way to model uncertainty is the use of

probability, and a suitable model for existing

domains is the Bayesian Networks (BN). These are

Directed Acyclic Graphs representing a distribution

function of joint probabilities of variables in a

domain of interest. Each Bayesian network consists

of nodes (random variables) and edges connecting

these nodes. These links represent the influence

from one node (ancestor) in relation to another node

Reasoning on Uncertainty in Smart Environments

241

(descendent) generating a directed arc. Each node

has a Conditional Probability Table (CPT) to

calculate the influence of a parent node "x" in

relation to its influenced node "y", and the joint

probability distribution is measured by the influence

of every parent in relation to a leaf node (Friedman

et al., 1997).

For Semantic Web applications, BNs have the

potential to provide a powerful, compact structure

for probabilistic inference mechanisms. However,

BNs have some key limitations. The first is that the

number of variables must be known in advance (i.e.

number of nodes is fixed). However, many domains

require reasoning about numbers and types of related

entities, where the relationships among entities

cannot be specified in advance or are uncertain. The

second limitation is that the language used to specify

BNs is not powerfull enough to express significant

problems with repeated structure. The third one is

that a BN is a directed acyclic graph, and hence no

native support for recursion is provided (Costa et al.,

2011).

2.2 Probabilistic Ontologies

The area of probabilistic ontologies appeared with

the objective of using the expressive power of the

First-Order Logic and the treatment of uncertainty

supported by Bayesian Networks. In general, there

are two approaches to the generation of probabilistic

ontologies. The first consists in representing

uncertainty by probability values described as

annotations, such as (Yang and Calmet, 2005).

However, annotate ontology with numerical

probability is not enough, and some information is

lost with the lack of representation capturing the

structural constraints and dependencies between

nodes (Laskey, 2008). The second alternative

consists of using a First-Order Probabilistic

Language, which combines aspects of probabilistic

representation with first-order logic (Howard and

Stumptner, 2014).

According to Costa (2005), a probabilistic

ontology must be able to properly represent: the (i)

types of entities that exist in the field; the (ii)

properties of these entities; the (iii) relationships

between entities; the (iv) processes and events

happening with these entities; the (v) statistical

regularities that characterize the domain; the (vi)

inconclusive knowledge, ambiguous, incomplete,

unreliable related to domain entities; and (vii)

uncertainty over all previous forms of knowledge. It

should be noted that the term entity refers to any

concept that can be described and reasoned in an

application domain. Probabilistic ontologies are used

to describe comprehensively the knowledge of a

domain and associated uncertainty, structured and

shareable, preferably in a format that can be read

and processed by the computer (Fenz, 2012).

Howard & Stumpter (2014) compare those

languages (First-Order Probabilistic Languages) in

relation to (i) aspects for handling uncertainty, (ii)

structural support related to types of inheritance (iii)

types of fields on which the language may represent

and (iv) reasoning techniques on a group of entities

and relationships. An overview of this comparison is

presented in Table 1.

In Table 1, attribute means uncertainty about the

attributes of entities and relationships. Numeric

indicates the uncertainty of numerical data entities in

a domain. Reference means uncertainty about the

relationship between domain entities. Existence

means uncertainty about the existence (or not) of

entities and their relationships in the area. Types

mean when an entity of interest is identified, but it

can be one or more of possible subtypes. Analyzing

data present in Table 1, the languages Probabilistic

Relational Models (PRMs), Object-Oriented

Probabilistic Relational Modeling Language

(OPRML) e Multi-Entity Bayesian Networks

(MEBNs) are compared, and in this paper we choose

to use the MEBNs language to represent uncertainty.

Table 1: Languages for Uncertainty Representation.

PRMs MEBNs OPRML

Uncertainty Attribute X X X

Numeric X X X

Reference X X X

Identity X X

Existence X X X

Type X

Inheritance Simple X X

Multiple X

Domain Static X X X

Recursive X X X

Dynamic X X

Adapted from Horward & Stumpter (2014).

To make possible the use of MEBN in Semantic

Web an OWL extension was created through an

upper ontology called Probabilistic Ontology Web

Language (PR-OWL). That extension expresses a

probability distribution on interpretations of any

first-order theory. PR-OWL was built to be

interoperable with non-probabilistic ontologies.

