In this paper we discuss the possibility to extend PORTIA, a persuasion system currently applied in human-agent dialogs, to support ambient persuasion. We have identified a fitness center as an appropriate smart environment in which ambient persuasion strategies can be applied. According to the Ubiquitous Computing vision, in the fitness center the user is surrounded by several connected devices that cooperate in the persuasion process, each of them with the most appropriate strategy, mode of persuasion, style of communication and ability of exploiting the kairos principle. To this aim we propose a multi-agent system able to support this distributed and intelligent approach to persuasion that allows to follow the user during the gradual change from the initial attitude to sustain of long term behaviours.
As stressed in Stock et al. [
In this paper, we present an approach to ambient persuasion [
Ambient Intelligence solutions may provide a great opportunity for achieving the
aim of distributing and embedding persuasion and coaching strategies into the
environments that the user attends, according to the Ubiquitous and Pervasive
Computing vision [
Changing habits in the context of well-being is influenced by several -rational and
emotional- factors depending on the context, that can be intended as: ‘What the user is
doing, Where is the user, With whom, When’ [
To this aim, we propose a multi-agent architecture which includes different types
of agents: (i) Sensor Agents –used in order to provide information about sensors
parameters and context features (i.e. temperature, heart rate, humidity, presence of the
user in a room, etc..); (ii) Device agents -typical of the environment- that manage the
active devices in the environment (e.g. cardio fitness machines, public displays,
mirrors, etc.) and convey to the user the training according to the context and the aim
of the environment; (iii) D-Me agents [
In order to show how this architecture works, we will consider a fitness center as a
suitable place to test the approach. In fact, a fitness center is equipped with enough
technology for simulating a smart environment, the users are already confident with
the technology during their workout and, moreover, most of them want to be
constantly motivated in order to reach their goals concerning an healthier lifestyle
[
The paper is structured as follows. In Section 2 we discuss the relation between ambient persuasion and wellbeing. Section 3 describes the proposed architecture of the system. Then, in Section 4, we illustrate a scenario example that is used to show the functioning of such a system. A final discussion and future work directions are reported in Section 5.
Persuasion is a form of social influence and a ubiquitous part of contemporary life. It
is a relatively new trend in the research community that shows a growing interest also
into intelligent information technologies, and for better or for worse, persuasive
technologies are already part of the everyday technological landscape (see examples
in [
Currently, the most common persuasive systems used in fitness centers (at least in Italy) employ the feedback mechanism to show to users effects of the exercise (i.e. Polar Cardio or Cardio Fitness machines and so on) and are generally isolated without the ability to communicate and cooperate with other devices in order to achieve the common goal of taking care of the user. In addition, a fitness center has professionals responsible for this purpose, as personal trainers and wellness coaches: they have a very important role in helping the user to change their habits and find the motivations to work hard for achieving their goals. However, beside that they can be expensive or unavailable when users need them, many people feel shame and fear of being judged by their human coach: sometimes this can be a motivation for changing attitude, sometimes it may compromise the success of the coaching strategy, increasing the user's attitude at overcoming barriers -especially emotional- and decreasing selfesteem. Several coaching systems have been implemented on mobile devices (see for example, My Weight Loss Coach for Nintendo DS, Nokia Fitness Coach for Nokia phones and the so many sport trackers like Endomondo for the most popular mobile platforms, or CardioTrainer for Android, or Sports-Tracker for Nokia) aimed at monitoring, supporting and tracking users’ progress and improving their energy balance. Again, in many of them the user has to input data about her workout, eating behaviour, etc. On the contrary, in other systems, as Nintendo Wii fit, My Body Coach by BigBen Interactive, or Your Shape: Fitness Evolved, the new edition of Ubisoft's training software for Microsoft console, the user is monitored and motivated during the exercise even though these are not integrated with other daily activities and situations of the user.
According to [
Again, an intelligent environment is a social place and therefore, people may share
personal experiences with others that have the same problems, goals, needs [
Finally, the system should be perceived not has having a pure functional intelligence but has being an emotionally and socially intelligent actor that may monitor the user and intervene appropriately at the right moment.
