Scholarly article on topic 'COALAA-GEN: A General Adaptive Approach for Ambient Assistive Applications'

COALAA-GEN: A General Adaptive Approach for Ambient Assistive Applications Academic research paper on "Computer and information sciences"

CC BY-NC-ND
0
0
Share paper
Academic journal
Procedia Computer Science
OECD Field of science
Keywords
{adaptativity / multi-dimensionality / cooperation / "assistive applications" / "multi-agent systems"}

Abstract of research paper on Computer and information sciences, author of scientific article — Nadia Abchiche-mimouni, Antonio Andriatrimoson, Etienne Colle, Simon Galerne

Abstract This paper presents the design and the implementation of general and adaptive framework for ambient assisted living applications. The challenge is to deal with a dynamic environment in order to provide an adequate service to an elderly or a sick person at home in a cooperative way. It is necessary to take into account constraints such as, the degree of urgency of the service and the intrusion degree of the system. The evolution of the degree of intrusion based on the degree of urgency and the availability of the different communication devices of the ambient environment are particularly targeted. Through an adaptive approach based on coalitions of agents, the multi-agents system ensures answers to various and/or unforeseen situations. The originality is to employ a declarative method for modeling the coalitions formation process by means of a rule-based system.

Academic research paper on topic "COALAA-GEN: A General Adaptive Approach for Ambient Assistive Applications"

(8)

CrossMark

Available online at www.sciencedirect.com

ScienceDirect

Procedia Computer Science 96 (2016) 324 - 334

20th International Conference on Knowledge Based and Intelligent Information and Engineering

Systems

COALAA-GEN: A general adaptive approach for ambient assistive

applications

Nadia Abchiche-mimouni*, Antonio Andriatrimoson, Etienne Colle, Simon Galerne

IBISC Lab. University of Evry

Abstract

This paper presents the design and the implementation of general and adaptive framework for ambient assisted living applications. The challenge is to deal with a dynamic environment in order to provide an adequate service to an elderly or a sick person at home in a cooperative way. It is necessary to take into account constraints such as, the degree of urgency of the service and the intrusion degree of the system. The evolution of the degree of intrusion based on the degree of urgency and the availability of the different communication devices of the ambient environment are particularly targeted.

Through an adaptive approach based on coalitions of agents, the multi-agents system ensures answers to various and/or unforeseen situations. The originality is to employ a declarative method for modeling the coalitions formation process by means of a rule-based system.

© 2016 The Authors.PublishedbyElsevierB.V. This is an open access article under the CC BY-NC-ND license

(http://creativecommons.Org/licenses/by-nc-nd/4.0/).

Peer-review under responsibility of KES International

Keywords: adaptativity; multi-dimensionality, cooperation, assistive applications, multi-agent systems

1. Introduction

Adaptivity is widely studied as a capability that makes a system able to exhibit a cooperative and intelligent behavior. Moreover, software increasingly has to deal with ubiquity, so that it can apply a certain degree of intelligence. Our specific context is to assist an elderly or a sick person in loss of autonomy at home by providing assistive applications based on cooperation among a robot and communication objects (CO). Maintaining such people at home is not only beneficial to their psychological conditions, but helps to reduce the costs of hospitalizations.

Corresponding author. E-mail address: nadia.abchiche@ibisc.univ-evry.fr

1877-0509 © 2016 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license

(http://creativecommons.Org/licenses/by-nc-nd/4.0/).

Peer-review under responsibility of KES International

doi:10.1016/j.procs.2016.08.145

Ambient assistive robotics can be defined as an extension of ambient intelligence, which integrates a mobile and autonomous robot and its embedded sensors on the one hand, and the CO present in the house on the other. The interaction among the components in such systems is fundamental. In (Andriatrimoson and al. 2012) the authors presented a coalition-based multi-agent system (COALAA) for implementing an ambient assistive living framework that takes advantage of an ambient environment (AE): a robot and its embedded sensors, cooperating with a network of communicating objects. The aim was to provide a service to the person in an adaptive way. A coalition of agents proposes a set of data and the way of combining these data in order to offer the desired service to the person. Adaptation is needed because the context is dynamic and difficult to predict. Depending on the context, the same service can be achieved by different combination of the data. A multi-agent system (MAS) reifies the sensors and the robot, allowing the cooperation among the agents by means of coalitions formation. In order to provide a service such as fall detection, the agents combine the data according to their availability and the relevance. Moreover, the system has to deal with privacy and intrusion level so that one minimise causing inconvenience.