However, the probabilistic definitions of an ontology

have to form a theory about the fragments of the

complete or partial valid world (Carvalho et al.,

ICEIS 2016 - 18th International Conference on Enterprise Information Systems

242

2013).

In Figure 1, the concepts of the upper ontology

are presented, where ellipses represent general

classes while arrows represent the main relations

between these classes. A probabilistic ontology must

have at least an individual's MTheory class, which is

formed by a group of MFrags that collectively form

a valid MTheory. In the next section, we discuss the

Multi-Entity Bayesian Network theory (Carvalho et

al., 2013).

Figure 1: PR-OWL Adapted from Costa (2005).

2.3 Multi-Entity Bayesian Network

Multi-Entity Bayesian Network (MEBN) represents

the world as composed of entities that have

attributes and which are associated with other

entities. Knowledge about the attributes of the

entities and their relationships are represented as a

collection of MEBN fragments (MFrags) organized

in MEBN theories (MTheories). An MFrag

represents a distribution of conditional probability

for instances of random variables about their parents

(parent node) in the same MFrag. An MEBN theory

is a set of MFrags that collectively satisfy

consistency constraints, ensuring the existence of a

single joint probability distribution on instances of

the random variables represented in each MFrag.

MEBN integrates the semantics of the standard

theoretical model of first-order logic with random

variables, as formalized in Bayesian Networks

(Howard; Stumptner, 2014).

Such as in a BN, one MFrag contains nodes that

represent random variables arranged in a directed

graph whose edges represent relations of direct

dependence. An isolated MFrag can be compared

with a standard BN with known values to their root

nodes and local distributions of its nodes that are not

root. A node in an MFrag may have a list of

arguments in parentheses. These arguments are

placeholders for the authorities in the field (Costa,

2005).

According to Laskey (2008), an MFrag consists

of three types of nodes: (i) the residing nodes (object

property in OWL-DL) have local distributions that

define how their probabilities depend on the values

of their parents in the graph. In a complete MEBN

theory, each resident node has exactly one MFrag

where their local distribution is set; (ii) the input and

context nodes can influence the distribution of the

resident nodes, but their distributions are set in their

own MFrags. Finally, (iii) context nodes represent

conditions that must be satisfied for the influences,

and local distributions of an MFrag can be applied.

These conditions are boolean values, which may

have true, false or absurd values.

An MEBN does not specify a standard for

Conditional Probability Table (CPT). However, as in

a default BN, a CPT summarizes statistical

regularities that characterize a domain. These

regularities are captured and coded in a knowledge

base using a combination of expert opinions and/or

learning from observation. For more information

about MEBN and PR-OWL, we suggest reading the

following works (Laskey, 2008) (Costa, 2005)

(Carvalho et al., 2013).

Most projects focused on Ambient Assisted

Living or, more specifically in Smart Homes (SH)

are interested in proposing models to provide

services. A reactive requirement drives the execution

of services, always after a fact has occurred.

Therefore, they do not show proactive behaviour to

adapt to the user environment. Related works are the

ones shown in Table 2: Soprano (Sixsmith et al.,

2009), Persona (Tazari et al., 2010), SM4ALL

(Rasch et al., 2011), Uranus (Coronato, 2012),

CoCaMAAL (Forkan et al., 2015).

Table 2: Related Work.

Per Sop SM4 Ura CoC

AAL X X X X X

SH X X X

Sensor X X X X X

Event X X X

Situation X X X X

Action X X X X

Activity X

Uncertainty

These works show a semantic context model, but it

is observed that these models have a reduced

expressiveness. Those who address the uncertainty

show a hybrid model to generate probability using

techniques without using Semantic Web

technologies. Therefore, it is important to develop a

model to support reasoning about uncertainty in the

AAL domain because none of the related work

addresses these characteristics, fully supported by

Semantic Web technologies, to sensitive systems on

the Situation-Awareness in Smart Environments.

