The ageing process of population of the First World countries is increasing the interest towards solutions to improve the quality of life of elderly or disabled people and their families, providing an economically sustainable healthcare. Thus Ambient Assisted Living (AAL) and Ambient Intelligence (AmI) are moving towards the Arti cial Intelligence eld to develop modular, adaptable and intelligent systems to cope with changing needs that characterize the life of people with chronic diseases. These systems should nd smart and simple ways to improve the patient quality of life, enhancing at the same time the communication between the system, the assisted person, his family and the medical sta , acting also as a lter of information that provides useful and signi cant data. This paper introduces the basic architecture of an expert system for AAL: the Virtual Carer (VC). It has to be understood as an IT system modelling a distributed, reliable and modular sensor network composed by biometric and ambient sensors, being able to communicate with an assisted person, to monitor his health conditions and to control the environment around him. The main goal of the system is to help an elderly patient with his daily activities ensuring his security. The proposed system is based on a multi-agent architecture to ensure its exibility and interoperability: new devices or sensors can be added simply adding new agents, thanks to the standardization of agent communication. The system includes also a reasoning part based on Belief-Desire-Intention (BDI) paradigm, trying to model the behaviours of a human caregiver.
It is a well-known fact that a shift in the distribution of a country's population
towards older ages is ongoing in almost every country in the world, to the point
that the world has never seen as aged a population as currently exists globally.
Due to their higher dependency ratio (i.e. the ratio between the number of people
over 65 years old and those of working age), this is becoming a major problem
in Europe, USA and Japan [
The increasing median age of the population has signi cant social and
economic implications. For example, it results in a rise in the number of chronic
diseases causing the increasing of health-related emergencies and the growing of
the healthcare spending [
Through the AAL Joint Programme, the European Union aims to foster the
emergence of AAL services and systems for ageing well at home, in the
community, and at work [
Every AAL system is based on pervasive devices typically used in Home
Automation systems, and on Ambient Intelligence (AmI) technologies to
integrate devices and build a safe environment for the assisted person [
AAL and AmI systems have to adapt themselves over time to cope with
the changing needs and health conditions of the assisted person. They need to
perceive variation in habits that can signal health-related problems or stressful
situations. Thus, applying Arti cial Intelligence techniques on this domain seems
to be a promising direction for research [
Our Contribution In this paper we propose a Multi-Agent System (MAS) for Ambient Assisted Living: The Virtual Carer. It should be understood as a system to manage a distributed sensor network composed by ambient and biometric sensors, being at the same time an interface layer between the network, the assisted person, his relatives and the medical sta . Furthermore it includes a goal-oriented reasoning module, i.e. the Virtual Carer Agent (VCA), based on Belief-Desire-Intention (BDI) paradigm.
The adoption of the MAS architecture derives from the natural features of
these kinds of systems. As stated in [
The MAS architecture allows to distribute software agents and the platforms that manage them over one or more networks; thus each sensor can be associated with an agent, making the system resilient to sensor failures and disconnections. The adoption of the FIPA standards as the Agent Communication Language (ACL) allows to add and remove sensors and devices in a transparent way, simply wrapping them with an agent.
