Using Argumentation to Understand Ambiguous Situations in

1 downloads 0 Views 146KB Size Report
ments where the pads are located (beds and sofas in this scenario). Finally, a set of in- .... Other inactivity times for rooms, sofa and bed are also managed here.
Using Argumentation to Understand Ambiguous Situations in Intelligent Environments a

Andrés MUÑOZ a,1 , Juan A. BOTÍA a and Juan Carlos AUGUSTO b Dpto. de Información para la Ingeniería y las Comunicaciones, University of Murcia, 30100 Murcia, Spain. {amunoz, juanbot}@um.es b School of Computing and Mathematics, University of Ulster, BT37 0QB Newtownabbey, United Kingdom [email protected] AbstractIntelligent Environments provide the infrastructure to deliver Ambient Assisted Living such that elderly people can remain at home and independent for longer. An important part of an Ambient Assisted Living provision is the monitoring, detection and follow up of unsafe situations. One problem these systems have to perform that task is that they only observe the world through incomplete, sometimes unreliable, sensing devices. This naturally leads to uncertain, inconsistent and incomplete perceptions of the world. This paper shows how argumentation, a well established field in computer science, can help to disambiguate some of those confusing (sometimes even inconsistent) situations. Keywords. Intelligent Environments, Ambient Assisted Living, Argumentation, safety monitoring

1. Introduction The work proposed in this paper is concerned with elderly people who live alone. We pursue the design of a system centered on Intelligent Environments [4], which is able to unobtrusively detect undesired events the elderly (the occupant of the house henceforth) might suffer. Such events must be detected (e.g. examples are falls or strokes) in order to trigger an emergency process which may help in improving the quality of the service offered to the occupant and eventually saving lives. The basic information system used here is based on a sensor network deployed through the occupant’s house. Such sensors include pressure, activity and open door sensors. Lectures from such sensors are gathered into a small PC in which the logic in charge of detecting anomalies related to the occupant’s state. Such kind of set-up is a suitable one because is unobtrusive (i.e. the occupant may even forget that she is being monitored, as sensors are disguised with the environment). However, using very simple sensors implies lack of information to draw some complex contexts (e.g. if the occupant does not attend a door bell ring, may it indicate she is suffering some problems or she is simply having a nap?). This paper proposes an 1 Corresponding

Author: [email protected]

argumentation process [3] to disambiguate such confusing situations. It starts modeling the environment, the occupant’s daily patterns in the house and her context by using Semantic Web ontologies [1]. Then, rules to classify each different context are defined. In case that conflicting contexts are entailed by these rules when new events are generated from the environment, an argumentation process is launched to decide which the actual context is. Until now, approaches to tackle with such uncertain situations included fuzzy logic based reasoning [7], Markov-based models [6], and so on. All these techniques rely on the application designer to specify the numeric parameters of the decision models (i.e. parametrical specification) and their structure. Argumentation frees from generating the parametrical specification. Only the structural is needed in this case. The rest of the paper is structured as follows: Section 2 describes the running scenario that will be used throughout the paper in order to illustrate the Ambient Assisted Living (AAL) system proposed here. Section 3 explains the information models for representing events and situations. Section 4 shows the AAL system architecture and the argumentation process, which functionality is demonstrated by means of a prototype. Finally, section 5 provides some reflections about this work.

2. A SmartHome scenario The scenario developed in this paper describes a smart home equipped with sensors and monitored by an AAL system. In this scenario, the occupant is located in the living room sofa when the door bell rings. The occupant does not open the door in the expected time, which may be seen as an undesirable situation. However, this person may be having a nap or simply she does not want to open the door because she is relaxing on the sofa. As a result, the AAL system may take the decision of sending an alarm to the carer because of the unattended door bell ring or it may detect that the occupant does not want to be disturbed, being this fact more relevant than the unattended ring. This scenario is taken as reference to illustrate how the use of Argumentation in the AAL system could help to disambiguate and give a suitable explanation for this kind of problems. Firstly, some informal definitions are needed. Thus, we will use the word “situation” to refer to the occupant’s safety state at any instant of time, according to the sensor data and occupant’s profile information in such an instant. Moreover, a situation will be “safe” or “unsafe” when the sensor data and occupant’s information fulfill the set of conditions gathering the requirements of the carer and/or occupant herself for defining it as such. Now, the following assumptions are made about the sensors in the house. A general supposition is based on the existence of a middleware which is responsible for capturing the sensor signals and transforming them into events to be managed by the Assisted Living system. This middleware will be explained in section 4.1. Moreover, we assume that the assessment information provided by the AAL system is relevant when the occupant is the only person detected at home. Therefore, if the occupant leaves the house or more than one person is known to be at home, the system is either switched off or its notifications are ignored until the occupant is detected to be the only person at home again. These cases are handled by sensors in the main door not explained here. The rest of assumptions are specific for the sensors which are employed in this paper: Movement detectors. There is one of this type of sensors in each room, tracking for activity in it. When activity is detected, a signal is sent to the middleware, and then it is