Reasoning on Uncertainty in Smart Environments

243

3 DEALING WITH

UNCERTAINTY IN SMART

ENVIRONMENTS

Systems for smart environments need to know the

user environment, and where necessary, implement

assistance actions. Currently, the main source for

real-time data collection is data obtained from

sensors. They collect raw data without semantic

characterization and with error rates adding

uncertainty in the collected data. Therefore, it

becomes necessary to consider this data associated

with data from other entities such as people, rooms’

characteristics, and electrical networks state, among

others. It is possible to group data to generate useful

information from a higher level, i.e., detecting of an

environmental or user situation, as cold or

emergency. In this paper, the vision of the

environment perceived by the system is obtained

from the data captured by the sensors; these inputs

are aggregated with environmental contextual

entities generating useful information. The actions

triggered by the system for the environment are

realized by Web services, which are associated with

objects like smart phones, televisions, microwaves

ovens, and others in the living environment. In

Figure 2, is presented the information and decision

flow.

Figure 2: Adapted from Machado et. al, 2013.

The system starts collecting raw data from sensors

and aggregates the information associated with those

entities to generate higher-level information that is

used to characterize the state (situation) of the

environment. With this characterization, the system

can do a decision to act in the environment using the

capabilities (services) provided by the available web

services automating the environment according to

the situations of interest and user preferences.

3.1 Ontology Network for AAL

Other works on situation-awareness for smart

environments are directed to modeling concepts that

relate to the situation itself, paying little or no

attention to the modeling of other concepts of an

intelligent environment. In this article, we try to

approximate the home automation model (essentially

describes semantic relations between the structure of

the physical environment) with the user model and

higher-level information such as events, situations

and actions.

The ontology network for AAL was described

using Web Ontology Language using the Protégé

software. The probabilistic fragments that make up

the Reference Model for Systems to Predictive

Situations-awareness in AAL was developed using

UnBBayes software. Intra networks relations were

established with OWL-DL owl:subclass and

owl:import resource and other relationships were

implemented using DOOR ontology (Allocca;

D'Aquin, 2009).

A simplified form of each modelled network is

presented in Figure 3; however, for easy viewing

some entities and their relationships were omitted.

This structure may be modified to incorporate new

concepts, allowing the inclusion of new entities in

different domains. The objective is to construct a

model that describes an automated home

environment, fully controlled by a middleware to

support a home care environment.

Analysing Figure 3, the user performs Actions

(Human Action). These result in External Events

collected by the system. Events start and finish the

Current Situation involving the user in the current

time. The same events influence a Predictive

situation that may involve the User in the future.

Using the information about Current and

Predictive Situation, the system can select

Automated Actions to handle situations of interest.

For example, if is necessary to handle the situation

including the interaction with a User, the system

should choose an Automated Action Type Regarding

Person.

This action will be performed by a functionality

provided by a device of the type Interaction with

Person. This functionality must be sensitive to the

disability presented by the User. The Automated

Actions produce Internal Events and analysing these,

the system can detect if the Current and Future

Situation change or will change in relation to a User.

ICEIS 2016 - 18th International Conference on Enterprise Information Systems

244

ACTION

AUTOMATED

ACTION

HUMANAC TIO N

INTERNAL

EVENT

EXTERNAL

EVENT

CURR EN T

SITUATION

PREDICTIVE

SITUATION

generate

start

end

influence

USER

DEVICE

USER_PROFILE

FUNC TIONALITY

INTERACTION

WITHPERSON

INT ERACTIONWITH

ITSELF

INTERACTIONWITH

ORGANIZATION

REGARDINGA

ORGANIZ ATION

REGARDINGA

PERSON

REGARDINGA

ITSELF

executedBy

hasProfile

isSensitiveTo

isSituationOf

willBeSituationOf

HEAL TH

CAPACITY

PREFERENCES

ROLE

SPECIALITY

VISUA L

HEAR ING

MOTOR

COGNITIVE

DESABILI TY

VISITANTCAREGIVER PATIENT

FORMAL

INFORMAL

HEAL TH 

CONDITION S

HEAL TH PROVIDER

STATE

useDevice

ACTIVITY

TASK

GOAL

hasTask

hasFunctionality

APPLIANCE

ACTUATOR

SENSOR

isComposedBy

ObjectProperty

subConceptOf

Concept

Figure 3: Ontology Network for AAL Environments.

