A SOCIAL ROBOT FOR FACILITATING HUMAN RELATIONS

IN SMART ENVIRONMENTS

Berardina De Carolis, Nicole Novielli, Irene Mazzotta and Sebastiano Pizzutilo

Dipartimento di Informatica, Università degli Studi di Bari “Aldo Moro”, Bari, Italy

Keywords: Social robots, Social facilitator, Social network analysis.

Abstract: This paper describes how a robot may use social network analysis measures for facilitating social relations

when acting as a “host” in a smart environment. The robot’s main goal consists in welcoming people,

facilitating contacts and information sharing among people present in the environment. It uses knowledge

on the structure of the social network for selecting the most appropriate strategy to create new relations or to

spread information in the most effective way. To this aim, a multiagent system has been implemented for

simulating and evaluating the functioning of the social facilitator.

1 INTRODUCTION

According to several research studies, Ambient

Intelligence (AmI) should be used not only for

providing task-oriented services but it should

include also social-oriented assistance in order to

facilitate human contacts, information broadcasting

and sharing, and to enhance community-oriented

services (Shadbolt, 2003). We developed a

multiagent system that aims at facilitating the

communication, interaction and sharing between

people in a Social Network (SN) located in the same

physical place (e.g. recreation places for elderly

people, meeting rooms, parties, fitness centers). As

interface, we employ a robot acting as Social

Facilitator (SF) among people present in the

environment and between them and the environment

services: social robots, in fact, can provide an

appropriate medium for facilitating relations in

public spaces due to their presence (Sakamoto et al.,

2007). In particular the SF can be seen as a “host”

that welcomes and guides people in the

environment, in order to facilitate the establishment

of social relations between people that do not know

each other, and to favour information spreading and

sharing among peers. To this aim, the SF has to

exhibit social competencies and, therefore, it has to

understand the existing relations among peers, using

Social Network Analysis (SNA) (Scott, 1991).

The system has been tested through a simulation

of a “recreation center for elder people”. Results

obtained so far show that the adopted approach is

successful for improving social relations in the

place, integrating isolated people and connecting

subgroups.

The paper is structured as follows: Section 2

presents the architecture of the system and provides

an example of application. A framework for the

system evaluation, conclusions and future work

directions are discussed in Section 3.

2 SYSTEM ARCHITECTURE

The system has been implemented using JADE

(

http://jade.tilab.com/) and JUNG (http://jung.source

forge.net/index.html

) and it is composed by: Peer

Agents, representing people in the smart

environment, the SNA Agent, handling information

about the SN, the Social Facilitator (SF), which has

the main goal of increasing the cohesion of the

social network, the Sniffer, which observes

exchanges among peers. Let’s see the structure of

these agents in more details.

Peer Agent: represents a person in the smart

environment. It maintains dynamically the model of

the interests, preferences, friendship, liking and

disliking relations of the person it represents. The

profile is updated dynamically according to the

creation of new relations among peers. It is

structured as follows: Id of the peer in the network,

Personal data (name, email, age, profession),

Relations with other people in the environment,

380

De Carolis B., Novielli N., Mazzotta I. and Pizzutilo S..

A SOCIAL ROBOT FOR FACILITATING HUMAN RELATIONS IN SMART ENVIRONMENTS.

DOI: 10.5220/0003142403800384

In Proceedings of the 3rd International Conference on Agents and Artiﬁcial Intelligence (ICAART-2011), pages 380-384

ISBN: 978-989-8425-41-6

 2011 SCITEPRESS (Science and Technology Publications, Lda.)

Liking and disliking of other members, Spoken

language (used a peer when communicating),

Known languages (the other languages that the peer

knows), Disabilities in hearing, seeing, speaking and

moving, Interests (e.g. hobbies, recreational

activities, and so on, expressed as a confidence

level), Preferences (e.g. food, movies, etc.), Privacy

issues (the profile can be made accessible to nobody,

to the environment, to everybody or only to friends).

To test our system in a real environment we have

associated a RFID to each person that enters in the

place. In this way we can dynamically monitor the

presence of that person in the environment and

activate the corresponding peer agent. At the

moment the member’s profile is acquired explicitly

through an interface when the RFID is associated to

a person.