transformed into an event from the specific sensor with the value “on”. If no activity is detected, the generated event contains the value “off”. Due to this data, the system could determine the occupant’s location and the amount of inactive time in each room. Pressure pads. They are installed in home elements where the occupant sits or lies down, such as beds or sofas. While the pressure on the pad does not surpass a particular threshold, it sends a signal which is converted by the middleware into an event from this sensor with the value“off”. Instead, when the pressure on the pad exceeds this threshold, an event is generated which contains the value“on”. Door and bell sensors. They are aimed to detect a door bell ring which has not been attended by the occupant. Therefore, if the door bell has been pressed and the door is not open within an expected amount of time, for example 30 seconds, the middleware converts the door bell signal in an inactivity event for that door. The next section defines a model for representing all these sensor events, together with some other inactivity events. These are introduced in the system as a result of exceeding the normal inactivity time that has been calculated for a room or element.

3. Representing events and situations with OWL ontologies The events described in the previous section are represented by means of an ontology. Hence, these events are formally represented through a set of concepts and relationships between them. As section 4.2 will show later, it is possible to define safe/unsafe situations by reasoning about the instances of these concepts and relationships. There are several languages for describing ontologies in a computational system, such as OWL (Web Ontology Language) [5]. In this paper, the ontologies for representing events and situations are written in OWL. Since ontologies are used here in a high level of abstraction, no specific knowledge is needed to understand their functionality in the system. Figure 1a depicts the events model in an OWL ontology. There is one concept for each movement detector, according to the room where it is placed. As explained in section 2, there are two types of events for each movement detector. Thus, a Mov_R_On concept represents activity in the room R, while Mov_R_Off is the concept defined for the event generated when no activity is detected there. Analogously, the pressure pads follow the same concept hierarchy. Instead of rooms, the hierarchy is based on the elements where the pads are located (beds and sofas in this scenario). Finally, a set of inactivity events are also defined in the ontology. They represent inactivity in the different rooms or elements that should be watched for detecting unsafe situations, such as lying unconscious on the bed or sofa. These events are introduced in the system by the middleware, when it detects that the normal inactivity time for each type of room or element has been surpassed (see section 4.1). The Inactivity_DoorBell concept indicates that the door ring was pressed and the door has not been opened. Figure 1b shows the OWL ontology for modeling situations in the system. The Situation concept is specialized in Assessment, Location and Context subtypes. The first subtype indicates a situation as safe/unsafe, whereas the second one adds information about the occupant’s location. Eventually, the context subtypes represent specific situations according to the combination of location information, events and/or the occupant’s profile information. For example, the Relax_Ctx concept denotes a

(b) (a) Figure 1. The Events (a) and Situations (b) model structures in an OWL representation (partial)

context where the occupant is detected in the living room, the sofa pressure pad is “on” and there is a pattern which says that now the occupant should be watching TV or having a nap, by considering his profile. On the other hand, the Unatt_DoorBell_Ctx represents a context where the inactivity door bell event has occurred. Due to this formal representation, a reasoning process about the location and context information can be executed to classify a situation as safe or unsafe. Obviously, a situation could have only one assessment and location subtypes at the same time (e.g. Unsafe and BathRoom_Loc). If the same situation is classified as safe and unsafe, or more than one location has been detected in it, an inconsistency arises in the system. This inconsistency is automatically detected thanks to the formal definition given by the OWL language, which allows to define the unsafe/safe concepts as disjoints. Similarly, the location concepts are disjoint among them. Section 4 explains how to deal with these possible conflicts.