3.2 Reference Model for Predictive

Situations in Smart Environments

Represent uncertainty in ontologies is not the same

as to build a probabilistic system. In this particular

case, the probabilistic part refers to the semantic

relationships modelled using the PR-OWL settings

that collectively form a MEBN theory. There is no

need for all the relations of an ontology to be

probabilistic, however the parts modelled with PR-

OWL extension should form a valid MEBN theory.

The semantic relationships that extend PR-OWL

resources are shown in Figure 4.

TASK USER

hasTask(us,tar)

EVENT

PREDICTIVE

SITUAT ION

influence(ps,t)

willBeSituationOf(ps,t,us)

AUTOMATE D

ACTION

hasReactiveAction(ps,aa)

OWL‐TIME

influence(ps,tPrev)

runningTask(tas)

automatedAction

Perfomed(ps)

Figure 4: Probabilistic Ontology for proactive domain.

In this figure the probabilistic ontology for

predictive situations is represented graphically. In

this probabilistic ontology, the semantic relationship

contains parameters derived from entity instances of

ontology. The definition of the semantic relationship

with parameters is inherited from the PR-OWL

model. In this case, the instances of entities are

random variables that will feed the reference model

to Predictive Situation and form a valid theory. In

this probabilistic ontology, events are affected when

the user is involved by some Task running

(runningTask(tar)) and/or Automated Actions taken

in relation to Predictive situation

(AutomatedActionPerformed (ps)).

The recursion for the model is promoted by the

Temporal Entity defined in OWL-TIME ontology

(Hobbs et al., 2004). This concept can be used as a

discrete concept, so representing subsequent steps

orderable of "T0" to "Tn" rather than continuous

scales. Therefore, an event can influence a predictive

situation at an earlier time by the relationship

influence(ps, tPrev) and influence a predictive

situation at time t. The willBeSituationOf(ps, t, us)

gives the probability of a Predictive situation ps on a

time t involving user us. Therefore, it can perform

prediction of a situation over time.

So the question that the probability part of the

ontology must answer is: “what is the probability of

a situation at a specific time, involving the User in

his living environment?”

In the Figure 4 is show the MBEN theory of

reference for predictive situation-aware systems.

The systems using this model can answer the

question by resident node willBeSituationOf(ps, t,

us). This reference model is a repeatable structure

(template) that generates Bayesian Networks

according to user situation. The Specific Situation

Bayesian Networks are generated according to the

semantic relationships that are linked to the user's

instance within the ontology. Therefore, at runtime,

when exists a need to generate a Bayesian network,

the structure of the network is dynamically

generated according to the reference model and the

Reasoning on Uncertainty in Smart Environments

245

user context that are represented in the ontology.

The model is composed of fragments, described

below, and the random variables related to the

context nodes are: "tar" to task, "us" to User, "aa"

for Automated Action, "ev" for Event, "ps" for

Predictive Situation and "tPrev" and "t" for

orderable Discrete Times (instances of OWL-

TIME), the resident nodes will be described as

<resident node name> = {states}.

The Task MFrag present the probability of a

specific task to be in execution in the environment,

and this fact is represented by runningTask(tas) =

{True, False} resident node. The Automated Action

MFrag presents the probability of an automated

action to be executed in relation to a predictive

situation, and this fact is represented by

automatedActionPerformed(ps) = {Automated

Action} resident node. The TaskUser MFrag

describes, through the resident node

hasTask(us,tas)={True, False}, the probability of a

user be involved by a task. The Predictive Situation

MFrag generates the local probability distribution of

a particular predictive situation involving a user at a

specific time by the resident node

willBeSituationOf(ps, t, us) = {True, False}. Such

distribution is calculated through of the influence of

the resident node influence (ps, t)={events}, whereas

the states of resident node influence must be

instances of events. So this way is possible to

calculate the probability of a particular event

happens when being influenced by task and

automated action. This context node applies the

probability distribution that an Event influences a

predictive situation in time t, when affected after by

the input nodes runningTask(tas),

automatedActionPerformed(ps) and its own

distribution in a previous time influence(ps, tPrev).

