HAL Id: hal-01400997
https://hal.archives-ouvertes.fr/hal-01400997
Submitted on 1 Jun 2017
HAL is a multi-disciplinary open access
archive for the deposit and dissemination of sci-
entic research documents, whether they are pub-
lished or not. The documents may come from
teaching and research institutions in France or
abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, est
destinée au dépôt et à la diusion de documents
scientiques de niveau recherche, publiés ou non,
émanant des établissements d’enseignement et de
recherche français ou étrangers, des laboratoires
publics ou privés.
A Multi-context BDI Recommender System: from
Theory to Simulation
Amel Ben Othmane, Andrea G. B. Tettamanzi, Serena Villata, Nhan Le
Thanh
To cite this version:
Amel Ben Othmane, Andrea G. B. Tettamanzi, Serena Villata, Nhan Le Thanh. A Multi-context
BDI Recommender System: from Theory to Simulation. Web Intelligence, Oct 2016, Omaha, United
States. �10.1109/WI.2016.0104�. �hal-01400997�
A Multi-context BDI Recommender System:
from Theory to Simulation
Amel Ben Othmane
∗
, Andrea Tettamanzi
†
, Serena Villata
†
, Nhan Le Thanh
†
∗
Universit
´
e C
ˆ
ote d’Azur, ADEME, Inria, CNRS, I3S France
amel.ben-othmane@inria.fr
†
Universit
´
e C
ˆ
ote d’Azur, CNRS, Inria, I3S, France
{andrea.tettamanzi, serena.villata, nhan.le-thanh}@unice.fr
Abstract—In this paper, a simulation of a multi-agent rec-
ommender system is presented and developed in the NetLogo
platform. The specification of this recommender system is based
on the well known Belief-Desire-Intention agent architecture
applied to multi-context systems, extended with contexts for
additional reasoning abilities, especially social ones. The main
goal of this simulation study is, besides illustrating the usefulness
and feasibility of our agent-based recommender system in a
realistic scenario, to understand how groups of agents behave in a
social network compared to individual agents. Results show that
agents within a social network have better collective performance
than individual ones. The utility and the satisfaction of agents is
increased by the exchange of messages when executing intentions.
I. INTRODUCTION
Nowadays, recommender systems must cope with an in-
creasing demand of complexity, for example, an application
for recommending routes in a traffic scenario should deal
with different contextual information (e.g., information about
the user location) and other non-logical components linked
to human behavior like desires, beliefs or emotions. For
this reason, multi-agent systems are considered as suitable
alternatives for modeling and simulating this kind of real-
world scenarios, where different entities autonomously interact
in a dynamic and uncertain environment. In particular, one
of the most popular agent architectures, the Belief-Desire-
Intention (BDI) model, seems to be particularly suitable for
this task. Under this model, the mental state of the agent
is composed by a set of beliefs, desires and intentions that
consist of informational, motivational, and deliberative states,
respectively.
Recently, the Artificial Intelligence (AI) community is
putting much effort on the investigation and evaluation of
recommender systems based on intelligent agents. Such a kind
of systems have been applied so far in different fields such as
health-care [1], tourism [2] and traffic and transportation [3].
A complete taxonomy of recommender agents can be found in
Montaner et al. [4]. However, few works only combine BDI
agents and recommender systems. Among them, we refer in
particular to the approach of Casali et al. [5] who propose a
BDI recommender agent in the tourism domain.
In this paper, we propose a multi-agent simulation to evalu-
ate the overall behavior of the multi-context BDI recommender
system we presented in [6], where different strategies are
applied. The proposed framework aims at recommending a
plan for a user taking into account different contexts. For this
purpose, we give an overview of the different theories used to
define contexts, and explain how all those contexts are relied
together to define the whole behavior of the system. In order
to evaluate the goodness of the proposed recommendation, we
compare the performance of the system with two different
strategies, namely the solitary agent strategy, and the social
agent strategy.
The reminder of this paper is organised as follows: Sec-
tion II provides an overview of our multi-context BDI formal
framework, highlighting the main features of the system. The
simulation and the results are discussed in Section III. Finally,
some conclusions are drawn.