4. An advanced architecture for reasoning about events in a SmartHome Figure 2 depicts the architecture for the Assisted Living system proposed in this paper, which is composed of three elements. Thus, the OWL Ontology Repository contains the events and situation models given in section 3, together with an ontology for representing the occupant’s daily patterns such as sleeping or relaxing times (i.e. when the occupant does not want to be disturbed). On the other hand, the Event Management System is a middleware for managing the sensor data and transforming it into events (see section 4.1). Finally, the Argumentation-based Reasoning System continuously evaluates the situation generated from the gathered events as safe or unsafe. To this end, an inference process is performed taking into account the possible conflicts that may arise in it. These conflicts are then solved by argumentation techniques (see section 4.2). As shown in figure 2, the sensor data is collected by the Event Management System. It generates a situation (possibly with location information) that is sent to the Argumentation-based Reasoning System. Here, an assessment report of that situation is finally produced as the output of the system. This report indicates if the situation has been entailed as safe or unsafe, along with an explanation for this decision. The reports can then be monitored by the Assisted Living system to decide when it is convenient

Occupant’s Profile Ontology Events Model Ontology Situation Model Ontology

OWL Ontology Repository

SSA agent Movement detectors

Location Inference Module Events KB

SSA-rules

Events + Location

Pressure pads

Situation Report (Safe /Unsafe)

Inactivity Inference Module

Event Management System

USA-rules

USA Agent

Argumentation-based Reasoning System

Door and Bell sensors

Assisted Living System

Figure 2. Assisted Living system architecture.

to send an alert to the correspondent carer (e.g. when a situation is evaluated as unsafe for more than a specific amount of time, or the explanations for the unsafe results are classified as critical, such as falls). 4.1. Event Management System The Event Management System (EMS) is a middleware for managing all the events in the SmartHome. It is responsible for transforming the sensor data into events according to the OWL Events ontology and adding these events to the Events KB. Moreover, the EMS contains a Location Inference module which entails the occupant’s location in the house according to the movement detectors events. For this purpose, the module uses a set of location rules and a rule-based inference engine. For example, the following rule infers the occupant in the hall: R-Hall-Loc : 𝑆𝑖𝑡𝑢𝑎𝑡𝑖𝑜𝑛(?𝑠) ∧ 𝑚𝑜𝑣_ℎ𝑎𝑙𝑙(?𝑠, ?𝑚𝑜𝑣) ∧ 𝑀 𝑜𝑣_𝐻𝑎𝑙𝑙_𝑂𝑛(?𝑚𝑜𝑣) ⇒ 𝐻𝑎𝑙𝑙_𝐿𝑜𝑐(?𝑠)

The rule checks if there is a situation with an “on” movement event in the hall. If this condition holds in the Events KB, the rule establishes the occupant’s location in the hall. There is one analogous rule to R-Hall-Loc for each movement detector. Regarding the rule-based inference engine, the module uses the general rule engine provided by the Jena framework (http://jena.sourceforge.net). Note that the Location Inference module may infer several locations if two or more “on” movement events are detected at the same time. As we have assumed that the AAL system is only working when the occupant is known to be alone in the house, this transitory conflict is due to a sporadic false positive in any detector or the occupant is moving from room A to B and the “on” movement event in room A arrives before that the “off” movement event in room B. In these cases, the module can not determine the location information. However, the conflicts disappear as movement events become stable again, and the location information is restored again. Apart from inferring the occupant’s location, the EMS also generates the inactivity events shown in figure 1 through the Inactivity Inference module. For example, it generates the unattended door bell ring event when the calculated time for opening the door

has been surpassed. Other inactivity times for rooms, sofa and bed are also managed here. There exist several approaches for calculating these times. However, this subject is beyond the scope of this paper and it will not be studied here. The interested reader is referred to [2] for detailed aspects of these approaches. 4.2. Argumentation-based Reasoning System The Argumentation-based Reasoning System (AbRS) is composed of two software agents. Each agent is responsible for a group of rules which evaluates the situation created from a set of events and the occupant’s profile. Hence, the Safe-Situation Assessment (SSA) agent maintains the rules for evaluating a situation as safe, whereas the Unsafe-Situation Assessment (USA) agent defines the rules for detecting unsafe situations. Advantages of this separation of rules reside in a more clear and flexible definition of the conditions for each type of situation, thus avoiding complex rules that are difficult to validate and allowing a more complete range of situations. However, due to this flexibility, it is highly improbable to have mutually exclusive sets of rules in the SSA and USA agents (i.e. different contexts will be fulfilled for entailing the same situation as safe and unsafe). As a result, both agents need to reach an agreement about the situation’s safety level. To this end, an argumentation process is carried out between them. Let us see an example of argumentation by following the scenario described in section 2. Suppose that the USA agent classifies the unattended door bell ring context as unsafe, by means of the rule R-Unsafe. Moreover, this agent uses the rule R-UnattDB to entail such context, which checks the current situation for a door bell inactivity event: R-UnattDB : 𝑆𝑖𝑡𝑢𝑎𝑡𝑖𝑜𝑛(?𝑠) ∧ 𝑖𝑛𝑎𝑐𝑡𝑖𝑣𝑒(?𝑠, ?𝑒𝑣𝐷𝐵) ∧ ∧ 𝐼𝑛𝑎𝑐𝑡𝑖𝑣𝑖𝑡𝑦_𝐷𝑜𝑜𝑟𝐵𝑒𝑙𝑙(?𝑒𝑣𝐷𝐵) ⇒ 𝑈 𝑛𝑎𝑡𝑡_𝐷𝑜𝑜𝑟𝐵𝑒𝑙𝑙_𝐶𝑡𝑥(?𝑠) R-Unsafe : 𝑈 𝑛𝑎𝑡𝑡_𝐷𝑜𝑜𝑟𝐵𝑒𝑙𝑙_𝐶𝑡𝑥(?𝑠) ⇒ 𝑈 𝑛𝑠𝑎𝑓 𝑒(?𝑠)