The SNA Agent monitors the evolution of the

network by gathering information about each

member. It uses this information to compute

measures useful for understanding the various roles

and groupings in the network, e.g. who is the leader,

which agents act as connectors, which are the

isolated peers and so on. The main behaviours of

this agent are:

a. Information Gathering: in this phase the SNA

collects information about each member of the SN

from the corresponding Peer agent. Moreover, for

debugging purposes, it provides a graphical

visualization of the current state of the SN.

b. SN Elaboration: it calculates the requested

sociometric measures (Haythornthwaite, 1996;

Scott, 1991). In the proposed application context the

following measures are considered:

 Density indicates the degree of cohesion of the

network. It allows verifying that SF actions are

actually improving social relations in the

community;

 Connectivity of a node expresses the number of

nodes that should be deleted from the network to

disconnect two persons. It can be used to find the

highly popular members that can help to foster

information spreading or the integration of isolated

peers in the network;

 Geodetic distance represents the shortest path

between two nodes. It can be used by the SF to put

in contact an isolated peer with the closest network

member with which he shares some interests;

 Centrality can be used to evaluate the importance

of each member of the network (i.e. the leader of a

group). Centrality can be specialized as: Degree

centrality, to measure the number of direct

connections; Closeness centrality, to identify the

shortest path to other nodes; Betweenness centrality,

to identify intermediary members between important

portions of the network. For example, people with

high closeness centrality are very effective for

spreading and monitoring information flow while

members with high betweenness centrality can be

seen as good brokers and therefore have a great

influence on information flow between subgroups;



Similarity among members of the network can be

used for integrating isolated peers. Among the

several similarity measures that can be used, we

decided to employ the Pearson correlation

coefficient (Rodgers and Nicewander, 1988), taking

into account the profession and the interests present

in the member profile;

 Clique Analysis allows identifying subgroups in

the SN. It can be used for connecting subgroups.

c. Communication: the SNA agent exchanges

messages with the other agents and in particular with

the SF about the overall SN situation.

Social Facilitator Agent: it acts as an intermediary

among the network members. Its main goal is to

select communication forms and artefacts according

to the situation of the social network. It has been

modelled and implemented as a BDI agent (Rao and

Georgeff, 1991), whose beliefs are facts about the

state of the SN. Beliefs are represented as BEL SF

u, where u is a fact concerning:

 one of the measures provided by the SNA agent

(e.g. Density(n, 0,76) represents the fact that the

density of the network n is 0,76);

 friendship relations among network members

(e.g. Friend(u

) indicates that there is a binary

friendly relation between the members u

and u

);

 dislike relations (e.g. Dislike(u

) indicates that

there is a unidirectional relation indicating that u

dislikes u

);

 predicates about members (e.g. Predicate(u

,z)

where z is a value or a fact; for instance Is-

Interested(u

,fishing)).

The SF Goals are of two types: Persistent Goals (P-

GOALs) denote the agent’s nature and mission, and

guide its reasoning while Contingent Goals are

triggered by the situation.

The SF has the P-GOALs of taking care of

people present in the environment and increasing the

number of relations among members in the

environment. The first goal may be formalized as:

(P-

GOAL SF (BEL SF (welcomed(u

)))) AND (P-GOAL SF (BEL SF

(confortable(u

)))) AND (P-GOAL SF (BEL SF

(NOT(isolated(u

))))) where u

represents a member of

the SN. The second goal may formalized as the

A SOCIAL ROBOT FOR FACILITATING HUMAN RELATIONS IN SMART ENVIRONMENTS

381

achievement of the maximum density of the SN,

(P-

GOAL SF (BEL SF (Density(net,max)))). This goal can be

achieved by modifying the initial state of the SN

until all members are connected. However reaching

a density equal to 1 may not be always possible

since, for instance, there can be members that really

dislike each other or that do not have anything in

common or that pursue different goals. The SF will

abandon a persistent goal when it has been achieved

or when it believes it is not possible to achieve it.

Contingent goals are triggered by contextual

needs (e.g. satisfying a request of a member or a

request of the environment of spreading important

information as quickly as possible, solving conflicts,

etc.). To achieve these goals, the SF executes

conditional plans stored in a library (Cavalluzzi et

al., 2003). At this stage of the project we have

defined plans for the three contingent goals: a)

integrating isolated peers, b) connecting subgroups,

c) spreading information.

a. Integration of isolated Peer: The SF should

integrate isolated peers in existing groups or it

should connect isolated peers among them for

creating a new subgroup. To integrate an isolated

peer, the SF puts her in contact with another member

by promoting a conversation.