In this work, we propose an improvement of COALAA by: (1) embedding a rule-based reasoning (RBR) module in the agents in order to reason about the coalition formation criteria, and (2) extending the scope of the adaptiveness to ethical, functional, computational, functional, methodological, and control dimensions. Due to these reasons, our new approach can be considered as a general approach for implementing adaptation in ambient assisted applications. New CO can be added in a dynamic way and the way of forming the coalitions can be tuned by the user by introducing new rules in the system.

The rest of the paper is organized as follows: the section 2 gives a review of related work in adaptation and ambient assisted living domains. The Section 3 presents a general adaptive approach COALAA-GEN, highlight its benefits and show its results validation. Finally, Section 4 concludes with some improvements and perspectives.

2. Related work

The ambient assisted living program catalogue (Ruijiao and al. 2015) shows the extent and the diversity of the projects in ambient assisted living domain. The European Commission currently funds no fewer than 160 projects through the AAL Joint Program. In the context of ambient intelligence, the CO of the AE play a "facilitator" role in helping the robot in the Ambient assistive living. Conversely, sensors and robots can be seen as CO, which are used by services to the person in loss of autonomy. Several projects have been interested in combining home automation, pervasive sensors and robotics, for the safety of the patient at home.

The CARE (Eduardo and al., 2003) project is a Research and Development activity running under the Ambient Assisted Living Joint Program, which is co-funded by several European countries. Its main objective is fall detection and person monitoring at home by Smart camera. As part of this project, algorithms essentially based on a biologically inspired neuromorphic vision sensor for fall detection have been developed. The system aims to define a level of reliable supervision by avoiding as much as possible interactions with the person in her/his own home.

ProAssist4Life (Remagnino and al., 2005) is a German project of situation of helplessness detection System for elderly. This project consists in developing an unobtrusive system that provides permanent companionship to elderly people living in single households or in retirement facilities. Multisensory nodes mounted on the ceiling of a room register an individual's movements. One multisensory node contains six motion sensors, one brightness sensor, and one oxygen sensor. According to data provided by various physiological sensors, the system is based on a predictive approach based on finite state automata modeling the previous activities of the patient.

These two projects deals in priority with fall detection, but none uses adaptation features. Another project developed at the University of Camerino is named ACTIVAge (Corradini and al., 2011). In order to keep people at home also, this project aims to provide services and teleservices based on the context. The system consists of adaptive planning solver based web services orchestration and choreography with decision making algorithms. A knowledge base is used to model persistent data of the ambient environment. Adaptation in this project is based on a multi-level ontology meta-model, allowing the description and the annotation of environmental and emotional data. The problem with such cognitive approaches is that they are time consuming. They are not suitable in case of emergency situations. A survey of cognitive assisted living ambient systems can be found in (Ruijiao and al., 2015). Ambient intelligence tries to adapt the technology to the needs of individuals and their environment based on ubiquitous computing. In this context, MAS comprises one of areas that make a strong contribution to the paradigm of adaptability in ambient assisted applications. Over the last decade, various projects

have been addressed the adaptability of technology at home based on MAS in order to improve the security and automating medical staff's work. The IDorm project (Rombach and al., 2011) is one of the pioneering projects in ambient intelligence. It is designed to assess an ambient environment composed of three categories of objects: static CO associated with the building, a robot and mobile devices. IDorm architecture is based on a MAS that manages the operations of all the environment sensors and the robot. One agent, named the sensor agent, controls the sensors, and another controls the robot. The sensor agent receives the different measures from sensors and controls actuators, which are linked to sensors like a pan-tilt camera. The robot agent acts as a data server and coordinates exchanges of information between the user and the robot. It controls the navigation of the robot by combining different functions such as the obstacle avoidance or the search for targets. This continuous project has been upgraded to iSpace where learning capabilities are added to the system. The new version of the system learns about the behavior of the users. While this contribution is very important because it is one of the precursors in the domain, the adaptiveness and the distributed control, inherent in MAS have been insufficiently exploited. Only two agents are present in the MAS.