In addition to the listed features, we took into account also pro-activeness, imagining a system that can autonomously take the initiative, guided by goals. Thus the reasoning module of the system, i.e. the VCA, is based on BDI paradigm. The VCA's main goal is to model logical structures, reasoning and behaviours similar to those of a human caregiver, transforming data from sensor agents in logical predicates representing its knowledge base. To perform actions on the monitored ambient and interact with the assisted person, VCA can collaborate with agents controlling actuators. Thus the system can monitor the health conditions of the assisted person and facilitate his daily activities. 1.2
Paper Structure The rest of the paper is organized as follows. The next section deals with some related works on MAS, AAL, HCI and HRI; then the core of the proposed system is described. Section 4 deals with the description of the whole agency, describing the implementation of a simple simulation scenario and highlighting some features of the Virtual Carer. In the end conclusions are drawn and future work suggested. 2
As stated in [
BDI agents act on the basis of practical reasoning [
A di erent research area is focused on Internet of Things scenarios for AAL
and remote healthcare [
Also sensor networks, HCI, HRI and robotics applications are gaining an
increasing attention, nding in AAL a natural eld of application. In [
The Virtual Carer is a MAS modelling a distributed, reliable and modular sensor network composed by biometric and ambient sensors, and integrating a BDI core agent (Virtual Carer Agent, VCA). The Virtual Carer is an IT system able to communicate with an assisted person, to monitor his health conditions and to control the environment around him. To cope with a highly variable environment the VCA models reasoning mechanisms and behaviours similar to those of a human being following the BDI paradigm. The large amount of information used by VCA is represented by logical predicates, forming its Knowledge Base (KB). VCA chooses the right actions to perform on the environment with an inference engine applied on its KB. The VCA works as follows: { it analyses data provided by sensors and devices forming the system (i.e. by agents controlling them) and updates its KB; { when its KB is updated, VCA generates new knowledge using its inference engine; { with backward reasoning rules applied on its KB, VCA selects the main goal and chooses (through backtracking) the plan to satisfy it; { VCA carries out the actions of chosen plan, collaborating with agents responsible for actuators, in order to directly act on the environment.
Figure 1 shows the basic architecture of the system. The system is composed
by the BDI Virtual Carer Agent, a Register Agent and a number Actuator
Agents and Sensor Agents.
The main goal of this agent is to provide assistance to the elderly, through the
execution of its plans on the basis of the context (its belief base), according to
the BDI paradigm. Figure 2 shows the Actor Model in the i* language [
Goals can be satis ed thanks to several tasks: the VCA has to identify the position of the elderly, his health conditions and to detect anomalies using available sensors. All the detected anomalies are registered through the Register Agent. In order to complete these tasks the VCA has to register with the Directory Facilitator (DF) of the whole agency, to get its services available for other agents. 3.2
Sensor Agents Sensor Agents can be distinguished in two types: Ambient Agents and Health Agents. Ambient Agents are responsible for reading the value of ambient sensors as, for instance, the temperature of a speci c room. They check if the recorded values are in a predetermined range: if the value is outside of the range, the Ambient Agent communicates the anomaly to the VCA ( gure 3). Presence sensors are managed by Ambient Agents, too. For example, we can consider a Passive Infrared (PIR) sensor: the Ambient Agent controlling it merely informs the VCA about the presence of someone in the monitored room. Health Agents, instead, manage sensors measuring values related to the health conditions of the assisted person. These agents are responsible for the activation and deactivation of the devices composing the system, such as speakers, lights, monitors and so on. As highlighted in gure 4, they must be able to receive request messages (typically from the VCA) in order to execute on/o commands on the devices. Hence, their main goal is the switching of the state of the controlled device, after a request is received. The Register Agent is responsible for the database storage of all the information provided by sensors and of the anomalies detected by the VCA. Figure 5 contains the Actor Model for the Register Agent. Its goal is to register anomalies in the database.
Agency
The typical interactions between the VCA, Sensor Agents, Register Agent and Actuator Agents are shown in gure 6. The VCA receives alarms from Sensor Agents when a value is out of the predetermined range; in addition, the VCA can directly require to read a value because the chosen plan requires data to satisfy the goal. The Register Agent receives and stores in a database values from Sensor Agents and anomaly reports from the VCA. The VCA can require values from the database to the Register Agent. The VCA sends request to the Actuator Agents in order to execute actions corresponding to the adopted plan.