The selection of the most appropriate node, among

those similar to the isolated one, is made as follows:

after receiving the ordered list of similar nodes, the

SF evaluates the appropriateness of a node by

considering its centrality, connectivity and

betweeness centrality. Then, the SF selects the

member that is more popular by calculating a rank as

rank=sim*sum(a*f(centrality),b*f(connectivity),c*f(betweeness)) (1)

where coefficients a, b, c allow tuning the function

according to the situation. In our evaluation scenario

we gave a higher priority to centrality and

connectivity than to betweeness by setting the value

of a and b to the double of the value of c.

Once the node has been selected, the SF has to

find an artefact for promoting a conversation with

the isolated one. To this aim it proposes arguments

considering the minimum gap between the

confidence values among their common interests.

If they do not have any common interest the SF

tries with another member (with the rank

immediately lower) otherwise it will decide to

connect the isolated peer to the member with the

most popular member (highest value of centrality).

The dialog management strategy adopted by the SF

is an extension of the methodology proposed in (De

Carolis and Cozzolongo, 2007).

b. Groups Connection: The SF may decide to

connect two different groups to facilitate the

interaction among their members. In this case the

strategy involves selecting (i) members with the

highest betweeness centrality in the two groups and

(ii) a topic taking into account interests of the two

subgroups. As group modelling strategy (Masthoff,

2004) for understanding interests of subgroups, we

applied a weighted average of preferences. Then, if

there is a common node between two groups, this is

used as a bridge for promoting common arguments;

on the contrary, the ones with the highest leadership

(calculated as in (1)) can be put in contact with each

other, using the same strategy described for the

integration of isolated peers.

c. Spreading Information: The strategy we

implemented so far is the following: a list of peers

belonging to every group of the SN is created

according to their degree of betweeness centrality.

Then the SF starts contacting those belonging to the

largest groups and selects among them the node that

is closest in terms of distance to this one, and so on.

If there are isolated peers that have not been

integrated in the SN yet, the SF will contact each of

them and communicate the information. In all plans

the SF communicates with the SNA for requesting

data and measures concerning the situation of the

network or for informing the SNA of its action

effects.

Sniffer Agent: its main goal is to constantly monitor

the SN through overhearing (Fan and Yen, 2005;

Busetta et al., 2001). The Sniffer has to understand

the shallow dialogue dynamics of the networks: this

monitoring activity should be conducted

continuously to have, at every time of the

interaction, the updated image of what is going on in

the SN. The Sniffer will apply conversational

analysis techniques, enabling the SF to both (i)

prevent (or even solve) conflicts and (ii) favour

fruitful exchanges among peers with similar features

and goals. In this perspective it is also important to

understand what is the task of each interaction

among couples or groups of peers (e.g., Information

Seeking, Negotiation etc.) and what is the attitude

the interlocutors are showing towards each other

(e.g., cooperative vs. individualistic, or warm vs.

cold, etc.). The history of the interaction will serve

as a basis for conversational analysis. In particular,

our Sniffer agent will employ Hidden Markov

Models for dialogue pattern analysis, using an

approach similar to the one described in (Novielli,

2010).

ICAART 2011 - 3rd International Conference on Agents and Artificial Intelligence

382

2.1 An Example

Let’s consider the scenario of a recreation center for

elderly people. The SF, represented by AIBO,

interacts with 14 people; 13 of them are divided in 3

subgroups that are in 3 different rooms of the place:

G1=(A,B,C,D), G2=(E,F,G,H,I) and G3=(L,M,N,O);

some members know only some elements of each

group as described in Figure 1.

The initial density of the graph is 0.37. X just

arrived in the center and can be considered an

isolated peer. The SF constantly monitors the state

of the SN by asking information to the SNA and it

receives the information regarding the fact that X is

isolated. Therefore the SF tries to integrate it by

applying the strategy described in Section 2.

Figure 1: The SN sociogram.