This reference model is a repeatable structure

that instantiates specific Bayesian Network

according to existing evidence. The reasoning

process on the Reference Model for Predictive

situations is intended to the generation (queries) of

Specific Situation Bayesian Network (SSBN).

Determining the values (probability) of a set of

queries the systems can query the probability of a

situation will be situation of a User at a specific time

in the future. Thus, the system has the possibility to

choose what proactive actions to be triggered in

relation to the probability values generated by virtue

of a query generated of the Specific Situation

Bayesian Network.

4 CASE STUDY

A scenario is the complete description of a

contextualized user’s routine. This case study is

based in a scenario that demonstrates how the

probabilistic model developed in this article is

applied in smart environments. To analyze the

model, it was used a Pervasive Application and a

prototype of Situation as a Service (SIaaS)

middleware described in Machado et, al. 2014.

Thus, the scenario described is typical of an

Ambient Assisted Living environment, consisting of

an unwanted situation. Different aspects of user’s

interaction with the object’s residence are identified

to generate events that can determine the beginning,

the end or influence the rise of unwanted situations.

These features result in the execution of proactive

actions to handle this situation. The scenario is

intended to generate data for the detection of

unwanted situations and actions performed to deal

with these situations, and the events generated

(effects of these actions).

Imagine John’s, a 75 years old citizen who has

some aging associated diseases such as diabetes,

hypertension and lightweight dementia. John's

residence consists of a living room and kitchen,

among other space. This type of patient tends to

forget in what activity is immersed, it is common to

start an activity and forget that he was doing

another, for example, cooking and watching

television, or even confuse the time of day and go to

sleep. Therefore, the family of John buys a

Pervasive Application to assist John in daily

activities. The focus of this scenario is the activity of

cooking, where John interacts with the stove device

(task use stove). According Blasco et al. (2014)

older people are a group with highest vulnerability

to accidents, especially in their homes. The vast

majority of domestic accidents are related to

activities in the kitchen: kitchen utensils, cutlery and

appliances are the most dangerous utensils. As a

result of these accidents, older people lose

confidence in their abilities, lowering their self-

esteem and, consequently, in many cases, deciding

to live in a nursing home.

Therefore, to avoid such features, in John's

kitchen was installed a smart stove with the

following features: (i) identify which user used / are

using the stove, by detection of digital; only after

this step the stove frees up resources, e.g. turning on

the stove; and (ii) which was the last time (time) in

which a person neared the stove, or even (iii) the

functionality to automatically turn off the stove.

Therefore, imagine that John is watching

ICEIS 2016 - 18th International Conference on Enterprise Information Systems

246

television in the living room, but moves to the

kitchen and decide to cook. John organizes the

preparations, he turns on and put a pot on the stove

and, after it, he listen news on TV that is of his

interest. So, he comes out of the kitchen and back

into the living room, sitting on sofa in front of the

television. An application that seeks to assist users

affected by the state of Senile in their daily activities

should interfere in everyday life as little as possible.

Therefore, in this scenario, the application has

interest in being notified if John forgot the stove

turned on, so putting him and his family in an

unwanted situation as a dangerous situation in

cooking when using the stove. So, the application

can make decision for trigger proactive actions and

to manipulate the unwanted situations. Table 3

presents the actions and events produced by the

scenario.

Table 3: Actions of appPervCook.

Actions

ac1: Warn User; ac2: Notify Caregiver that stove is turned on;

ac3: Turn off the stove automatically

Events

ev1: User forgot stove on; ev2: Stove off;

ev2.1: Sensor detected presence near to the stove

The ev1 event starts the unwanted situation, and the

ev2 and ev2.1 events finish the situation. The actions

can be used with a reactive or a proactive way. To

handle this situation, a reactive behaviour of the

application could be making decision to turn off the

stove automatically or alert the caregiver to perform

this action or wait more time. For a proactive

behaviour, the application could have the ability to

predict whether John, when turn on the stove, he will

forget this on. This characteristic is promoted by the

Reference Model for Predictive situation-aware

system. The local probabilistic distribution for each

MEBN fragment must be generated with the help of

an expert in Senile (in a real scenario), or by

machine learning. This model can be started with

historical data of people affected by this disease, or

even after John turned on the stove a few times.