ALZ-MAS (Corchado and al., 2008) is an Ambient assisted Application based MAS aimed at enhancing the assistance and health care for Alzheimer patients. ALZ-MAS takes advantage of the cooperation between intelligent agents and the uses of context-aware technology providing a ubiquitous, high-level interaction with adaptable interfaces between users, system and environment. System structure has five different agents that are embedded into deliberative agents based on the BDI (Rao and Georgeff 1995, Wooldridge and Jennings 2009) architecture. ALZ-MAS makes use of RFID, Wifi-Network and ZigBee devices, providing the agents automatic and real time information about context. The devices agent interacts with the Zigbee to control physical services. Furthermore the admin agent processed all information obtained. The essential aspect in ALZ-MAS is the use of a set of technologies which are integrated to agents so that to provide multiple services. Moreover, a cooperation agent is employed to plan and schedule the tasks defined for the nurses connected to the system. The strength of this approach is that it provides a middleware, which is a BDI-based MAS system, allowing it to be easily improved in terms of reasoning ability. But this system is used for the nurses and does not interact with the person at home.

CAMAP is a framework presented by Ferrando and his colleague in (Ferrando and Onaindia, 2013). This work lays on adaptive and intelligent behavior of ambient assistance. CAMAP system aimed to apply multi agent planning based on argumentation-based defeasible logic for deciding the action that meet the needs of the patient. The argumentation (Amgoud 2009) and the partial planning methods (Durfee and Lesser, 1991) are used as a mechanism for dealing with context adaptation. The overall goals of CAMAP protocol is to collaboratively and progressively refine the base plan until it becomes a solution plan. For these reason, CAMAP seems to be relevant to find optimal solutions and could be less efficient if one have the obligation to obtain a result.

Because of their simplicity, we have focused on coalition formation in MAS in this work (Huhns 1987; Huhns and Singh, 1997; Wooldridge and Jennings, 2009). Indeed, such mechanisms ensure adaptive solutions without taking into account the optimality of the solutions. So, they are able to provide quick solutions, which can be refined if needed.

As previously said, the present approach takes into account ethical, computational, functional and methodological, and control dimensions. This can be illustrated by the fact that the evolution of the inconvenience, that is intrusion level of the system, is based on the degree of urgency and on the availability of different CO. So, the problem may be seen as an optimization problem where optimization criteria values change during a session. Moreover, the criteria to consider change too. The purpose is to design a dynamic and adaptive way for: (a) selecting the agents participating to the tasks to achieve and, (b) taking into account several criteria in a dynamic way. The general adaptive coalition-based approach presented in below fit these constraints.

3. Designing a general coalition-based approach architecture

The principle of coalitions aims at temporarily putting together several agents for reaching a common goal. (Sims and al., 2003; Chalkiadakis and Boutilier, 2012; Soh and Tsatsoulis, 2001; Soh and Tsatsoulis 2002) have illustrated the relevance of coalition-based approaches for adaptiveness. The methods are various: either incremental or random or centralized. But, all of them proceed in two stages: (1) the formation of agent coalitions according to their ability to be involved in achieving a goal and (2) the negotiation stage among the coalitions in order to choose the one that provides the closest solution to the goal. The interests of the coalition-based formation protocols are the

flexibility with which coalitions are formed and straightforwardness of the coalition formation process itself. The coalitions can get rid of dynamically reorganize with local and simple rules defined in the agents.

3.1. COALAA: a coalition-based approach for ambient agents

COALAA is MAS based on a coalition-based approach for ambient agents. Each agent in COALAA encapsulates a CO and decides in a local and proactive way when and how to contribute to the required service to the person. A more general notion than a service, called an effect has been introduced. An effect can be a particular lighting at a precise place of the residence or the localization of a robot. The MAS configures itself for providing a solution according to the availability of the CO and the respect of criteria. Note that the goal is not to find the optimal solution but a solution close enough to the required effect. In the coalition formation protocol, the obligation to obtain the required effect and an intrusion level depending on the urgency of the situation, are the most important considered criteria. They are also used during the reorganization of the agents while trying to achieve a desired

effect. The effect obligation criteria is used in priority while the level of intrusion is modified only if needed, i.e., to acquire new data and thus to activate the sensors (ex. tilt-camera) likely to cause discomfort to the person.

Figure 1. Architecture of COALAA

As shown in the Figure 1, several kinds of components are necessary to deal with the complexity of COALAA. An effect is modelled in the form of a triple < t ; c ; f > where:

• t € T, c € C

• T is a set of task labels: localization of a robot or a person, enlightening.

• C is a set of criteria: accuracy, efficiency, time constraint, and neighbourhood.

• F is a list of influencing factors: intrusion level, urgency degree.

The designer of the system statically assigns the criteria, while the influencing factors are assigned by the end-user.

