Interacting with intelligent systems in general and ambient intelligent systems in particular, requires that these systems have the ability to build a trust relationship with the users. The ability to explain its own behaviour is one of the most important abilities that such a system can exhibit to gain trust. We argue that explanations are not just an addition to an ambient intelligent system rather it is an approach to the design and implementation of such a system. Explanations are useful both for the reasoning process itself and as a means of communicating with the users. In this paper, we present a knowledge intensive approach for identifying different contexts and generating a course of action depending on the context found. We explore the use of explanations both as a means of reasoning and as a means of communication with the user.
Ambient intelligent systems are described by their ability to be aware of its users,
perceive their needs and respond intelligently to them [
Further, the nature of an ambient intelligent system dictates that the way we traditionally interact with computer systems are substituted with multi modal interfaces and in particular behavioural interfaces. It is not only the input devices where behavioural interfaces are used; the main output device for an ambient intelligent system is its behaviour. Thus, a user now runs the risk of having to master the methods of behavioural psychology to use an ambient intelligent system. Relying solely on observing the behaviour of an ambient intelligent system will, at best, give us a limited understanding of its behaviour. It will not allow us to consider any internal organisation. Yet, to persuade the user of a system’s usability and credibility, explanations are in order
The work presented here argues that explanations are not just an addition to an ambient intelligent system; rather it is an approach to the design and implementation of such a system. Explanations are a useful approach to both the reasoning process from a system centric perspective, as well as a means of communication viewed from a user centric perspective.
This paper is structured as follows: First, an overview on how explanations and case-based reasoning are related is given. Secondly, an introduction to the use of case-based reasoning in ambient intelligence, as well as a short overview of the system employed here, is presented. Thirdly, a description of the theoretical framework underpinning our approach to the use of explanations and case-based reasoning in ambient intelligence is given. A summary and pointers to future work ends the paper. 2
Originally, case-based reasoning did emerge from an understanding of
reasoning as an explanation process [
Sørmo et al. [
Roth-Berghofer has explored some fundamental issues with different useful
kinds of explanation and their connection to the different knowledge containers of
a case-based reasoning system [
Building on the last two works, we have earlier started to investigate a
combined framework of user goals and explanation kinds [
Weiser, who coined the term ubiquitous computing, did explicitly state that artificial intelligence was unimportant when realising the visions described [11, p.3]. This view that proper embedding of computers is sufficient does also to a large degree saturate the field of pervasive computing. However, the insight that achievement of the visions described in pervasive computing does indeed require some degree of reasoning has lead to the developments jointly labelled as ambient intelligence.
The IST Advisory Group to the European Commission (ISTAG) describes
ambient intelligence as human beings surrounded by intelligent interfaces,
supported by computing and network technology embedded in everyday objects.
More importantly, the environment should be aware of the presence of a person,
perceive the needs of this person and respond intelligently to them [
The ability to be aware of the environment, reason about ongoing situations
and decide about appropriate behaviour is closely linked with being
knowledgeable about the world. Case-based reasoning in general, and knowledge intensive
case-based reasoning in particular [
Zimmermann reports on case-based reasoning used to generate
recommendations based on the user’s context in a mobile environment [
Along the same lines of adapting solutions to users is the work by Ma et al., where case-based reasoning is used to adapt the behaviour of smart homes to user preferences [15]. The cases are represented as frames, where the findings are: the user, the environment, the time, and the values of the active devices. When new cases are instantiated; the similarities between the new and existing cases are calculated and a setting for the appliances in the house is selected.
Recently B´enard et al. have investigated the use of case-based reasoning as a mechanism for selecting suitable behaviour in different situations [16]. They propose an agent-based architecture that uses perceived information, or context, as the findings of a case and the proposed action as the solution. The existing cases in the case base are pre-classified situations modelled by a domain expert. The authors briefly describe how the case-based reasoning process approaches Perceived
Goal
Directed World Knowledge
Awareness
Sensitivity Situations Context
Previous Cases
New Case
Retrieved Case
Goal Goal Context
Decomposer Agent
Application Agent Application Agent Application Agent Context Middleware
CREEK Agent
Agents the problem of behaviour selection. Initially a new context is compared to the findings in the existing cases and the best matching case is retrieved. Secondly, the retrieved case is adapted to fit the new context.
