-Ambient Assisted Living (AAL) systems' main goal is to augment live quality of elderly people, by using ICT based systems. In this paper, we are concerned with the artificial reproduction of a physical environment (i.e. a house) and an elder (i.e. the attended) living in such environment. An agent based social simulation system is used for such purpose. Such simulator will allow the integration of ubiquitous computing appliances, services and applications in such environment. A realistic reproduction of human behaviour in the simulator helps, in this context, in the validation of silent monitorisation, diagnosis and action based applications. Proofs are given in the paper which demonstrate the level of reality reached by comparing the artificial behaviour with real ones. Index Terms-Ubiquitous computing, Ambient Assisted Living, behaviour simulation, user modelling.
The main thesis presented in this paper is the following:
agent based social simulation (ABSS) [
A main difficulty one may find in the development process
of an AmI system is that of testing and validation. Testing is
the process of executing a program with the intent of finding
errors in the code [
The proposal is articulated by means of a methodological
work. Such a methodology is a set of procedures which
guides the developer in the definition, creation, testing and
validation of the AmI system. It is based on a methodology
previously described by Gilbert et al. [
The rest of the paper is structured as follows. Section II introduces the methodology used for the engineering of AmI services. Section III introduces the computational models employed to artificially reproduce the behaviour of the attended. Such behaviour is based on a probabilistic and hierarchical automata which governs the activity and location of the modeled elderly in each instant of time. In section IV, the validation approach is presented. It is based on the statistical contrast of artificial data traces obtained by simulating the automata just mentioned with similar traces coming from real users in similar context.
II. AVA, AN AGENT BASED METHODOLOGY FOR THE
VALIDATION OF AMI SYSTEMS
This section explains an agent based methodology for the
validation of AmI systems called AVA. This methodology
proposes the development of ABSS in order to validate AmI
applications. AVA is an extension of the methodology
promulgated by Gilbert et al. [
Gilbert et al. [
The development of an AmI system model is performed
in step 5. The model design associates the real system with
a representation of this system (the model) [
Step 6 deals with the implementation of the AmI system in
a simulation language. The implementation from the concepts
of the model is not a trivial task [
After building the simulation, this must be executed (step 7). Quick, cheap and numerous experiments can be performed thanks to the ABSS. These executions produce large amounts of data regarding the behaviour of users, environment and application models. Forensic analysis, step 8, is an offline analysis to be conducted on the data stored from the previous step. The analysis should consider whether the AmI application functionality is correct. Furthermore, this step must validate that the behaviour of users and environment models is consistent with the observed reality (steps 2 and 3). Without this validation, the theories generated from simulations have no relation to reality (as they are based on non-descriptive models). Therefore, the functionality of the AmI application model is linked to the users and environment models. Section IV deals with the validation of the users model for an AAL system
One of the innovative points in the AVA methodology is the use of simple simulations as a means to validate descriptive simulations. Step 9 checks if any elements of the simulation are too complex. In this case, complexity means that it is difficult to assess whether the behaviour of an element is the expected one. For example, some behaviours of users can be so complex that they need to be evaluated in isolation. An example of this type of behaviour would be the resolution of collisions on the motion of a large number of agents. This behaviour is a problem to be studied itself and its validation would be much more complicated with additional elements in 2OpenABM Consortium website: http://www.openabm.org/ the simulation (more users’ behaviours, a realistic environment model, a realistic model of an AmI application, etc). In these cases, the methodology proposes to consider these complex elements as an object of study itself (step 10) and repeat the AVA methodology for them (step 11). The reuse of models and code in this new iteration will be direct because a more descriptive simulation is available as result of the previous steps. Once the complex element is validated in a simpler simulation, which is the final result of the methodology, the next step for the overall simulation (step 12) can be performed.