The information handled by the system is classified into two types. This so-called persistent information, related to the application domain, puts together data about the structure of the residence and the features of the CO. The second type concerns volatile data mainly the measures provided by the sensors and the orders sent to actuators. The volatile data are distributed in each agent, while persistent data are stored in an ontology named AA (Ambient Assistance) (Kivela and Hyvonen 2002; Arnand and al., 2003). The AA ontology contains four categories of information related to the application domain: The Home category for defining the structure of the environment, the CO category for knowing their characteristics and their operating mode, the User category for defining the user profile and the Task category that puts together the tasks and services achieved by the system.

The Gateway is a module for the standardization of information exchanged between the ambient environment and the MAS. Its role is to make the agents manipulating the common information format. This standardization is necessary because of the heterogeneity of protocols from different manufacturers.

3.1.1. Agent internal architecture

The agents of the MAS are created according to the ontologies concepts. Each agent is assigned an internal architecture able to take in charge the agent adaption and reactivity by using three main parameters that are: neighborhood, history, and ability. The neighborhood sets the list of agents that are close to this agent at a given time, according to the topological distance. The history stores previous perceived information that comes from the sensors. This is a simple succession of perceived data, which helps to consider the timescale during the process of coalition formation. The ability identifies the skills of the agent, which are directly related to the encapsulated CO.

3.1.2. Agent behaviors

In the process of the coalition formation, an agent may be either initiator or candidate. Any agent whose ability can partially meet the desired effect can be a coalition initiator. The initiator exchanges messages with other agents, potential members of the coalition, called candidate agents. The Protocol is based on exchanges of messages between the initiator agent and candidate agents. As soon as the overall ability of the coalition is close to the desired effect, the initiator agent is pending the negotiation phase. At the end of the coalition formations, each initiator agent that is the referent of a coalition is negotiating with other initiators agents to choose the winning coalition. The coalition whose ability is the closest to the desired effect is the winning coalition.

The concept of ability is generic. In the localization application example, it is instantiated by the measures precision. The principle is simple. Each initiator agent sends a message that contains the ability obtained by its coalition. On receipt of this message, each initiator agent compares the ability of the coalition it received to its own one. If its ability is lower than that received, the coalition will be no more considered, otherwise, it is a winning coalition up to receiving a new message. Apart from the desired effect, the formation of coalitions uses other criteria such as the topological neighborhood to reduce the response time or the obsolescence of a measure when the desired effect depends on sensor data. Thus, the first step is the identification of candidate neighbors according to its own location in the environment (defined by the topological distance) and the desired effect. The aim of this strategy is to ensure that a result will be provided (cf. obligation of result criteria). For that purpose, the first selection criteria considered is the topological distance. Once all candidate agents are known, each initiating agent continues the selection of candidates based on the recent measures criteria. When no coalition is able to meet the desired effect, a new search for a successful coalition is restarted after having relaxed the constraints on certain criteria. Indeed, it is possible to increase the level of intrusion of the system despite of the tranquility of the person at home. This authorization to increase the level of intrusion allows, for example, operating a pan-tilt camera of the robot in order to acquire new measures and restart the process by finding a winning coalition.

The protocol of coalition formation is composed of two distinct steps. The first step consists in forming coalitions of agents according to their ability. The second step is a negotiation and refining phase so that the best one, in satisfying the desired effect criteria, is chosen. Figure 2.a. summarizes the agents' behaviors. The baseline algorithm is described in Figure 2.b. After initialization, the exchanges among agents follow three main actions: formation of all possible coalitions for each referent, selection of the best coalition according to the coalition precision, deployment of the winning coalition.

The agents' interaction semantic is based on speech act theory, introduced by John Searle (Searle 1969), allowing the agents to assign a semantic to each message by defining a message a type. The most important types are: Request, Response, Initiate, Acknowledge, Accept and Negotiate.

S Receive a frame from the

environment.

BMavMur v sxnfl ',-n unavlour

S Recover the sensor associated

identifier. S Access the Ontology and update

ability.

Figure 2.a. Agents behaviours

Figure 2.b. Baseline agents' algorithm

3.1.3. Discussion

COALAA shows the feasibility and the relevance of coalition-based MAS for ambient assisted scenario. It has addressed the very important problem of fall detection by exploiting the availability of CO present in the houses. To fit the obligation for the system to give a result, CAOLAA requires the user for manually assign a priority to the criteria and the bounds for the values of the criteria. The next section illustrates this weakness and shows a way of solving the problem.