Sensor and Actuator Agents are implemented using the JADE framework [
Simulation scenario the light on/o (AL1, AL2, AL3, AL4) are provided. In addition there is a speaker (SP1, SP2, SP3) for each one of the rst three rooms, and a monitor to visually communicate with the assisted person in room 4. Pressure sensors are supposed to be in each place where the person can sit or lie down (e.g. couch, bed, seats), and contact sensors are supposed to be placed on every door and window, to monitor their open/close state. Body sensors are used to monitor the vital parameters (e.g. heartbeat, body temperature) of the assisted person. For each sensor and device in the described scenario the proper Ambient, Health and Actuator Agents are created. At the beginning of the simulation the VCA's belief base includes facts representing the environment structure and pointing out the available sensors and devices. Figure 8 shows an example of facts describing how the home environment for the simulation scenario is divided in rooms.
partOf(couch,hallway). partOf(bed,bedroom). partOf(livingRoom,home). partOf(bedroom,home). partOf(hallway,home).
partOf(bathroom,home).
Another extract of the belief base is shown in gure 9. The listed facts indicate the areas covered by the available devices. For example the video device is signi cant for the couch; the speaker 2, instead, covers a whole room. covers(tv,couch). covers(speaker1,bed). covers(speaker2,hallway).
covers(speaker3,bathroom).
In addition to the facts in the belief base, the VCA knowledge base is composed also by rules to infer new knowledge and thus to activate plans. The rules allow the VCA to localize the elderly in a speci c room, to establish which actuator covers the room and to decide if it has to be activated for an emergency plan. Some examples of these rules are shown in gure 10.
In this simulation, the VCA has several emergency plans: the one shown in gure 11 is executed when an Ambient Agent, responsible for a temperature admissible(D):- inside(A) & videoDevice(D) & covers(D,A). admissible(D):- inside(A) & audioDevice(D) & covers(D,A). covers(D,R1):- partOf(R1,R2) & covers(D,R2). inside(R1):- partOf(R1,R2) & inside(R2). sensor, reads a value out of the predetermined range. In this case the VCA communicates the anomaly to the Register Agent and activates the plan "notify" in order to evaluate if the monitored person has to be noti ed with some alarms. A similar plan is activated when the body temperature of the assisted person is out of a security range.
+emergency1(V)[source(A)] : true <.print("Room temperature out of range"); .print("from sensor ", A); .print("value ", V); .time(H,M,S); .send(register,tell,anomaly(A,V,H,M,S)); .print("Information sent to Register Agent"); ?em1(N); New = N + 1; -+em1(New); !notify; !canc.
The system was complemented by a BDI agent, the Elderly Agent, modelling behaviours and activities of the person assisted by the Virtual Carer System. For simulation reasons the Elderly Agent is able to send messages to other agents of the system: for example, it informs the PIR agent when it moves in the controlled room, and sends requests to the Health Sensors to know the value of speci c health parameters. To simulate the di erent behaviours of the assisted person, the Elderly Agent was provided with di erent plans. Figure 12 shows a plan including the transition from the hallway to the bedroom, and the request of the body temperature value.
In a rst simulation scenario, we modelled the situation in which the assisted person goes to bed and measures its body temperature. The simulation works as follows. At rst, the Elderly Agent initializes the VCA, providing it the information about its initial position; then, after 10 seconds (i.e. the time necessary to get to bedroom), it informs the PIR Agent of its presence in the nal position (this simulates the detection of the elderly by the PIR sensor). The PIR Agent controlling the bedroom noti es the VCA about the presence of a person in the +-scenario5 : true <.send(carer,tell,init(hallway)); .wait(10000); .send(pir1,tell,at); .send(carer,achieve,queryP(tempB)). bedroom; thus VCA has a new belief (i.e. the presence of the assisted person in room 1), and activates the plan ( gure 13) to turn the light on in the bedroom and to turn it o in the hallway. Finally, the Elderly Agent sends a message to the VCA to know its body temperature value. VCA activates a plan that includes all the actions necessary to require the data from the proper Health Agent and to communicate it to the Elderly Agent.
@ent[atomic] +enter[source(A)] : true <?inside(M); ?spir(A,R); M \== R; -+inside(R); !powerOn(R,M); .print("light on in "); .print(R); .print("light off in "); .print(M); -enter[source(A)].