Thus, making it possible to identify a behavioral

pattern and identify the probability of John forget

the stove on while using it. In the fictitious scenario,

if John turn on the stove, there is possibility of forget

it on. In Table 4, is presented the distribution for

each fragment MEBN.

The resident node runningTask(tas), describes

the probability that there is a running task in the

environment. For this node, the article used values

that do not affect other nodes residents (50% true

and 50% false). The node hasTask(us, tas)

establishes that 92% of the time there running task

the user is involved. The resident node

automatedActionPerformed(ps) shows that directly

warn John (ac1) was executed 15% of the time.

Notify the caregiver (ac2) by 50% and automatically

turn off the stove (ac3) by 35%.

Table 4: Local probability distribution for resident nodes

of appPervCook.

esident: automatedActionPerformed(ps)

[ac1=0.15, ac2=0.5, ac3=0.35]

esident:

runningTask(tas)

[true=0.5,

false=0.5]

esident: willBeSituationOf (ps,t)

If any ps has (influence = ev1) [true=0.99,

false=0.01] else If any ps has (influence = ev2)

[true=0.01, false=0.99] If any ps has (influence

=ev21)

[true=0.45,false=0.55]else[true=0.5,false=0.5]

esident:

hasTask (us,tas)

[true=0.92,

false=0.08]

Resident: influence (ps,t)

If any ps has (automatedActionPerformed = ac1) If any tas has

(runningTask = false) [ev1=0.05, ev2=0.35, ev21=0.6] else

[ev1=0.75, ev2=0.05, ev21=0.2] ]else If any ps has

(automatedActionPerformed = ac2) If any tas has (runningTask

= false) [ev1=0.02, ev2=0.38, ev21=0.6] else [ev1=0.45,

ev2=0.05, ev21=0.5] ]else If any ps has

(automatedActionPerformed = ac3) If any tas has (runningTask

= false) [ev1=0, ev2=1, ev21=0] else [ev1=0, ev2=1, ev21=0]

]else [ev1=0.34, ev2=0.33 ev21=0.33]

The local distribution to the resident node

influence(ps, t) describes how Automatically

Executed Actions and Running Task influenced the

establishment (arise) of the Events “forgot stove

turned on (ev1)", "Stove turned off (v2)" and "

Sensor detected presence near to the stove (ev2.1) ".

Therefore, if: (a) John is warned (ac1) and the

task is not running, John forgets the stove turned in

5% of cases, do not forget 35% and there is presence

near to the stove 60%. If the task is running, John

forgets stove turned in 75%, does not forget 5% and

there is presence near to the stove 20%; (b)

Caregiver notified (ac2) and there is no running task,

John forgets the stove turned on 2% of time, do not

forget 38% and there is presence near to the stove

60%; If running task, he forgets in 45% of time, do

not forget to 5% and presence by 50%; (c) Stove

turned off automatically (ac3), in 100% of cases, the

stove turned off (v2) and, for a default distribution,

are distributed on average 33.33% for all events.

The resident node influence(ps, t) applies your

local distribution probability into

willBeSituationOf(ps, t, us) node at time t, so if John

forgot the stove (ev1), then 99% cases the dangerous

Reasoning on Uncertainty in Smart Environments

247

situation is valid and only 1% false. If John turned

off the stove (v2), then there is 99% chance of

dangerous situations does not exist and, if it was

detected presence near stove (ev2.1), then there is a

dangerous situation in 45% of time. Thus

demonstrating that even if someone neared the stove

(checking or not the cooking activity), there is a

probability of danger.

Using the structure of the reference model

defined in Section 3.1 and local probability

distributions given in Table 4, the system can

generates the Specific Situation Bayesian Network

for the scenario as following. Figure 5 presents the

SSBN not recursive (only T0 time). It was adding

the evidences that John is using the smart stove and

the dangerous situations will be situation of John at

T0 on 46,39% of chance.

Figure 5: SSBN at T0 and John using Stove.

In Figure 6, it was added the T1 time (ontology

instance). Therefore, it was possible to determinate

that John will be involved by Dangerous Situation at

T1 with 46,4%.