3.2. COALAA-GEN: a generalized criteria management for agents coalitions formation

Figure 3 shows an example scenario. A robot in the person's home; the patient has fallen. To move towards her/him and to guide its camera to the remote caregiver, the robot has to be located first. A visual contact will then help the remote caregiver to perform a correct diagnosis of the situation. Depending if the robot is the room P1 or the room P2, the CO required are different.

Figure 3: Fall detection scenario

The Figure 4 briefly shows how the MAS solving this problem. More details can found in (Andriatrimoson 2012). Three kinds of CO are involved: a robot pan-tilt camera, a fixed camera and a presence detection sensor. Three respective ambient agents encapsulate these three CO: a Presence Detector Agent (APD), a Fixed Camera Agent (AFC) and a Pan-Tilt Camera Agent (APTC). Visual markers like Data matrix are associated with each camera. Following the fall of the patient, a request for a localization effect is generated in the form of a triple < t; c; f > where t is a localization task which matches with the desired effect, c matches with a singleton containing the precision criterion needed for the localization task and f matches with a set containing two influencing factors that are the intrusion level and level of urgency. In the considered scenario, we have considered a precision equal to 0.1, a level of urgency equals to 3 (three levels of urgency are considered: low=1, medium=2, high=3) and an intrusion level initialized to 0 (the less intrusion level). So, the triple becomes: <Locate ; f0:1g; f3; 0g>. The Interface agent (AI) has received the desired effect and then broadcasts the request InitCoal (<Locate; f0:1g; f3; 0g>) to all the agents of the MAS. As son as each agent receives the desired effect, it checks its ability. As all sensors in the environment have a precision that is not better than the desired effect, each agent initiates a coalition with immediate neighborhood. In this figure, only interactions with APD agent are shown. Assuming that all agents are topologically close, APD broadcast a coalition formation request by sending an InitCoal message. Each agent receiving the initialization message checks if its ability is adequate with the request of coalition formation. If yes, it sends an acceptance message labeled AcceptCoal to be a candidate. Such a message contains the precision of the agent. APD adds progressively answer acceptance, and accumulates the abilities, which are the precision in the considered localization task. By this way, it calculates the overall ability of the coalition until it reaches that of the desired effect. Then, it sends ACKCoal acceptance to confirm the membership of the candidate to the formed coalition. The next step is to activate the coalition. The robot moves to the place designated by the coalition and

guides its pan-tilt-camera to the remote caregiver. First of all, the distant user has to verify that the person is in his field of vision, so it can perform a correct diagnosis of the situation and adopt an adequate action. Conversely, if the person is not well located the system restarts searching for a new result, after having increased the intrusion level. This allows the cameras to be moved randomly so that the chances of getting a visual marker are increased. The consequence will be improving the precision of the result (i.e. the precision of the coalition).

Figure 4: Interaction diagram

3.2.1. Agent rule-based reasoning module

The previous scenario shows that criteria management is critical. Indeed, obtaining a successful coalition depends on the order in which the criteria have been considered. In the above scenario, if the first considered criterion was the level of intrusion (instead of the precision), then the first result would have been the correct one. Then, the question could be the following: why can one not have a management criteria step integrated in the coalition formation process? This is the main contribution of this work. We have introduced into each agent of COALAA a rule-based reasoning (RBR) module responsible of determining a priority of the criteria to consider according to the context. The RBR is also responsible of assigning and adjusting the criteria values. The RBR is used (instead of the algorithm described in Figure 2.a) for interleaving the execution of the behaviors in a dynamic way.

A RBR system is composed of a knowledge base (KB) and an inference engine. The KB contains a set of rules and a set of facts. The rules are given in the form of implications. The knowledge facts (KF) describe the state of the world. The inference engine is a special interpreter that controls the triggering of the rules according to the KF.

The form of a rule is: IF <antecedent> THEN <consequent>. Antecedent is the condition that must be satisfied to trigger the rule. Consequent is the performed action when the rule is triggered. Antecedent is satisfied if the condition matches the facts in the KF. Some examples are given below.