II. THE MULTI-AGENT FRAMEWORK
The specification of our agent model is based on Multi-
Context Systems (MCS) [7], to allow for a separation of the
definitions of the different formal components or units. A MCS
is defined as a group of interconnected units ⟨{C
i
}
i∈I
, ∆
br
⟩
where each context is defined as a tuple ⟨L
i
, A
i
, ∆
i
⟩ where
L
i
, A
i
and ∆
i
are the language, axioms, and inference rules,
respectively. ∆
br
is a set of bridge rules, i.e., rules of inference,
which relate formulas in different units. A bridge rule is of the
form:
C
1
∶ φ, C
2
∶ ψ → C
3
∶ θ
and it can be read as: if the formula φ can be deduced in
context C
1
, and ψ in C
2
, then the formula θ is to be added to
the theory of context C
3
.
The advantage of adopting MCS is illustrated below, where
we use MCS for the specification of our extended BDI agent
based on [7]. As visualized in Figure 1, our multi-context BDI
agent is defined as follows:
Ag = ({BC, DC, GC, SC, P C, IC, CC}, ∆
br
)
where BC, DC, GC represent respectively the Belief Con-
text, the Desire Context and the Goal Context which model
an agent mental attitude. P C, IC and CC are functional
contexts that represent respectively the Planning Context, the
Intention Context and the Communication Context. SC is for
the Social Context, and it models social influence between
agents. We first extended the classical BDI model with others
Fig. 1. The Multi-context BDI Agent Model
contexts. The Goal Context, for example, is introduced based
on [8], where goals are considered as a list of desires that,
besides being logically consistent, are also maximally desir-
able. Second, in order to represent and reason about graded
notions of beliefs, desires and goals, we use the classical
propositional language with additional connectives as language
L
i
, following [8], [9], where uncertainty reasoning is dealt
with possibility theory [10]. The behavior of these contexts is
handled by means of internal deduction rules ∆
i
and axioms
L
i
which contain axioms from classical propositional logic,
and from necessity and possibility measures of possibility
theory. The detailed formalization of each context can be found
in [6].
Concerning the Social Context, we assume that an agent
has the tendency to be socially influenced by other agents to
adopt a certain mental attitude if it has similarities with the
latter without the need to be in an explicit social relationship
with it, e.g., to have the same goals or to be in the same
location. Consequently, if an agent a
i
is similar to another
agent a
j
, a direct link is created with the latter. Between them,
we consider a trust relationship and the trustworthiness of a
i
towards agent a
j
about an information φ is interpreted as a
necessity measure τ ∈ [0, 1].
For specifying and reasoning over plans, i.e., the Planning
and Intention contexts, we propose to adopt the 5W (Who,
What, Why, When, Where) vocabulary
1
, which is relevant
for describing different concepts and constraints related to
plans and allows spatial and temporal reasoning over plans
and intentions.
On the one hand, the behavior of each context is handled
by axioms and inference rules. On the other hand, the overall
behavior of the system is handled by bridge rules like Rule
(2) linking GC to DC, and expressed as follows:
(2) ⊧ GC ∶ G(a
i
, φ) = δ
φ
→ DC ∶ D
+
(a
i
, φ) = δ
φ
It can be read as follows: if an agent a
i
has as goal φ with
a satisfaction degree δ
φ
in a GC then it positively desires φ
1
http://ns.inria.fr/huto/5w/
with the same degree δ
φ
in a DC. For more details about the
proposed model, we refer the reader to [6].
To show the applicability of our multi-agent BDI frame-
work, an experimental model is presented and evaluated in
the next section using the NetLogo Platform.
III. THE SYSTEM SIMULATION
In agent-based systems with spatial reasoning and social
behavior, a visual output is needed to display the agents’
moving and interaction in two or three dimensional spaces.
The Netlogo graphical user interface offers the possibility to
design agents with different shapes and positions. Each agent
in the simulation is a multi-context BDI agent whose behavior
is described in the previous section. An agent represents a user
with different desires and beliefs that are randomly initialized.
The aim of the simulation is to compute a recommendation
based on a user initial set of beliefs and desires, and to see
how our agent will adapt the recommendation in two cases:
● the agent is part of a social network (social agent), i.e.,
it has links with other agents similar to it,
● the agent is considered as a solitary agent, i.e., it has no
interaction with other agents.
Plans are a list of activities that consist of moving from one
destination to another. Each destination contains some rewards
that an agent will obtain when it reaches that destination. The
amount of rewards for each agent is random. Once rewards
are gained, an agent will broadcast information about the
number of remaining rewards in the correspondent destination
to similar agents. These agents will decide to accept or not this
recommendation according to the trust degree in the sender,
and whether there is any information in their knowledge
base (desire or belief base) that contradicts this one. If an
agent decides to accept the recommendation, then it adds this
information to its desire base, and then trigger the recalculation
of its intentions according to the updated desire base.