Let’s assume that A, B, H, D, L are, in the order,

members that are more similar to X (in the current

setting, nodes with a similarity level above or equal

to 0,5 are considered). Taking into account their

rank (see Table 1), the peer represented by the node

D results as the most suitable to be contacted by the

SF and for promoting a conversation about a

common interest.

Table 1: Selection of the best candidate of the sociogram.

ID Centrality Connectivity Betweeness Similarity(X) Rank

A (3) 4,17 (3) 4 (1) 0 0,8 10,4

B (8) 8,26 (8) 6 (8) 1 0,75 30 2

D (12) 120,44 (12) 10 (12) 55 0,65 39 1

H (1) 2,08 (1) 2 (1) 0 0,7 3,5 9

L (11) 114,26 (8) 6 (11) 54 0,5 24,5

Let’s suppose that the best common interest

between X and D is “Art”. Then the SF introduces X

to D, promoting a conversation about Art. An

example of possible move by the SF is: ‘Good

Morning D, I’m pleased to introduce you X. She is

just arrived. X, like you, is interested in art.’ In case

of conversation between D and X failing because of

lack of interest in the proposed argument or because

of some kind of resistance to start the friendship

relation, the SF tries with the next candidate node in

the list. When the integration succeeds the SN and

the SF beliefs are updated accordingly.

3 CONCLUSIONS

The described framework allows simulating and

evaluating the behaviour of a robot acting as social

facilitator in a smart environment. The system has

been evaluated through the simulation of 25

different scenarios in the described domain. For each

scenario we formalized the peer profiles and,

consequently, the structure of the social network and

a set of rules describing the dynamic of the

interaction among peers. Randomly we assigned a

level of resistance in order to establish a friendship

relation (0, no resistance – 1, resistance). Results

show that in the majority of cases (16) the social

facilitator’s strategies successfully increased the

density of the network. In the rest of cases, the

integration of isolated peers and the subgroups

connection failed due either to absence of common

arguments or to antipathy (simulated through the

resistance variable) towards the isolated peer or

among leaders of subgroups.

In our future work, besides enriching the

formalization of the peer agent mental state with

extra-rational factors (personality traits, moods,

emotions), we intend to endow it with the ability of

reasoning by taking into account these factors to

simulate social intelligence towards other peers.

REFERENCES

Cavalluzzi, A., De Carolis, B., Carofiglio, V., Grassano,

G., 2003. Emotional Dialogs with an Embodied Agent.

User Modeling 2003, pp. 86-95.

De Carolis, B. Cozzolongo, G., 2007. Planning the

Behaviour of a Social Robot Acting as a Majordomo

in Public Environments. In AI*IA 2007, pp. 805-812.

Busetta, P., Serafini, L., Singh, D., Zini, F., 2001.

Extending Multi-Agent Cooperation by Overhearing.

In Sixth ICCIS, pp. 40-52.

Fan, X., Yen, J., 2005. Conversation Pattern-based

Anticipation of Teammates' Information Needs via

Overhearing. In IEEE/WIC/ACM Conference on

Intelligent Agent Technology, pp. 316-322.

Haythornthwaite, C., 1996. Social network analysis: An

approach and technique for the study of information

exchange. Library and Information Science Research,

18 (4), pp. 323-342.

A SOCIAL ROBOT FOR FACILITATING HUMAN RELATIONS IN SMART ENVIRONMENTS

383

Masthoff, J., 2004. Group modeling: Selecting a sequence

of television items to suit a group of viewers. UMUAI

14(1), pp. 37-85.

Novielli, N. 2010. HMM modeling of user engagement in

advice-giving dialogues. JMUI 3(1-2) pp. 131-140.

Rao, A. S., Georgeff, M. P., 1991. Modelling rational

agents within a BDI-Architecture. In SICPKRR, pp.

473-484.

Rodgers, J. L., Nicewander, W. A., 1988. Thirteen ways to

look at the correlation coefficient". The American

Statistician 42, pp. 59-66.

Sakamoto, D, Ono, T, 2006. Sociality of robots: do robots

construct or collapse human relations? In 1st ACM

SIGCHI/SIGART, pp 355–356.

Scott, J., 1991. Social network analysis: A handbook.

Sage, London.

Shadbolt, N., 2003. Ambient Intelligence. IEEE Intelligent

Systems, 18(4).

ICAART 2011 - 3rd International Conference on Agents and Artificial Intelligence

384