On the other hand, the SSA agent classifies a relax context as safe, by means of the rule R-Safe. In this case, the agent uses the rule R-RelaxCtx to entail the relax context, according to the scenario description. Thus, this rule checks that the occupant is located in the living room, the sofa sensor is activated, and the situation is within the relax time defined in the occupant’s profile ontology. This last condition is evaluated by the function isInRelaxTime(). R-RelaxCtx : 𝐿𝑖𝑣𝑖𝑛𝑔_𝐿𝑜𝑐(?𝑠) ∧ 𝑠𝑜𝑓 𝑎_𝑝𝑟𝑒𝑠𝑠𝑢𝑟𝑒(?𝑠, ?𝑒𝑣𝑆𝑜𝑓 𝑎) ∧ ∧ 𝑆𝑜𝑓 𝑎𝑃 𝑟𝑒𝑠𝑠𝑢𝑟𝑒_𝑂𝑛(?𝑒𝑣𝑆𝑜𝑓 𝑎) ∧ 𝑖𝑠𝐼𝑛𝑅𝑒𝑙𝑎𝑥𝑇 𝑖𝑚𝑒(?𝑠) ⇒ 𝑅𝑒𝑙𝑎𝑥_𝐶𝑡𝑥(?𝑠) R-Safe : 𝑅𝑒𝑙𝑎𝑥_𝐶𝑡𝑥(?𝑠) ⇒ 𝑆𝑎𝑓 𝑒(?𝑠)

Now, let us denote the situation in section 2 as 𝑆𝑖𝑡. Suppose that 𝑆𝑖𝑡 has been classified as 𝐿𝑖𝑣𝑖𝑛𝑔_𝐿𝑜𝑐 by the EMS since the movement sensor detected the occupant there. Suppose also that 𝑆𝑖𝑡 contains the events 𝑒𝑖𝑑3 and 𝑒𝑖𝑑4, representing the sofa pressure “on” and the unattended door bell events, respectively. Finally, the occupant’s profile indicates that the current time in 𝑆𝑖𝑡 is within her relaxing pattern interval. Then, this information is passed to ORE-AS, a tool for simulating argumentation between agents which has been developed in the University of Murcia. This tool implements conflict detection among the knowledge generated by the agents and an argumentation dialog directed to solve these conflicts. Here we describe ORE-AS as the AbRS implementation, while the theoretical models that underlie the conflict detection and argumentation

process are explained in [8, 9]. Thus, ORE-AS is launched with the situation 𝑆𝑖𝑡 and USA and SSA agents. As a result, the tool detects a conflict since 𝑆𝑖𝑡 is classified as safe and unsafe at the same time. Figure 3 shows the arguments generated by each agent for these entailments. While the USA agent uses event 𝑒𝑖𝑑4 to infer the unattended door bell context (Arg_U2), and in turn the final unsafe situation (Arg_U1), SSA agent entails 𝑆𝑖𝑡 as a relax context by the rule R-RelaxCtx (Arg_S2), and finally as safe (Arg_S1).

Figure 3. Arguments generated by the USA agent (upper part of the figure) and by the SSA agent (lower part).