A. Experimental Setup
Table I summarises the parameters that can be varied for
different use cases. As shown in Figure 2, agents are initially
randomly distributed in the space (patches in NetLogo). They
also have different profiles (desires, trust degrees, positions,
etc. . . ). Links are also created randomly between agents ac-
cording to an initial link number defined by the user at the
beginning of the simulation (on the left-hand side of Figure 2).
We used Netlogo 5.3.1 version to implement our simulation.
For the BDI behavior and the communication context we
use two available NetLogo libraries [11], one for BDI-like
agents and the other for ACL-like communication, allowing
the development of goal-oriented agents that communicate
using FIPA-ACL messages. We developed the rest of the
behavior of the agents using the NetLogo language with some
extensions. The objective of the simulation is to assess the
effects of these agents on the system as a whole (and not only
to assess the effect of individual agents on the system).
Fig. 2. The User interface of our multi-agent simulation in Netlogo. The person icon represents an agent which represents a user. Flags represent destinations
in which agents can go. Labels represent an agent intention which consists of two elements: the name, mapped to a NetLogo command, and a done-condition,
mapped to a NetLogo reporter. Intentions are stored in a stack, and are popped out when to be executed. If the done-condition is satisfied, the intention is
removed and the next intention is popped out consecutively. The figure shows also, on the right side, how the graphs are updated dynamically as the program
runs.
TABLE I
THE SCALE AND DISTRIBUTION OF PARAMETERS IN THE SIMULATION.
Parameters Scale Distribution
Number-of-agents 0-100 Random
Desires 0-50 Random
Beliefs 0-100 Random
Intentions 0-10 Random
Links 0-100 Random
Gain 0-50 Random
B. Experimental Results and Discussion
The model and experimental data were analysed using the
RNetLogo extension [12]. Once the experiment is set up, each
agent will have a list of random desires, beliefs are empty
at the beginning. According to these desires and the afore-
mentioned behavior, an agent calculates the recommendation
which has a plan as output. This plan will become the agent’s
intention, and the agent will execute it. In the case of a solitary
agent, it will execute its plan without any change. Only a new
belief from an external source that does not contradict the
agent initial belief can make it change its intention. In the other
case, i.e., a social agent, similar agents will communicate a set
of proposed recommendations with the aim to influence the
others to change their beliefs or desires. If the recommendation
is accepted, the agent will recalculate its intentions based on
the recommendation, and it will follow a new plan. Metrics
such as utility or satisfaction are calculated using the following
equations:
utility(p) =
∑
i∈G
S
g(i)
∑
j∈D
initial
d(j)
where G
S
is a set of goals satisfied by a plan p, and D
initial
is a set of initial desires of an agent.
satisfaction-degree(p) = max{G(φ
i
), i ∈ [0, n]}
where n is the number of goals satisfied by a plan p. The
utility measure estimates how much the user needs (desires)
match the recommendation (plan). The satisfaction degree,
as its name suggests, computes the user satisfaction about a
recommendation based on its initial degrees of desires.
The mean gain of the agents is also reported, and results are
showed in Figure 3. We can see that agents within a social
context, i.e., agents that communicate in order to influence
each other, accrue more gain most of the time in comparison
with those without a social context. These results show that
a social population could have a greater social welfare than a
non social one when agents have similar interests.
For comparison, we calculate the average satisfaction degree
and utility over time for 50 agents in the case of individual
and social agents. One may expect that the probability of
gaining utility will increase with exchanging messages. Fig-
ure 4 confirms this expectation. It shows that utility augments
considerably within social agents compared to the utility
within individual agents. We notice that the average utility
is the same over time for individual agents. We can deduce
that exchanging ones beliefs and desires increases, on average,
the agents utility.
In Figure 5, we can see the average satisfaction of the agents
about the recommendations they received (plans). The average
is higher within social agents than within individual ones.
We can conclude that agents get more satisfaction collectively
from exchanging information than alone.
These results provide for agents further motivation to en-
gage in communications with similar trustworthy agents and
support our modeling choices. It is also interesting to note
Fig. 3. Mean gain of agents with and without a social context.
Fig. 4. Mean utility of agents with and without social context.