In the light of these premises, we present an agent-based system that tries to apply the principles of ambient persuasion in a smart fitness center. 3
According to the ambient persuasion model proposed by Kaptein et al. [
In order to generate the most appropriate persuasive message to the user, we extended PORTIA and used its reasoning and argumentation model. As far as the sustain phase it is necessary to reason on which motivational arguments have to be adopt for continuing to motivate the user according to the situation. To this aim, we started an empirical study aiming at exploiting the knowledge and rules that human personal trainer and fitness professionals use.
Before illustrating the architecture and the functioning of the system, let us introduce a brief overview of PORTIA. 3.1
PORTIA is a user-adapted persuasion system capable of simulating the persuasion
process used by humans to convince someone to perform a given action. In this paper
we provide a brief overview of the system. For a more detailed description of
PORTIA, please refer to [
PORTIA considers three knowledge bases: the User Model, the Persuasion and the Argumentation Knowledge Bases.
The User Model. Understanding the presumed weight of user’s goals is crucial to select the most promising persuasion strategy in a given context. User Model is employed to reason about the user’s presumed characteristics. Rather than acquiring this knowledge through direct questions, PORTIA attempts to implicitly infer it, with some level of uncertainty, from information about user’s personality traits and living habits. The User Model includes a specific knowledge and a general knowledge component. The former collects facts about the user (evidence).The second represents criteria to infer the user’s goals and abilities under conditions of uncertainty in the form of Elementary Belief Networks (EBNs) that are belief networks with only one leaf node representing uncertain implications. In particular, user’s rational and emotional goals can be inferred respectively from knowledge about user’s habits and personality traits.
The Persuasion Knowledge Base is employed to model rational and emotional
strategies. The Persuasion model is defined in term of goals and beliefs from the
Persuader’s perspective that may employ rational as well as emotional strategies (but
also a mixture of them) to induce the user to perform a given action. Persuasion
strategies are represented with EBNs too. In particular, emotions may be introduced
in the persuasion process in two forms: by selecting an emotional goal or by
activating, through arousal of user’s emotion, an intermediate goal which is
instrumental to the final one. The PORTIA’s persuasion strategies are summarized in
Table 1. For more details see [
PORTIA’s Persuasion KB
General induction of intentions [(VGoal U gi)∧(AGoal U gi)∧ (Bel U Implies(a,gi)) ∧(Bel U CanDo(U,a)) ] →? (Int U Do(U a)) [i] It may be summarized as follow: “If User has the goal g (VGoal U g) and it is really relevant at this time (AGoal U g) and he believes that doing the action a implies achieving g in a more or less near future (Bel U Implies(a,gi)), and he believes that has the ability to do a (Bel U CanDo(U,a)), then probably user intends to do a (Int U Do(U a))” (from Miceli et al, 2006).
Rational induction of intention Emotional induction of intention gi ∊ {Rational goal set} gi ∊ {Emotional goal set} It focuses on rational goals like ‘to be in good health’, It focuses on rational goals like ‘to make friends’, to be ‘to have a good appearance’, and so on. in good mood’, and so on.
Activation of goal strategy Activation through a belief or an emotion of an intermediate goal which is instrumental to the user’s goal. It considers two possible applications: Rational Activation strategy or Emotional one.
Induction of beliefs Argumentation about means-end implication. It represents the action-goal relation.
Appeal to Expert Opinion
Appeal to Popular Opinion
Appeal to Position to Know
Appeal to Friendly Personal Experience
Others
The Argumentation Knowledge Base is employed to translate each strategy into an
argument. Items to include in the argument correspond to the variables associated
with nodes of EBNs, and the way these items are combined in the message (order in
which to present them and relationships among the various parts) is represented into
Elementary Argumentation Plans (EAPs) that are a coherent translation of EBNs.