Figure 6: SSBN from T0 to T1 and John using Stove.

The difference between the last net and it was

0.01%, but if it is relevant, depends on the

interpretation of an expert on the detected situation.

In Figure 7 is presented that there is an evidence

that John, at T0, forgot the stove turned off (ev1), so

in T1 there is a probability of 72,5% of chance that a

Dangerous Situation will be the situation of John in

T1.

In Figure 8 it was used more evidences. Thus,

the network was used to answer the following query:

"What is the probability that John to be in a

dangerous situation in T2?". In the Figure, is

presented the SSBN with the evidences (axioms

when applied in SSBN result in gray nodes)

provided by the scenario. Therefore, the current time

is T1, John is involved by tasks watching tv and use

stove, these tasks comprise the cooking activity.

Figure 7: SSBN from T0 to T1, John using Stove and Ev1

detected at T0.

The local probability distribution of the Table 4

describe that John is involved by some task, it is a

positive influence to dangerous situation. There is

evidence that John at T0 (previous time) forgot the

stove on (ev1 = 100%) and the current time (T1)

already was detected presence of some people near

the stove (ev21 = 100%). With these evidences, the

system has the intention to know if in T2 John will

be involved by a Dangerous Situation. According to

figure 8, there is 69,8% of chance that John is in

Dangerous Situation in T2.

Figure 8: SSBN with the evidences.

5 CONCLUSIONS

Systems for intelligent environments may become

proactive when technologies support the reasoning

over uncertainty in runtime. Multi-Entity Bayesian

Network made these characteristics possible through

of the generation of Specific Situation Bayesian

ICEIS 2016 - 18th International Conference on Enterprise Information Systems

248

Network. This paper presented the Reference Model

for Predictive Situation-Awareness for give dynamic

support in relation to reasoning over uncertainty for

detection unwanted situation with Multi-Entity

Bayesian Network Theory. This model provides

essential support for the prediction of situations in

real environments. The reference model enables the

Bayesian network structures as well as the

probability values for predictions are generated at

runtime. Therefore, the SSBN are generates on the

moment that the user is living in their residence.

Ordinary Bayesian Network are not dynamic this

way, because need of an expert to model their

structures.

The contributions of this paper include the use

of Semantic Web Technologies for reasoning about

uncertainty, as well as the reference model for

predicting unwanted situations. Further work is the

identification of a top ontology to increase the

coverage context model and using the reference

model for different scenarios.

REFERENCES

Allocca, C, D'Aquin, M. 2009. Door: Towards a

formalization of ontology relations. In. International

Conference on Knowledge and Ontology

Development, 2009. Proceedings of the International

Conference on Knowleadge and Ontology

Development. Madera: [s.n.], p. 13-20.

Bettini, C., Brdiczkab, O., Henricksen, K., Iindulzkad, J.,

Nicklase, D., Ranganathanf, A., Rriboni, D. 2010. A

survey of Context Modelling and Reasoning

Techniques. Pervasive and Mobile Computing, 6(2), p.

161-180.

Blasco, R., Marco, Á., Casas, R., Cirujano, D., Picking, R.

2014. A smart kitchen for ambient assisted living.

Sensors, v. 14, n. 1, p. 1629-1653.

Carvalho, R.; Laskey, K.; Costa, P. C. 2013. PR-OWL

2.0–bridging the gap to OWL semantics.

In: Uncertainty Reasoning for the Semantic Web II.

Springer Berlin Heidelberg, p. 1-18.

Coronato, A. 2012. Uranus: A middleware architecture for

dependable AAL and vital signs monitoring

applications. Sensors, v. 12, n. 3, p. 3145-3161.

Coronato, A., De Pietro, G. 2013. Situation awareness in

applications of ambient assisted living for cognitive

impaired people. Mobile Networks and Applications,

v. 18, n. 3, p. 444-453.

Costa, P. C. G., Carvalho, R. N., Laskey, K. B., Park, C.

2011. Evaluating uncertainty representation and

reasoning in HLF systems. In: Information Fusion

(FUSION), 2011 Proceedings of the 14th

International Conference on. IEEE. p. 1-8.