Next, USA agent tries to persuade SSA agent by claiming the assertion (𝑆𝑖𝑡, 𝑡𝑦𝑝𝑒, 𝑈 𝑛𝑠𝑎𝑓 𝑒). To this end, arguments are exchanged by means of a dialog between the agents (see figure 4). Note that arguments Arg_U1 and Arg_S1 are rebutting, i.e. their conclusions are conflictive. In order to decide which argument is more relevant, ORE-AS is given a preference criteria about the premises in the arguments (i.e. the assertions supporting the conclusion). In this scenario, the preference criteria are defined by semantically annotating the Context concepts in the OWL Situation ontology. Hence, it is decided that the relax context is preferred to the unattended door bell context when assessing the situation’s safety. As a result, Arg_S1 defeats (i.e. is more relevant than) Arg_U1. Finally, the USA agent checks if Arg_S2 can be defeated in order to reject the relax context. Since Arg_S2 is based on events and information that is also shared and accepted by the USA agent, it finally accepts both arguments Arg_S1 and Arg_S2, retracting its initial claim. Consequently, 𝑆𝑖𝑡 is eventually classified as safe. 5. Conclusions Intelligent Environments are strongly dependent on their sensing capabilities. These systems have such potential that their application are being researched intensively. One such application is the implementation of systems which can provide Ambient Assisted Living. One fundamental service expected from such a system is that they can increase safety for the occupant of a house. However, given the intrinsic ambiguity, incompleteness and inconsistency of the provided input data, the implementation of this service a major challenge. This paper explain a safety scenario and the problem of implementing a sensible decision making mechanisms based on typical sensor data. We show how an argumentation-based architecture can offer a valid alternative to provide sensible qualitative reasoning in the presence of inconsistent and incomplete sensor data, as a first step to the development of more flexible and sensible Intelligent Environments.

Figure 4. Argumentation dialog between the USA and SSA agents. The initial proposal (𝑆𝑖𝑡, 𝑡𝑦𝑝𝑒, 𝑈 𝑛𝑠𝑎𝑓 𝑒) claimed by the USA agent is labeled as out, therefore it is not accepted in the system.

Acknowledgements The authors would like to thank the Spanish Ministerio de Ciencia e Innovación for sponsoring this work with the FPU grant AP2006-4154, the Fundación Séneca for the grant “Programa de Ayuda a los grupos de excelencia 04552/GERM/06 ”, and the Project CARDINEA, TSI-020302-2009-43.

References [1] [2]

[3] [4] [5] [6]

[7]

[8]

[9]

T. Berners-Lee, J. Hendler, and O. Lassila. The Semantic Web. Scientific American, 284(5):34–43, 2001. J. A. Botía, A. Villa, J. T. Palma, D. Pérez, and E. Iborra. Detecting domestic problems of elderly people: simple and unobstrusive sensors to generate the context of the attended. In First International Workshop on Ambient Assisted Living, IWAAL, (to appear), 2009. D. V. Carbogim, D. Robertson, and J. Lee. Argument-based applications to knowledge engineering. Knowledge Engineering Review, 15(2):119–149, 2000. D. J. Cook, J. C. Augusto, and V. R. Jakkula. Ambient Intelligence: applications in society and opportunities for AI. Pervasive and Mobile Computing, 5:277–298, 2009. M. Dean, D. Connoll, F. van Harmelen, J. Hendler, I. Horrocks, D. L. McGuinness, P. F. Patel-Schneider, and L. A. Stein. Web ontology language (OWL). Technical report, W3C, February 2004. B. Kröse, T. van Kasteren, C. Gibson, and T. van den Dool. CARE: Context Awareness in Residences for Elderly. In ISG 2008 - The 6th International Conference of the International, pages 101–105, Pisa, Italy, 2008. Intelligent Systems Lab Amsterdam. G. LeBellego, N. Noury, G. Virone, M. Mousseau, and J. Demongeot. A model for the measurement of patient activity in a hospital suite. IEEE Transactions on Information Technology in Biomedicine, 10(1):92–99, Jan. 2006. A. Muñoz and J. A. Botía. ASBO: Argumentation system based on ontologies. In M. Klusch, M. Pechoucek, and A. Polleres, editors, Cooperative Information Agents XII, volume 5180 of Lecture Notes in Artificial Intelligence, pages 191–205. Springer, 2008. A. Muñoz, A. Sánchez, and J. A. Botía. A software architecture for an argumentation-oriented multiagent system. In Y. Demazeau, J. Pavón, J. Corchado, and J. Bajo, editors, 7th International Conference on Practical Applications of Agents and Multi-Agent Systems (PAAMS’09), volume 55 of Advances in Intelligent and Soft Computing, pages 197–207. Springer, 2009.

Suggest Documents