The work presented here builds on the experience gained by applying casebased reasoning to situation assessment in an ambient intelligent system [17]. The architecture developed and system implemented approaches ambient intelligence by separating the main responsibilities into three layers.
The architecture has been implemented as an ambient intelligent system in a hospital ward [18]. The personnel at the hospital ward are involved in many different activities, such as doing ward rounds, meetings and different forms of examinations. The system’s main purpose is to recognise ongoing situations and proactively acquired digital information relevant for the user.
Figure 1 depicts the functional system architecture. The Context
Middleware [19] provides an infrastructure that perceives the environment by
collecting and maintaining a coherent model of available context. The CREEK agent
implements and extends the knowledge intensive case-based reasoning method
CREEK [
The knowledge model utilised in this system is structured as a multi-relational semantic network, where each of the five different parts are integrated. Three parts of the knowledge model are universal for any CREEK application: the top-level ontology, called ISOPOD; the domain-specific model of general and factual knowledge; and the case base. In addition to these, two parts have been developed for the ambient intelligent system, namely the Basic Context Model and the Activity Theory Model.
The Basic Context Model is structured around a meronomy based on tradition in pervasive computing [19,20], which imposes a structure that facilitates easy development of context sensitive applications [21]. The Activity Theory Model captures knowledge regarding activities, which is one of the most important aspect of situations [22]. 4
As outlined in Section 2, the term explanation can have different focii in artificial
intelligence. In his introduction to the ECCBR-04 Workshop on Explanation in
case-based reasoning, David Leake [23] has identified three different facets of
explanation and case-based reasoning:
– Using explanations to support the case-based reasoning process
– Generating explanations by case-based reasoning
– Using cases for explaining system results to an external user
With our notion of user goals, we can subsume the last two facets as both being
targeted towards the user of the system. In our understanding, showing the case
to the user is a special case of ‘generating explanations by case-based reasoning’,
making use of the case-based reasoning assumption that similar problems have
similar solutions. Provided that the user has some knowledge about the
similarity function and that the case structure is understandable by the user, the
displayed case acts as an explanation to the user (see e.g. [
Example: a knowledge intensive case-based reasoning system can use its domain knowledge to explain the absence/variation of feature values. – Giving explanations of the found solution, its application, or the reasoning process to the user.
Example: in an engine failure diagnosis system, the user gets an explanation on why a particular case was matched.
When we now look at the use of context, we can again identify different views on context during the reasoning process and the use of the system. First, the system has to identify an unknown situation out of its internal state and the percieved state of the world. Here the system has to find out which situation it is in. Second, the system has to act according to the percieved state of the world and an assumed context. Following our earlier argumentation introduced in [18], we identify these two aspects as two distinct steps in the reasoning process: – Context Awareness: Trying to figure out which situation the system is in.
Example: An ambient intelligent system for supporting health personnel recognises that the user is on a ward-round because of the time of the day, the location, and the other persons present. – Context Sensitivity: Acting according to the situation the system thinks it is in.
Example: the same system fetches the newest versions of electronic patient records of all patients in the room from the hospital systems. When the user stands close to the bed of a patient, the system automatically displays them. Combing these views on explanation and on context, we end up with two dimensions of inquiry as depicted in Table 1. 5
We will further investigate the relationship between context and explanations by examining an example from our current implementation. At this time, our system is capable of assessing ongoing situations in a simulated hospital (cardiology) ward domain. The system’s main purpose is to identify ongoing situations and proactively acquire digital information required by the persons present. Recognise: In this step, we are using explanations to recognise the current ongoing situation. The system uses all available resources in its reasoning process. Let us assume that the ongoing situation is a ward round. Normally ward rounds take place in a patient room, however the current situation is occurring in the hallway. This discrepancy can be explained away by the system generalising that both locations can indeed contain patient beds. When CREEK retrieves a matching case, the system has no explicit knowledge stating that a hallway can contain hospital beds. The initial match is of a syntactical nature only. However, it can use its general knowledge and the reasoning mechanism of plausible inheritance to generate an explanation supporting the hypothesis that beds can be located in the hallway, for example because they are both some kinds of rooms, and beds are some kinds of objects which have a room as a location. Therefore, as all other parameters are consistent with a ward round, the system assumes that it is indeed a ward round situation.