EAPs are built on two theoretical grounds: Walton’s Argumentation Schemes [
PORTIA considers two main modules: the Reasoning module (REASONER), and the Argumentation one (ARGUER).In the Reasoning module, PORTIA exploits the information about the user (User Model KB), computes the degree of importance of the various -rational and emotional-goals on which focus the persuasion strategy, and evaluates the persuasiveness of different combination of strategies (Persuasion Strategies KB) and selects the most promising one with respect to the goal of inducing in the user the intention to do a certain action. For this purpose, PORTIA builds a complex Belief Network (BN) by dynamically chaining forward several EBNs. The BN is a representation of the user’s mental state that enables to apply a “what-if” reasoning form for evaluating the persuasive power of the strategies, and to select the most promising one.
In the Argumentation module, PORTIA has to construct the arguments to express the strategy selected in the previous step. To this aim, PORTIA explores the complex Belief Network and decides the items to mention, their presentation order and the rhetorical relations among them. Also, she has to decide whether to include an appeal to cognitive consistency (between the user’s of goals and beliefs, and his behaviour) as a form of encouragement to adopt a more consistent behaviour. That is, PORTIA has to translate the complex Belief Network into a coherent discourse plan. The discourse plan is dynamically built by combining the elementary argumentation plans (Argumentation Plans KB) that represent the elementary beliefs networks included in the Belief Network. The discourse plan is then translated into a natural language message used as an attempt to persuade the user. 3.2
In order to develop a system for ambient persuasion in the context described in the
Introduction, we propose an extension of a multi-agent platform implemented in
another project [
It is worth noting that the architecture that we propose has not been conceived with the sole purpose of persuading the user, but it aims at implementing smart environments that aims at improving the quality of life of the user. In these environments all agents exchange data and information in order to provide services (recommendations, information, motivation) suitable for helping users in achieving their goals.
For instance, the D-Me agent, by monitoring the user behaviour, knows about his/her meals, and, through social networks, may know who are his/her friends, etc. Then, it may communicate this data to the Coach Agent that may adopt the optimal persuasion strategy and arguments accordingly. Again, suppose that the doctor recommended to the user to loose weight. When the user goes to the food shop with the intention to buy a sweet cream, the D-Me agent may communicate this information to the Coach Agent that may act to persuade him not to do it by adopting the most effective persuasion strategy and arguments.
For this reason, one might argue: why not delegating the entire ambient persuasion process to the D-Me agent? The idea is to build a platform environment-independent that may be applied in a fitness center as well as in a virtual home and, possibly, enabling interaction between different environments so as to support the user at different times of the day. According to this perspective, we believe that D-Me agent should not have specific knowledge of the environment or the technological devices because persuasion strategies and arguments used by a Personal-Coach in the fitness center are probably different from those applied by Personal-Butler in the smart home. 4.
The following example is a simulation of the system’s behaviour in a typical scenario of the gym environment. Let us consider the following starting conditions.
Robert is a man below 40 years who regularly makes medical check-ups. He is a hypochondriac, too. He is probably an extravert because he feels comfortable around people. The doctor suggested him to make some physical activity regularly. He decides to go to the gym but he is quite sceptical about this. When Robert enters in the gym his D-Me agent has the permission to communicate to the other agents in the environment the anagraphical and physiometric data and other information about the user social network. Robert registers himself to the gym information system. Then, the system provides a personalized workout schema to Robert and, in order to persuade him to adopt the proposed workout, the Coach agent generates and communicates the following persuasion message, according to the BN in Figure 2: “Hi Robert, I am your personal coach and this is your personalized workout schema. You should do it because I know that you care for your health and training has a lot of benefits on your health. In fact, the World Health Organization says that this is very important for health and it is a specialized agency of the United Nations that acts as a coordinating authority on international public health, it is an authoritative voice. In addition working out may be a great opportunity to know new friends. In fact, this is well known to all who attend gyms and there is no evidence against it. Come on! I’m sure you can do it if you wish”.
Fig. 2. The BN used by the Coach agent to simulate the effect of selected strategy on the user’s mental state.