Instead of having a procedural control, each behavior is modeled by a production rule whose activation condition is precisely the context of its execution. The behaviors of the agents are associated with trigger conditions. These conditions represent the context that makes behaviors possible to be executed. Explicit chaining between the behaviors is no more needed since the inference engine triggers the rules. For example, the AcceptCoal behavior is chained with the InitCoal behavior. So, the InitCoal behavior is executed once the AcceptCoal behavior is terminated. Here are some examples of expressing these assertions in Jess rules (Java Expert System Shell, http://herzberg.ca.sandia.gov/, accessed on 2016 March, 10) syntax:_

(defrule check-ability "the agent accepts joining the coalition if it has the required ability" (Message InitCoal ?x) (Ability ?x) => (assert (Behavior AcceptCoal ?x)))

(defrule perform-ability "Create an accept message" (Behavior AcceptCoal ?y) => (bind ?m (createMessage (AcceptCoal ?y)))

These rules express that if the agent has in its working memory an InitCoal message and if the agent has an ability ?x, so the rule can be triggered. In this case, the core of the behavior associated with the rule is executed. Note that all behaviors are not controllable. Some of them, such as the message reception behaviors are automatically executed to threat the reception of the messages. In COALAA-GEN, the architecture of the agents has been modified with an embedded RBR module responsible of a declarative reasoning process. In the new agent

architecture, the inference engine replaces the procedural algorithm implementing the decision module. The KF represent the knowledge that has been extracted from the ontology, the perceived data and the exchanged messages among the agents. For that purpose, a set of rules is defined to determine, depending on the context, the most relevant criteria to consider first at each step of the coalition formation process. On the other hand, when the coalition proposed by the system is not a correct one, the RBR is in charge of determining the most relevant criteria to relax or to modify. The involved rules in this case are some kinds of heuristics that guide the coalition process in managing the criterion. For example, if a CO involved in the coalition does not include a CO whose precision is sufficient, it is advisable to relax the intrusion level. This increases the degree of freedom of the system regarding to its actions allowing the cameras to be activated or lights to be switched on.

Another use of the RBR for the management criteria concerns the addition of new criterion such as data freshness. It is sometimes more relevant to consider not sufficiently precise data if they are very recent. For example, a presence detector can only inform that the person is situated in a particular room. Suppose that a particular presence detector "informs" that the person is in the room R1 and a camera "shows" that the person is in the right corner of the room R2. Obviously, the information given by the camera is more accurate, but if it is too old it should be obsolete and will not help correctly locating the person. It is suggested here to consider the date of perceived information for determining the priority of the criteria.

3.2.2. Adaptiveness multi-dimensionality

The results are obtained in a real environment composed of heterogeneous sensors and markers. The platform includes several sensors of the market and dedicated sensors developed in our laboratory. The environment is composed of three rooms equipped with a set of sensors and the robot with its own sensors. The localization is based on goniometric measurements provided by robot on-board sensors and environment sensors. These can provide localization information allowing the localization of the robot in its environment using real-time data either from the robot on-board sensors or from the sensors in the environment. COALAA-GEN has been implemented using the Jade multi-agents platform (Bellifemine and al. 2007), where each agent embeds an instance of Jess. The production rules are given as a text file input parameter to the agents. A user interface allows introducing new rules.

3.2.2.1. Computational adaptiveness

COALAA-GEN and COALAA have been compared to the well-known CNP protocol (Smith 1980). The Figures below shows the obtained results. The tests have been performed with a dozen scenarios. Each scenario has been executed with CNP, COALAA and COALAA-GEN. For COALAA and CNP, different values for the criteria have been experimented. COALAA-GEN has been tested with the same collected data, without any user intervention for criteria management. The figures below summarize the results.

COALAA-GEN

Figure 5. Formed Coalitions

Figure 6. Response time

Figure 7. Exchanged messages

Figure 5 shows the number of formed coalitions depending on the number of agents present in the MAS. The preferred strategy in our approach is to obtain a maximum number of coalitions that meet the selection criterion. The goal is to maximize the number of solutions to meet the request to increase the chances of securing a result. The number of coalitions is less than or equal to the number of initiators. In terms of the number of formed coalitions, the Contract Net protocol is less efficient than COALAA. COALAA-GEN gives the result with fewer numbers of agents. This can be interpreted by the fact that "intelligent" criteria management helps the agents to be more relevant for coalition formation. The response times are compared (see Figure 6). This time corresponds to the time spent in calculating the coalitions, including the message exchanges. The fact that the number of coalitions that the CNP can form is lower than the number of initiators has a direct effect on the response time. It also impacts the number of

exchanged messages represented by the Figure 7. The curve representing the number of exchanged messages follows the same rate for CNP, COALAA and COALAA-GEN. However, COALAA-GEN shows a higher number of exchanged messages. Unlike the CNP, COALAA and COALAA-GEN avoid system crashes, by a progressive coalition formation, which in contrast increases the number of exchanged messages. In terms of time response COALAA, COALAA-GEN and CNP are almost similar; CNP is slightly better in terms of response time. But in terms of obtained COALAA-GEN is the best. Indeed, a failure can be catastrophic and thus the few milliseconds delay in the response time may be insignificant, if success to complete the task is assured. This is explained by the fact that COALAA-GEN continues reorganizing until a solution is found (even with deteriorated criteria).