The explanation used by the system in this example states that a hallway is a room and can therefore contain a hospital bed.
Elucidate: We now want to generate an explanation for the user that tells the user why the system assumes a certain situation. The system will make use of all available sources of knowledge in order to gain the user’s confidence in its capabilities. It will also have to consider the user’s goals when chosing a specific explanation. It has been shown that simply presenting the reasoning trace is not always sufficient (it can even be counter productive) [26,27]. The system might therefore generate an after-the-fact explanation, which for example justifies its assumption instead. Since the ward round situation is occurring at an unusual place, CREEK will point out the time of the day, the availability of the other expected participants, and the fact that hallways might contain beds, as the reason for its assumption instead of only displaying its generalisation of the location.
The explanation shown to the user is a justification of the system’s believe of being on a ward round.
Identify: After the system has successfully identified the context, it is using explanations to generate a plan for a reasonable course of action. Now, it is using only the knowledge sources important for the situation at hand (the context is acting as a focus lense [18]). When we now presume that the system has recognised that we are on a ward round, discussing medical conditions and treatments with several patients, it will try to prepare all the relevant information to be presented to the user. This includes all test results. The system can now ask other available artefacts for test results on the user, and the medical images database can offer a MR image whereas the patient record offers a textual description of the MRI. Because of limitations of handheld devices, the system will for example not be able to display high resolution MR images. When choosing which of the artefacts to query, the system will reject the medical image database and only query the electronic patient record database.
The explanation used by the system is based on the knowledge that a high resolution image displaying device is not available on a ward round. Explicate: Looking at the user centric part again, we are now in need of generating an explanation for why the systems takes a specific action. The system will take into account which situation it assumes it is in and the possible goals the user might have for an explanation. In executing its plan, the system proposes its user to visit the isolation room with patients which should be kept seperate. The user is surprised since he is not aware that any of the patient he should see at the ward round is in the isolation room, and no information on this was exchanged in the morning briefing. The system can then generate an explanation that shows the relevance of the proposal by pointing out that one particular patient had to be moved to the isolation room for medical reasons since the time of the morning meeting, and this information was available via the patient information system. This explanation would not be useful if it had not been established already that we are on a ward round and the aim was to visit the patients. Vice versa, if the system would generate a justification for its assumption of being on a ward round, this would still not satisfy the need of the user to know why he should go to the isolation room.
The explanation shown to the user is pointing out the relevance of performing a particular action, namely visiting the isolation room. 6
This paper has presented an approach to combine explanations and case-based reasoning in context awareness. It has been argued that explanations can be viewed from both a system centric and user centric perspective. It has further been described how explanations play a key role in ambient intelligent systems, where traditional interaction has been changed, and division of responsibility has been shifted from user initiated to mixed initiative.
The current implementation does support some of the goals seen from an user centric perspective; that is elucidating why it assumes that a given situation is occurring. The transparency goal is satisfied by CREEK’s ability to graphically represent the case-match and the underlying semantic network representation, as well as a textual representation of the overall match strength and individual feature matches. A textual representation of explanations used in the reasoning process is also shown to the user.
The conceptualization goal is supported by providing a means to the user to explore the relation of the case to the underlying knowledge base. By examining the relations of the case features to this underlying knowledge structure, the user can gain some insight into the conceptual model the system has of the domain at hand.
The justification goal is to some degree supported in our current implementation. CREEK does explain why two different features are similar by means of plausible inheritance [28].
The most important future development is to apply the facets of explanation to the context sensitive part of our system. Here the Decomposer and Application agents must be able to use a system centric perspective on identifying a suitable behaviour in a given situation; as well as explicating why a given behaviour was chosen to the user.