Figure 2 represents the result of the reasoning process of the persuasion model applied by the coach agent when tries to persuade the user to adopt the proposed workout. That is, the coach’s representation of the user’s mental state on which has been tested the effectiveness of the persuasion strategy. Reasoning component has propagated the user’s evidence received by the D-Me agent into the EBN-KB and has inferred that, although ‘to be in good health’ is the presumed most important goal to Robert, the associated rational persuasion strategy does not seem to induce in the user the desired level of intention to do the proposed workout schema and a mixed strategy could be more effective. Therefore, Coach agent selects the goal with the highest value among the emotional goals and infers that the two candidate goals on which the persuasion strategy focuses are the rational goal to be in good health and the emotional goal to make friends. Moreover, it selects the belief induction strategies through appeal to expert opinion and appeal to popular opinion, respectively, for the goals of being in good health and making friends.
In a dialog perspective, persuasion, rather then a predefined, integrated set of
propositions, is seen as a sequence of moves in which two parties (Persuader and
Receiver) are reasoning together on some argument [
For example, let’s suppose that Robert is hypochondriac and he does not want to do the proposed workout schema because he is afraid of getting sick. The Coach agent classifies this reaction as an object on the user’s capability and, after reasoning on the BN in order to select the most appropriate response, tries to reassure him because all the tests say he's fine: “I know you're scared but try to stay calm because your medical certificate says you're fine!” Finally Robert goes to the bike and began his training.
Another strategy applied by the Coach agent is to motivate and sustain the user during the exercises. Let’s suppose again that, despite the encouragement received, Robert starts going very slowly on the bike. Then the Device agent that monitor the user’s exercises, communicates to the Coach agent the new situation of Robert. Then, the Coach tries again to reassure the user by conveying a following message on the display of the bike device: “Robert, come on, do not worry, I'm tracking your heart and, at present, you are even below the threshold workout. Therefore, you can do more without fear of forcing your heart ... I monitor you, do not worry”.
Robert starts to push a little more and retakes the right pace.
Then, finally, the user finishes the exercise. The Device agent transmit this information to the Coach that tells Robert that everything went well but that, the next time, he could do more. 5
This contribution shows a preliminary work towards the development of a system for ambient persuasion in a fitness center based on a multiagent architecture. In the proposed architecture there are two agents that are central to the persuasion process. The D-Me agent, which manages the personal user profile, may transmit to the intelligent environment data about the user that may be important for adapting the persuasion strategy and the motivational messages. The Coach agent, which acts as a personal trainer, has the role of persuading the user to train and adopt a certain workout and also to sustain this behaviour during training. The Coach agent uses the reasoning and argumentation model of PORTIA for generating the persuasive message. While, for the generation of motivational messages to be provided in the sustain phase, we are collecting data from professionals expert in the fitness and wellness domain, such as personal trainers.
We are aware that in this domain the risk of producing a message that is not
appropriate to the situation because of an inferential error on the user goal, personality
traits, and so on may determine the selection of a wrong strategy or arguments and
consequently may cause distrust in the user. In this case it is necessary to endow the
environment with a formal model of trust [
At moment, in order to test the effectiveness of the proposed system we used the knowledge of two personal trainers. To these people we proposed some scenarios, like the one in Section 4 of this paper, with the aim of collecting examples of motivational sentences, arguments to be used to motivate the users. Now we are conducting an experimental study that involves a greater number of experts in the fitness domain aiming at understanding: - which are the features of the user relevant for adapting the motivational message; - when to intervene with a motivational message; - which are the strategies, at the reasoning and argumentation levels, most widely used according to the user features.
At present, the collected data give us some useful information for understanding how human personal trainers build in their mind the models of their clients and which are the features of the clients that influence their decision about how to motivate them.
For instance, the gender of the client seems to be important for choosing the arguments to use. Personality traits (mainly the levels of sociability and extraversion) influence the message style. While the cultural background, the age and the profession of the client influence the argumentation schema to be used to support some concepts and claims.
Moreover, from this initial analysis, seems clear that personal trainers initially classify their clients into stereotypes (Lazy, Super, Model, Normal, Sociable, …) that help in deciding how to motivate them initially. This capability is related to the level of experience of the trainer.
Then, in our future work we plan to analyze the collected data and build the initial knowledge of the coach agent relative to stereotypes and reasoning rules in order to generate motivational messages appropriate to the user and to the situation. Moreover, we intend to give to our coach the capability to learn from the user feedback in order to refine the rules driving the choice of the optimal strategy.