3.2.2.2. Methodological and functional adaptiveness

The genesis of the MAS is done automatically in COALAA-GEN. In spite of the fact that this has not been detailed in this paper, this is a very important feature of the system. In fact, modifying the AE, by adding or suppressing CO, automatically updates the ontology and triggers automatic MAS reconfiguration. In case of such modifications, the user does not need to do any specification to make the system adapting its architecture to AE dynamic updating. This ability of the system is qualified by methodological adaptiveness. We refer to functional adaptiveness while dealing with services that the system can offer to the user. The description of the ability of the CO used by the agents to construct services according to the "effect description" is included in the "task" ontology part. This allows the agents to perform an automatic detection of their ability to perform an effect.

3.2.2.3. Ethical adaptiveness

An original specificity of our system is that it deals with ethical dimension in an adaptive way. Adding ethical values as criteria for forming the coalitions ensures this specificity. The level of intrusion of the system is modeled in such a way that it is upgraded only in case of emergency and if the user wishes to. Moreover, the personal data are stored in the equipment of the house and are uploaded only if needed by the distant caregiver and if the user has agreed.

The degree of intrusion of each CO is modeled in the ontology as an attribute associated to the CO concept. The personal data are kept locally in the agent and are not stored in the distant ontology. But if the distant caregiver needs it (in case of emergency), the private data are uploaded, with a special status that is, volatile. This means that they are deleted from the distant storage as soon as they have been used. In the presented scenario, only two ethical criteria have been considered: the level of intrusion and the data privacy. They have been modeled as criteria for coalition formation. As shown by the rules below, adding new criteria is performed by adding new rules._

(defrule crit-manag-001 "add new criteria" (Crit ?type ?name) => (assert (Coal Crit ?name))

(defrule crit-manag-002 "assign new criteria for coalition formation" => (modify (Coal Crit ?name)))

3.2.2.4. Control adaptiveness

The fact that an inference engine has been employed instead of a procedural algorithm has a direct effect on the intelligence of the system. The behaviours are involved only when their associated rules are triggered, which are themselves triggered when some declarative conditions are met. Since the conditions of the rules can be modified without any procedural modification, the control of the execution of the behaviours is completely adaptive. The user can control and modify the execution of the behaviours even at rune time. Furthermore, the system is also able to detect missing information that is able to lead to the execution of a particular behaviour. This is ensured by backward chaining rules. The engine seeks steps to activate rules (when necessary) whose preconditions are not met.

The rule given below illustrates the control adaptiveness of COALAA-GEN.

(defrule ctrl-001 An alarm has occurred, but no behavior can be triggered" (Alarm ?x ?y) (not (Behavior ?z ?t)) => (assert (Backward ?z ?t))|

More generic the rules are, more the system intelligence can be improved. If it is easy to convince that adding new rules does not imply modifying the implementation of the system, suppression of rules is not trivial. The process of suppressing rules has to deal with consistency of the remaining data. More explanation is needed about the Jess engine, which is beyond the scope of this paper.

4. Conclusion and perspectives

We have introduced a new general approach for improving adaptiveness in ambient assistive applications. A RBR module has been embedded in the agent architecture to dynamically assign the criterion to consider during the coalition formation process and we proposed to deal with the adaptation at different levels. The adaptiveness has been considered according to four dimensions: (1) computational dimension: during the coalition formation process, (2) functional and methodological dimension: while service modeling, (3) ethical dimension: associating the intrusion level to the degree of emergency, (4) control dimension: for behaviors triggering and criteria management. We have compared the obtained results with those previously obtained without the RBR, and we have observed that the adaptiveness has been improved without any performance degradation. The feasibility of this general approach has been showed on a usage scenario to remove the doubt of a false alarm in fall detection. The first results illustrated with robot localization are promising. The validation of the system with a great data size is currently performed by the generation of statistical distributions of data that provide more meaningful results. Moreover, several improvements are under consideration. The efficiency can be improved by making the initiator agents revising the way of choosing their partners during the coalition formation process. This can be based on agent past-obtained results. The agents can infer the capabilities of their potential partners through repeated interactions such as done in (Chalkiadakis and Boutilier, 2012). We think that this can be compiled into a set of rules that will be added into the RBR module.