Costa, P. C. 2005. Bayesian semantics for the Semantic

Web. George Mason University Departament of

Systems Engineering and Operations Research,

George Mason University: Fairfax, VA, USA. p. 312.

Dey A, Abowd G. 1999. The context toolkit: Aiding the

development of context-enabled applications. In: Proc.

of the SIGCHI conference on Human factors in

compu-ting systems, Pittsburgh, Pennsylvania, US, pp.

434–441.

Fenz, S., 2012. An ontology-based approach for

constructing Bayesian networks. Data & Knowledge

Engineering, v. 73, p. 73-88.

Forkan, A. R. M., Khalil, I., Tari, Z., Foufou, S., Bouras,

A., 2015. A context-aware approach for long-term

behavioural change detection and abnormality

prediction in ambient assisted living. Pattern

Recognition, v. 48, n. 3, p. 628-641.

Friedman, N., Geiger, D., Goldszmidt, M., 1997. Bayesian

network classifiers. Machine learning, v. 29, n. 2-3, p.

131-163.

Hobbs, J. R., Pan, F., 2004. An ontology of time for the

semantic web.ACM Transactions on Asian Language

Information Processing (TALIP), v. 3, n. 1, p. 66-85.

Howard, C; Stumptner, M., 2014. A Survey of Directed

Entity-Relation--Based First-Order Probabilistic

Languages. ACM Computing Surveys (CSUR), v. 47,

n. 1, p. 4.

Laskey, K., 2008. MEBN: A language for first-order

Bayesian knowledge bases. Artificial intelligence, v.

172, n. 2, p. 140-178.

Machado, A., Pernas, A. M., Augustin, I., Thom, L. H.,

Krug, L., Palazzo, J., Oliveira, M. De, 2013. Situation

awareness as a Key for Proactive Actions in Ambient

Assisted Living. In: Proc. of the 15th International

Conference on Enterprise Information, p. 418–426.

Machado, A., Lichtnow, D., Pernas, A. M., Wives, L. K.,

de Oliveira, J. P. M. 2014. A Reactive and Proactive

Approach for Ambient Intelligence. International

Conference on Enterprise Information System, p. 501-

512.

Paganelli, F., Giuli, D., 2011. An ontology-based system

for context-aware and configurable services to support

home-based continuous care. In. Journal of the IEEE

Transaction Information Technology Biomedical, v.

15, n. 2, p. 324-333.

Rasch, K., Li, F., Sehic, S., Ayani, R., Dustdar, S., 2011.

Context driven personalized service discovery in

pervasive environments. World Wide Web, v. 14, n. 4,

p. 295-319, springer, Netherlands.

Sixsmith, A., Meuller, S., Lull, F., Klein, M., Bierhoff, I.,

Delaney, S., Savage, R., 2009. SOPRANO: An

Ambient Assisted Living System for Supporting Older

People at Home. In: Lecture Notes in Computer

Science. Springer Berlin: Heidelberg, vol. 5597, p.

233–236.

Strang, T., Linnhoff-Popien, C., 2004. A Context

Modeling Survey. In: Workshop on Advanced Context

Modeling, Reasoning and Management, UbiComp

2004 - The Sixth International Conference on

Ubiquitous Computing, Nottingham, England.

Tazari, M. R., Furfari, F., Ramos, J. P. L., Ferro, E, 2010.

The PERSONA service platform for AAL spaces.

Reasoning on Uncertainty in Smart Environments

249

In: Handbook of Ambient Intelligence and Smart

Environments. Springer US. p. 1171-1199.

Yang, Y., Calmet, J., 2005. Ontobayes: An ontology-

driven uncertainty model. In: Computational

Intelligence for Modelling, Control and Automation,

2005 and International Conference on Intelligent

Agents, Web Technologies and Internet Commerce,

International Conference on. IEEE, p. 457-463.

Ye J, Stevenson G, Dobson S., 2011. A top-level ontology

for smart environments. In: Pervasive and Mobile

Computing. vol. 7, no. 3, p. 359-378.

ICEIS 2016 - 18th International Conference on Enterprise Information Systems

250