Step 12 checks if the developer has found errors in functionality. If that is the case, the AmI application of step 4 must be modified in order to correct these errors and the process repeated. Besides the primary objective (validating the AmI application), is typical to find bugs of previous steps at this point in the form of implementation failures or unrealistic models.
The final decision, step 13, checks if actual elements of the AmI system can be connected to the simulator. The AVA methodology proposes to inject or connect real elements in the simulation progressively in order to make more realistic validations3. This process is called “reality injection” and the basic idea is that real elements can coexist with simulated elements. After connecting real elements, the methodology must be repeated from step 4 to improve the application and the models. The result is an exhaustive validation which is as realistic as possible. The obvious question is why models and simulations are necessary if real elements (as real users) can be injected. The answer is that a model, by definition, is somewhat easier to study than the modelled reality. The purpose of including real elements in simulations is to improve the realism of the models. Then, in subsequent iterations, the real elements will not be included because the pure simulations allow faster and cheaper tests.
Finally, if models are descriptive enough, the bugs found in the functionality of the AmI application model will correspond to failures of functionality in the real AmI application. These failures should have been corrected in each iteration of the AVA methodology. Therefore, the result of the methodology (step 14) is that the AmI target is exhaustively validated.
In this section, the particular models used, in the application of the AVA methodology in the AAL domain, are introduced. In section III.A, it is presented how the physical environment (i.e. the house and furniture) and sensors were defined. Section III.B refers to the production of realistic computational models for elders living in such environment and making sensors to react on their presence. Having such models (i.e. the house, sensors and persons) within a simulation, and its integration with the ubiquitous computing software, such software can be tested.
3Notice that this not involves necessarily a Participatory Multiagent Simulation. The real elements do not have to be humans playing the role of simulation components. These elements can be software applications, hardware or even parts of the environment.
For Multi-Agent Based Simulation (MABS), it is available Ubiksim4, a simulator developed by University of Murcia that works over MASON5. It has integrated an environment modeling tool based on SweetHome3D6, an application conveniently adapted for modelling attended people and their environment. It is possible to create houses over a 2D plane, also it offers a 3D navigable view (figure 2). This view is used for simulation’s visualization at real time.
In the physical environment generated by using the Ubik editor, a simple house (see figure 2) with a kitchen, a bathroom, a bedroom and a living room is modelled. Presence sensors are included in every room of the house. And a sensor for open door (it is necessary for knowing when the elder leaves the home) is also included in the outdoor. When a simulation is run, the person moves in the house and stimulates sensors when he is detected by them. Such sensors, through the ubiquitous computing software, generate events. And these events generate log entries. Such simulated log entries are used afterwards to check if the virtual elder behaves in a realistic manner (see section IV).
Notice that log entries (both in the simulator and the
real setting) are generated by the same monitoring service
which continuously checks if the elder may be suffering some
problem, by using a pattern recognition approach (more details
on this may be found in [
The target of the modelling activity is a typical aged person, who lives independently and alone in his own house. As he lives alone, the following situation may occur: he may suffer some health problem and stay immobilised in the floor for too much time before anybody comes and notices 4UbikSim: http://ubiksim.sourceforge.net, last access: 20 May 2010 5MASON Toolkit: http://cs.gmu.edu/∼eclab/projects/mason/, last access: 20 May 2010
6Sweet Home 3D: http://www.sweethome3d.eu/es/index.jsp, last access: 20
May 2010
that something is wrong. But it is possible to develop an
ubiquitous computing system which detects it and generates
some emergency response process [
The behaviour of simulated people are modelled probabilistically. In this approach, behaviours are defined as situations the agent should play in each moment. Transitions between behaviours are probabilistic. The underlying model is a hierarchical automaton (i.e. in a higher level there is a number of complex behaviours that the agent may play and once it is in a concrete state, within the state there is another automaton with more simple behaviours). So, the modelling of each behaviour is treated separately and the modeller is abstracted of unnecessary details. So, in the lowest level (basic actions), each state is atomic. An agent never conducts two behaviours of the same level simultaneously.