Fig. 5. Mean satisfaction degree of agents with and without social context.
how communities of agents (e.g., agents with similar interests)
likely to be influenced are more efficient collectively than
individual agents. However, it is also interesting to see how
the system will behave if some malicious agents communicate
incorrect information. An interesting approach is presented
in [13], where a score pair (trust, distrust) is used. We
are currently studying the possibility of merging the two
approaches.
IV. CONCLUSIONS
In this paper, we have presented an agent-based simulation
of the framework we proposed in [6]. The purpose of the sim-
ulation was to evaluate agent behaviors adopting two different
strategies (the social and the individual strategy) in order to
infer the quality of recommendations. Results show that agents
achieve a better performance collectively when they are in
“communities”, i.e., agents with shared interests (thus similar
to each other), then when they are acting as solitary agents.
We believe that the issues of trust and recommendation are
tightly related. Results show that exchanging beliefs or desires
with trustworthy agents can improve the whole performance
of agents. However, we ignore how the framework will behave
when errors are introduced in the communicated information.
For that reason, we need a mechanism for checking informa-
tion reliability and updating the trust value, accordingly. We
believe that approaches like [13] can be used to extend our
agent model with further reasoning abilities and, consequently,
to deal with information reliability as well. Our ongoing
work, then, is to expand the multi-agent simulation with those
abilities. For future work, we will investigate how to extend
the agent framework with temporal and spatial reasoning to
be more representative of real-world applications.
REFERENCES
[1] B. Lopez, S. Aciar, B. Innocenti, and I. Cuevas, “How multi-agent
systems support acute stroke emergency treatment,” in IJCAI Workshop
on Agents Applied in Health Care, 2005, pp. 51–59.
[2] M. Batet, A. Moreno, D. S
´
anchez, D. Isern, and A. Valls, “Turist@:
Agent-based personalised recommendation of tourist activities,” Expert
Systems with Applications, vol. 39, no. 8, pp. 7319–7329, 2012.
[3] B. Chen and H. H. Cheng, “A review of the applications of agent tech-
nology in traffic and transportation systems,” Intelligent Transportation
Systems, IEEE Transactions on, vol. 11, no. 2, pp. 485–497, 2010.
[4] M. Montaner, B. L
´
opez, and J. L. De La Rosa, “A taxonomy of
recommender agents on the internet,” Artificial intelligence review,
vol. 19, no. 4, pp. 285–330, 2003.
[5] A. Casali, L. Godo, and C. Sierra, “A tourism recommender agent: from
theory to practice,” Inteligencia artificial: Revista Iberoamericana de
Inteligencia Artificial, vol. 12, no. 40, pp. 23–38, 2008.
[6] A. B. Othmane, A. G. Tettamanzi, S. Villata, M. Buffa, and N. Le Thanh,
“A multi-context framework for modeling an agent-based recommender
system,” in 8th International Conference on Agents and Artificial Intel-
ligence (ICAART2016), 2015.
[7] S. Parsons, N. R. Jennings, J. Sabater, and C. Sierra, “Agent specification
using multi-context systems,” in Foundations and Applications of Multi-
Agent Systems. Springer, 2002, pp. 205–226.
[8] C. da Costa Pereira and A. G. Tettamanzi, “An integrated possibilistic
framework for goal generation in cognitive agents,” in Proceedings of
the 9th International Conference on Autonomous Agents and Multiagent
Systems: volume 1-Volume 1. International Foundation for Autonomous
Agents and Multiagent Systems, 2010, pp. 1239–1246.
[9] ——, “Syntactic possibilistic goal generation.” in ECAI. Citeseer, 2014,
pp. 711–716.
[10] D. Dubois and H. Prade, “Possibility theory: qualitative and quantitative
aspects,” in Quantified representation of uncertainty and imprecision.
Springer, 1998, pp. 169–226.
[11] I. Sakellariou, P. Kefalas, and I. Stamatopoulou, “Enhancing netlogo to
simulate bdi communicating agents,” in Artificial Intelligence: Theories,
Models and Applications. Springer, 2008, pp. 263–275.
[12] J. Thiele, “R marries netlogo: Introduction to the rnetlogo package,”
Journal of Statistical Software, vol. 58, no. 1, 2014.
[13] P. Victor, C. Cornelis, M. De Cock, and P. P. Da Silva, “Gradual trust
and distrust in recommender systems,” Fuzzy Sets and Systems, vol. 160,
no. 10, pp. 1367–1382, 2009.