Beside plans similar to the one just described, we simulated plans without an
explicit request by the Elderly Agent, in which the Virtual Carer had to infer the
right actions. For example, when a window remains open in the bedroom and
the assisted person is sleeping, the decreasing of room and body temperature
should activate the plan to wake up the person and ask him to close the window.
In this scenario, the Ambient Agent associated to the bedroom window informs
the VCA that the window state is "open". At a later stage the Ambient Agent
responsible for T1 sensor communicates to the VCA that the detected value
is out of the predetermined range; a similar message is sent by Health Agent.
VCA updates its KB with these new beliefs and activates a plan ending with a
message to the Elderly Agent, asking it to close the window.
The evaluation of an AAL system that aims at the interaction with the assisted
individual to help him and ensure its security should include several perspectives,
as described in [
At this stage of the work only the developers' point of view can be considered: the simulations we carried out are adequate to underline that MAS and BDI paradigm are useful when applied to AAL contexts. Agents, by their own nature, can be easily distributed over a network. The MAS approach and the adoption of communication standards as FIPA-ACL guarantee modularity to the system, in order to cope with the addition or removal of devices. Several agentbased frameworks, as JADE and JASON (both adopted for our simulations), provide techniques to respond to fault-tolerance requirements: the Main Container Replication Service (MCRS) ensures the execution of the agent platform, the Directory Facilitator (DF) Persistence supports the traceability of agent services. The BDI paradigm allows to quickly respond to changes in a dynamic environment: new data from sensors result in the updating of the belief base allowing the VCA to infer new beliefs and new desires. Thus goals are established and plans to achieve them scheduled, using the knowledge base rules and backtracking mechanisms to select the best plan for the current state resulting from the belief base.
Of course, the ideal test to fully validate the proposed approach should
include an AAL scenario with sensors and devices deployed in a real home
environment. In such a scenario the analysis of the e ectiveness of the system could
be carried out using participatory evaluations, as in [
We described a multi-agent architecture to model an expert system for AAL. The system, named Virtual Carer, has to deal with a dynamic environment, monitoring the health conditions of an elderly person. To achieve these tasks the system should model a distributed, reliable and modular sensor network including also those devices typical of home automation. The core of the proposed system is a BDI Agent, the Virtual Carer Agent (VCA) and it represents the reasoning module of the Virtual Carer.
The agents modelling various sensors and devices collaborate to simplify the daily activities of the assisted person and to trigger alarms when something is wrong with parameters regarding his health conditions. The multi-agent approach allows to distribute the agents and the platform over a network. The adoption of frameworks such as JADE guarantees: { the communication through FIPA-ACL messages, ensuring the exibility of the system. New devices can be added simply integrating agents in the proposed platform. { Recovery techniques (such as the DF Persistence or the MCRS) to achieve fault tolerance goals.
We tested the proposed architecture considering several simulation scenarios and representing the assisted person using another BDI agent. Several plans were taken into account, modelling di erent actions performed by the Elderly Agent (e.g. movements, direct requests, changing of health parameters). These simulations con rm that MAS and BDI paradigm improve modularity and adaptability of AAL and AmI systems. The multi-agent approach guarantees the modularity of the system, whilst the BDI paradigm permits to respond to changes in the environment and in the needs of the assisted person. Thus this work con rm the crucial role of Arti cial Intelligence to ensure a better quality of life in the modern society.
As future work more insightful tests have to be conducted: a real AAL
scenario, i.e. a daily-used home environment equipped with devices and sensors,
where an assisted person can live, is beyond our possibilities at this stage of
the work, but is the only way to fully validate the proposed approach.
Several technologies could be integrated in the Virtual Carer system: for example
those for video surveillance described in [
However the direction that this work seems to undertake is that one of an interface between the medical personnel and the patient, being a Virtual Nurse, i.e. a virtual hospital ward at the patient's home environment. Such a system should act as a lter of information for the medical sta , monitoring the health parameters of the person on the basis of his disease and providing only that information considered signi cant. The result could be an earlier dehospitalization for the patient, following the need of having hospital wards available for emergencies and trying to provide an economically sustainable healthcare.