The behaviours used for modelling elders are of three types: • Monotonous behaviours: the kind of behaviour the elder manifest always approximately in the same time slot, and on a daily basis (e.g. sleeping, having meal, medication and so on). • Non monotonous behaviours: the kind of behaviour the elder usually manifest, not bounded to a concrete time slot, and repeated within a non constant period (e.g. going to the toilet, having a shower, cleaning the house and so on). • Any time behaviours: such behaviours will sometimes interrupt others the elder is already doing, and will be generated regardless they were already generated in a temporal proximity (e.g. in his spare time).
A probabilistic automaton [
Notice that, when the elder is at any state, the necessity of changing to another state may arise. But this is not done immediately. Moreover, a number of different changes (i.e. transitions) may be pending simultaneously. Thus, a list of pending tasks (i.e. or events) is maintained. Such tasks are ordered by a static priority (e.g. going to the toilet goes before cleaning).
For generating transitions in real time, probability distribution functions are used according to the type of the behaviour and its features. These distributions are member Q = {a0 = NormalT ime, a1 = MedicationT ime, a2 = MealT ime, a3 = SleepT ime, a4 = Anomalous}
(a)
Q = {a00 = SpareT ime, a10 = MedicationT ime, a20 = MealT ime, a30 = SleepT ime, ax1 = T oiletT ime, a02 = ShowerT ime, a03 = CleanT ime}
(b) Q = {a200 = GoingT oF ridge, a201 = GoingT oCooker, a202 = Cooking, a203 = GoingT oT able, a204 = Eating}
(c)
of the exponential family [
Notice that monotonous behaviours must be activated at specific time hours or within specific time intervals. For example, in the case of MedicationTime, a new necessity of changing to such state will be generated exactly at the time to take medicines. In the case of MealTime and SleepTime, the necessity is generated within a time interval. The distribution that models these transitions is bounded in this time slot. It must be assured that agent eats and sleeps every day, because of that, if not transition is generated into the time slot, a transition is generated at the end time.
To provide more realism, an automata hierarchy is introduced representing each state by lower level automata which define more specialized behaviours.
As shown in figure 3(a), in level 0 the initial state is a0 =NormalTime. For every one of the other states {a1 =MedicationTime,a2 =MealTime,a3 =SleepTime} exist a list of times generated by a probability function. When a time counter arrives to one of these times, a transition to state owner of the list is added to pending tasks list. When the action is finished, the automaton returns to initial state if there is not another pending task. The final state is a4 =Anomalous, if it is reached, the execution will be stopped.
In level 1 (figure 3(b)) non monotonous behaviours are represented. There is an initial state in every refinement where the person does the main task of the upper level (eating, sleeping or taking medication). The state ax1 =ToiletTime is considered in every refinement of level 0. However, the states a02 =ShowerTime and a03 =CleanTime only may be activated in normal state of level 0.
In the lowest level (level 2), the new automata define some specialized actions refining every state of level 1. These new states do not give relevant information, but the agent gains more realistic behaviours.
With the sequence of actions in figure 3(c) the person cooks before eating. It refines state a20 of level 1, the agent is not going to be static in the kitchen, because it must be going to different places and spends some time in every state.
In figure 4 a base implementation for agent’s behaviour is presented. First of all, all possible automata’s states are iterated for initializing lists which contain time instants (lines 13-24). These time instants are generated by a probability distribution function, and they represent when a transition to its associated state is going to be launched. After of that, the automaton begins to run. At every time instant, if actual time is equal than first item of a time list, a transition is added to pending tasks list (lines 57-60). When there is one state with higher priority than actual state in pending tasks list, a transition is also generated. Then, if actual state is unfinished, it is stored as a pending task and a new state is reached (lines 46-50). When this new state is finished, leaved state may be resumed.