This approach is currently applied to cognitive stimulation and will shortly be applied for detecting the activity of the person. At a more long-term perspective, we will propose to physically wrap an agent in each CO, so that no time is spent to acquire information from a gateway.

5. Bibliography

Amgoud L. (2009). Argumentation for Decision Making. In Argumentation in Artificial Intelligence, Rahwan, Iyad (Ed.), p. 301-319 Andriatrimoson A., Abchiche-Mimouni N. (2012). E. Colle, and S. Galerne, An adaptive multi-agent system for ambient assisted living. In

ADAPTIVE 2012, IARIA, ThinkMind, July, pp. 85-92. Arnand R., Robert E. M., Roy H. C., and Dennis M. (2003). Use of ontologies in a pervasive computing environment. In Knowledge Engineering

Review, vol. 18, pp. 209-220. Bellifemine F.L., Caire G., Greenwood D., Developing Multi-Agent Systems with JADE. Wiley Eds. 2007.

Chalkiadakis G., and Boutilier C. (2012). Sequentially optimal repeated coalition formation under uncertainty. Autonomous Agents and MultiAgent Systems archive, vol. 24, no. 3, p. 441-484. Corchado JM, Bajo J, De Paz Y, Tapia DI (2008) Intelligent environment for monitoring alzheimer patients. In Agent technology for health care.

Decision Support Systems, pp 382-396. Corradini F., Merelli E., Cacciagrano D. R., Culmone R., Tesei L., and Vito L. (2011). Activage: proactive and self-adaptive social sensor

network for ageing people, ERCIM News, vol. 2011, no. 87. Durfee E. H., Lesser V. R. (1991). Partial Global Planning: A Coordination Framework for Distributed Hypothesis Formation. IEEE Transactions

on systems, man, and cybernetics, VOL. 21, NO. S, September/October, p. 1167-1183. Eduardo A., Kudenko D., and Kazakov D. (2003). Adaptation and Multi-agent Learning. Springer-Verlag Heidelberg (Eds.). Ferrando S. P. and Onaindia E. (2013). Context-Aware Multi-Agent Planning in intelligent environments. Information Sciences 227, p. 22-42. Huhns M. N. (1987). Distributed Artificial Intelligence. Pitman. Huhns M. N. and Singh M. P. (1997). Readings in Agents. Morgan Kaufmann.

Kivela A. and Hyvonen E. (2002). Ontological theories for the semantic web. In Semantic Web Kick-Off in Finland, May pp. 111-136. Rao A. S., and Georgeff M.(1995). BDI Agents: from theory to practice. In Proceedings of the First International Conference on Multi-Agent

Systems (ICMAS-95), pages 312-319, San Francisco, CA, June 1995. Remagnino, P., Hagras, H., Velastin, S. and Monekosso, N. (2005) Ambient intelligence: a gentle introduction. In: Remagnino, Paolo, Foresti,

Gian Luca and Ellis, Tim, (eds.). Ambient intelligence: a novel paradigm. New York, U.S.: Springer. Ruijiao Li, Bowen Lu and Klaus D. McDonald-Maier, Cognitive assisted living ambient system: A survey, Digital Communications and

Networks, http://dx.doi.org/10.1016/j.dcan.2015.10.003 D. Rombach D., Storf H., and Kleinberger T. (2011). Situation of helplessness detection system for senior citizens, October, p. 32-33. Searle J. (1969). Speech acts. an essay in the philosophy of language. Cambridge University Press.

Sims M., Goldman C., and Lesser V. (2004). Self organization through bottom up coalition formation. In the 2nd AAMAS, 2003. T. Scully, M.

Madden, and G. Lyons, "Coalition calculation in a dynamic agent environment," in the 21st ICML. Smith R. (1980). The contract net protocol: high-level communication and control in a distributed problem solver. In IEEE Transactions on computers, p. 1104-1113.

Soh L.-K. and Tsatsoulis C. (2001). Reflective negotiating agents for real-time multisensor target tracking. In IJCAI'01, p. 1121-1127.

Soh L. and Tsatsoulis C. (2002). Allocation algorithms in dynamic negotiation-based coalition formation. In AAMAS02 Workshop 7 "Teamwork

and coalition formation", pp. 16-23. Wooldridge M. and Jennings N. R. (2009). Agent theories, architectures, and languages: a survey. In Intelligent Agents, Wooldridge and J. Eds., Berlin: Springer-Verlag, pp. 1-22.