Another direction for future research we want to explore is to tie our findings back in to our earlier work on a design methodology for explanation-aware intelligent systems with a socio-technical perspective. Our research has shown that ambient intelligent systems can benefit from a psychologically plausible knowledge model, but we have not yet explored the relation to the different knowledge containers in detail.
Additionally, we want to augment existing design guidelines with methods for the analysis of social aspects that can lead to a better understanding than ad-hoc methods. It is also important to note that we have not yet fully utilised some aspects of our theoretical foundations in organisational psychology, like the notions of breakdown situations or functional organs in Activity Theory, or the use of Semiotics for the organisation of the user interface itself. 15. Ma, T., Kim, Y.D., Ma, Q., Tang, M., Zhou, W.: Context-aware implementation based on cbr for smart home. In: Wireless And Mobile Computing, Networking And Communications, 2005. (WiMob 2005), IEEE, IEEE Computer Society (2005) 112–115 16. B´enard, R., Bossard, C., Loor, P.D.: Context’s modelling for participative simulation. In: Proceedings of the 19th FLAIRS Conference, AAAI Press (2006) 613–618 17. Kofod-Petersen, A., Aamodt, A.: Case-based situation assessment in a mobile context-aware system. In Kru¨ger, A., Malaka, R., eds.: Artificial Intelligence in Mobile Systems (AIMS 2003), Universit¨at des Saarlandes (2003) 41–49 18. Kofod-Petersen, A., Aamodt, A.: Contextualised ambient intelligence through casebased reasoning. In Roth-Berghofer, T.R., G¨oker, M.H., Gu¨venir, H.A., eds.: Proceedings of the Eighth European Conference on Case-Based Reasoning (ECCBR 2006). Volume 4106 of LNCS (LNAI)., O¨ lu¨deniz, Turkey, Springer (2006) 211–225 19. Kofod-Petersen, A., Mikalsen, M.: Context: Representation and reasoning – representing and reasoning about context in a mobile environment. Revue d’Intelligence Artificielle 19 (2005) 479–498 20. G¨oker, A., Myrhaug, H.I.: User context and personalisation. In: Workshop proceedings for the 6th European Conference on Case Based Reasoning. (2002) 21. Kofod-Petersen, A., Mikalsen, M.: An architecture supporting implementation of context-aware services. In Flor´een, P., Lind´en, G., Niklander, T., Raatikainen, K., eds.: Workshop on Context Awareness for Proactive Systems (CAPS 2005), Helsinki, Finland, HIIT Publications (2005) 31–42 22. Cassens, J., Kofod-Petersen, A.: Using activity theory to model context awareness: a qualitative case study. In: Proceedings of the 19th International Florida Artificial Intelligence Research Society Conference, Florida, USA, AAAI Press (2006) 619– 624 23. Leake, D.B.: Workshop introduction: ECCBR-04 workshop on explanation in CBR. Slides, https://www.cs.tcd.ie/research groups/mlg/ecbrws2004/, last access 22.02.2007 (2004) 24. Cunningham, P., Doyle, D., Loughrey, J.: An Evaluation of the Usefulness of
Case-Based Reasoning Explanation. [29] 122–130 25. Aamodt, A.: A knowledge-intensive, integrated approach to problem solving and sustained learning. PhD thesis, University of Trondheim, Norwegian Institute of Technology, Department of Computer Science (1991) University Microfilms PUB 92-08460. 26. Majchrzak, A., Gasser, L.: On using artificial intelligence to integrate the design of organizational and process change in US manufacturing. AI and Society 5 (1991) 321–338 27. Gregor, S., Benbasat, I.: Explanations from intelligent systems: Theoretical foundations and implications for practice. MIS Quarterly 23 (1999) 497–530 28. Sørmo, F.: Plausible inheritance semantic network inference for case-based reasoning. M.sc. thesis, Norwegian University of Science and Technology (NTNU), Department of Computer and Information Science, Trondheim (2000) 29. Ashley, K.D., Bridge, D.G., eds.: Case-Based Reasoning Research and Development: Proceedings ICCBR 2003. Number 2689 in LNCS (LNAI), Trondheim, Norway, Springer (2003)