The configuration parameters for the whole simulation involve: • Temporal limit in every room before entering in anomalous state (tmax) • The probability distribution parameters according to the kind of behaviour • Time slots of routinary temporal behaviours, like eating or sleeping (inis, ends, s ∈ M ealT ime, SleepT ime)
From section II, it is clear that user modelling is a means for testing AmI services and applications without a real environment. This task is performed in the third step of the AVA methodology. Regarding to the validation, other important steps within AVA are the steps from 5 to 8. Model validation is needed in order to show that models are able to describe the users’ behaviours. The rest of the section shows how the model validation has been approached. But basically, activity data from real users is compared (in statistical form) with the same type of data produced by the artificial models.
Activity data from real users were obtained within a pilot project devoted to the validation of the Necesity system. Around 25 users, all elderly people living independently, were used in such Pilot project. In this paper, data coming from three users, under monitorisation during two months, were used. Data is in the form of a Necesity log. The logs offer the possibility to represent where the user is, at any moment, at the house (including also if he leaves the house or he is seated or sleeping). Thus, three different data sets, with log entries corresponding to sensor events are available. These data sets need further processing.
Validating the artificial models is assuring that the right probability distribution function is used to reproduce the transition between the different states (i.e. behaviours). Thus, data series are obtained from log data as they were a random number series generated by the corresponding probability distribution. Such series are used afterwards in a goodness of fit test. For example, in case of the non monotonous behaviours and anytime behaviours, preprocessing the log data involves the extraction of the time series of the moments in which each event is produced. These time series only include values within the typical awake period of the corresponding user. Obviously, while the user are sleeping, his daytime routines change.
The data preprocessing for the monotonous behaviours is slightly different. This kind of behaviour is usually produced in a bounded time slots. For example, having dinner, having lunch or sleeping are behaviour which occur during specific time periods of the day. So, the preprocessing involves the extraction of the behaviours events inside the time slots. The time slots can be slightly different for each person, but it is possible to define an approximation of them which will be valid for all (see in section III the configuration parameters). Finally, in each slot time, the time intervals between each ocurred event are measured. The extracted time series are composed of these time intervals.
Validation is one of the most important issues in a
simulation system. Validation consists in the determination that
the simulated model is an acceptable representation of the
real system, for the particular objectives of the model [
The models which describe social processes, as the model
proposed here, are generally hard to validate. In this approach,
the behaviour is probabilistically modelled. However, some
statistic tests should be done to assume that a probabilistic
model is reasonable to explain the data. This process is called
model diagnosis. And this section is devoted to make the
diagnosis of the models presented above. The most serious
problem that one usually faces in this kind of validation is the
lack of real data [
From these preprocessed data, some histograms for different behaviours and people are shown in Fig. 5. The sample density is shown with a black line. The dashed line shows the probability density function of the theoretical distribution that models that behaviour.
Graphs 5 (a) and (b) show two monotonous behaviours, sleeping and having dinner (having lunch is similar). In these kinds of behaviours the event that raises the behaviour occurs (a) (b) (c) (d) Fig. 5. (a) Minutes between 21:00 hours and the instant the attended C goes to bed, (b) Time between dinners for attended B, (c) Time between uses of the toilet for attended A, (d) Time between spare time for attended A during a time interval. The interval is usually the same for each attended person, i.e., a person usually goes to sleep or to have dinner at the same hours. Because of this, the curve of the behaviour can be fitted to a gamma distribution. The gamma distribution is usually employed as a probability model for waiting times, in this case, the waiting time until the next event of the behaviour. So, a gamma distribution is suitable for modelling monotonous behaviours.
A kind of non monotonous behaviour, i.e. going to the toilet, is shown in Fig. 5 (c). In this case, the behaviour is fitted with an exponential distribution. Non monotonous behaviours are also waiting time models, but they are defined in wake periods. This is a special case of the gamma distribution which can be modelled with an exponential distribution.
Finally, the group of anytime behaviours are behaviours which will be often interrupted for the other behaviours. Because of this, size of intervals in an anytime behaviour are small due the interruptions. This causes the characteristic curved heavy-tailed distribution, specifically, the Pareto II distribution, also known as Lomax distribution.
Gamma, exponential and Lomax distributions are members
of the exponential family of probability distributions. Actually,
they are special cases of the beta prime distribution (also
known as beta distribution of the second kind, Beta II). The
Beta II distribution nests many important distributions as the
gamma, the exponential or the Lomax distribution. The gamma
and the Lomax distributions are special cases of Beta II. On the
other hand, the exponential distribution is a special case of the
gamma distribution Γ(α, β) when α = 1, and a special case
of the Lomax distribution with some restrictions [
Now, in the rest of the section, empirical evidences which support using gamma, exponential and Lomax for monotonous, non monotonous and anytime behaviours are given.
Notice that it is possible to estimate the distance between
time series generated, as explained above, from real log
data and time series generated by simulation of theoretical
distributions. Such estimation is done by using some statistical
test, like the Kolmogorov-Smirnov (K-S) test [
The K-S test and the chi-square test are the most commonly
used and for large size sample both tests have the same power.
However, the chi-square test requires a sufficient sample size
in order to obtain a valid chi-square approximation [
The K-S is a goodness-of-fit test to indicate whether it is reasonable or not to assume that a random sample comes from a specific distribution. It is a form of hypothesis testing where the null hypothesis says that sample data follows the stated distribution. The hypothesis regarding the distributional form is rejected if the test statistic, Dn, is greater than the critical value obtained from a table, or, which is the same, if the pvalue is lower than the significance level. The significance level is fixed in this work at 0.05, which it is the value usually referred in statistical literature.
Table I shows the p-values obtained from the K-S test for each validated behaviour with the adequate distribution. The null hypothesis is that the behaviour sample data come from the stated distribution and it is rejected if p-value is lower than the significance level.
Person
A B C
From these results, none of the stated null hypothesis can be rejected. Therefore behaviour of the attended people could be fitted by the specified distribution, as it is described above.
Various approaches have been proposed to create
autonomous characters. For example, in [
The approach to behavioural autonomy presented in section
III is based in [
The decision of representing the automaton’s transitions
with probabilities instead of using a vector of strengths is
based in [
In [
VI. CONCLUSION
In this work, a general behaviour model for different people is proposed. Adjusting the model to specific persons would imply using the suitable configuration parameters for the corresponding probability distributions governing transitions between states of the probabilistic automata. This process is part of a more general task which is testing AmI services and applications. For such purpose, the AVA methodology was presented. It has been applied for the validation of an AAL system called Necesity. More specifically, this paper is focused in producing models of humans (i.e. step 3 of AVA) and validation of the models (steps form 5 to 8).
Such an approach requires, as a means to validate the artificial behaviours, for all the artificial models, a method to check if gamma, exponential and Lomax distributions are well suited. Using the K-S test is possible to quantify the distance between the empirical distribution functions of two samples. Then the K-S test is used to validate the behaviours of the artificial models, by using real data of real elders. A high level of the p-value (higher that the significance level) means that the behaviours are drawn from the same distribution (the null hypothesis is not rejected). Notice that, in most cases, the obtained p-values are higher than the significance level and the null hypothesis is not rejected. But, it is also necessary to remark that a good fitting of the configuration parameters will always be required. And this is due to inevitable heterogeneity in behaviour of people (including elders). As a conclusion, it can be said that the proposed model is suitable for modeling probabilistically the behaviour of simulated people, as the work states.
Future works include a deep study of source data in order to generate a taxonomy of elders in terms of their behaviour within their houses, depending on mobility and habits. Such a taxonomy would be useful for an automatic parameter tuning of the models of elders. The user of the simulator, instead of configuring parameters by hand, would simply choose between a catalogue of elderly people patterns of behaviour.