People are being in uenced by several means to purchase di erent products or services, e.g., targeted advertisements that are generated by computer systems. Agents in these systems are known as the in uencers of the environment. In this work, it is proposed an architecture based on a Multi-Agent System to show autonomous behaviors with social implications. The social behaviors observed are used to analyze the in uence of suppliers and the trend of products or services in commercial markets. The study case in this paper uses the color based fashion trends.
Technological advances in computer science o er di erent opportunities to create collaborative systems. In order to ful ll its purposes, these kind of systems require coordination and sharing capabilities. Arti cial Intelligence (AI) and Intelligent Distributed Systems are two principal contributors to the development of these environments.
Traditional AI systems are based on a centralized model and its components have the purpose of not a ecting other elements. Modern AI proposes an approach based on agents: entities that perceive their environment through sensors, and respond or act in such an environment through e ectors.
As Damazeau et al. [
In this work, it's proposed a MAS based architecture for the study of fashion trends, e.g., clothes' color. The approach taken initially works with a single solid color and, later, with the combination of two di erent colors. 2
There are several techniques to deploy agents' systems, but it is necessary to
consider an appropriate software architecture to assure the e ective performance
of the system. The most important aspect is how agents interact with each
other to achieve the objective of the system. This interaction depends on the
organizational structure that represents the relationship of the agents inside the
system, and the coordination mechanism to control the sequences of interaction
and con ict management among agents [
The problem of how to design architectures based on agents can be stated
in terms of the form of organizational structures between agents and how to
coordinate the interactions between them. For example, Sanchez implemented an
agent-based system for the simulation of epidemiological behavior of in uenza
AH1N1[
The architecture used in e-commerce systems, according to Zeng [
In a similar design, Aragon's [
The architecture of an agent is an essential element by itself. It allows to
decompose a system into smaller components and determine how is the relationship
between them and, therefore, how they should interact with each other and with
the environment. It can be found various kinds of architectures for the
modeling of agents, e.g., deliberative, cognitive, reactive, and hybrid architectures.
In particular, Muller [
This work relies in the features independently provided by Zeng and Aragon's architecture. On the one hand, Zeng's architecture gives especial importance to the user: its agents have well-de ned tasks to interact between user and system. On the other hand, this architecture is based on personalized search towards similar preferences of the customer, and this is a disadvantage in our case: the aim is to in uence other agents, or to be in uenced by the environment, and not to reinforce our own preference of color.
Aragon's architecture has the advantage that it incorporates agent's recommendations. Moreover, this architecture implements two business models, direct and indirect, which are used in e-commerce to nd a customer or notify a user that the recommendations are directed at him. However, agents in this architecture are intended to seek bids according to preferences or customer pro les. That intention is not desirable for the proposal in this work.
After analyzing the strengths and weaknesses of both architectures, a hybrid architecture is proposed which consists of the following agents: Interface or Environment, Color, Comparator, Expert and Recommender (see Fig. 3).
Interface or Environment. It is the intermediary between user and system. It will be the means by which the user can enter data, and feedback, into the system. Furthermore, it will broadcast to other agents to announce the task requested by the user, and one of them will accept the task, according to its abilities, skills and knowledge.
Color. In this system there will be only three such agents, each de ned by a color: red, green and blue. These colors represent the trends that will be handled. They will be responsible for in uencing neutral populations. However, the belief of these agents cannot change because the agents are\providers" of a color.
Comparator. The function of this agent is to perform, after a certain period, and if any of the o cers of color accept the task, to contact the expert agent to solve the task in progress. Once this agent has the solution, it should make a broadcast to o cers of color to apply the solution.
Expert. This agent is tasked to respond to Comparator. When Color agents are not able to handle user's requests, this entity knows how to create all shades from three color agents. Expert can say how to x the problem. This agent has a preliminary knowledge base on RGB colors.
Recommender. His goal is to present the proposals obtained by the agents of color, this is made directly with the agent Environment.
The simulation tool NetLogo Framework was used to implement the proposed
architecture. NetLogo is a programming language that allows the speci cation
of the behavior of each patch (plots), turtles (agents) and execution control.
The language is simple, expressive and functional. Each agent in NetLogo o ers
perception of its environment and acts on it, carries its thread of control and it
is autonomous [
The simulation included three suppliers agents which were de ned according
to the BDI (Beliefs-Desires-Intentions) architectural model [
The population without trend, as simulation's time progress, will know the existence of agents with a trend (suppliers). In each encounter with suppliers, individual agents will increase their belief in this trend. After some threshold, ve in this simulation, they will adopt the preference for that color, i.e., their beliefs have changed.
The algorithm was implemented for transmission of beliefs from initials agents to acquirers agents. Provider agents move randomly in the environment and, consequently, there will be a time to converge of two or more agents as neighbors. Transfer of beliefs is made to a purchaser agent when the other agent is a Color or a purchaser that has only the belief of a trend at the time. However, not necessarily the rst convergence will transfer the belief by the supplier.
What happens when the percentage of belief of two trends are equivalent? This gives place to a new generation trends. However, this agent is not able to share this new trend to other agents so, if it is feasible that in a future iteration it become part of the group of agents that share, but only one of the three initial trends will be used.
In Fig. 4, an initially de-trended population are in uenced by three supplier agents. After some time, it is observed how the agent with blue trend has been most successful compared to the other two supplier agents (red and green). In the graph located on the right side of Fig. 4, x-axis represents time, and y-axis represents the population for each trend. It is easy to observe the generated tendencies, and the result of the in uence of suppliers.
The fact that most of the population tendency is blue does not prevent the other two agents to continue to pursue their objective. This generates a greater number of collisions between them which develops a new trend. 5
Several simulations were performed, each one consisted of four runs, but the value of population's variable was changed among them. Results are shown below with the population size and time (ticks) as indicated. Four runs were needed in order to introduce new trends (colors).
In Table 1, it can be observed a test performed with a population of 30 agents. After 300 ticks, it appeared that the population with no tendency is reduced in each run, but in small amounts. In Fig. 5, it can be seen more clearly the percentage of the population that has no tendency after four runs, and the in uence of providers agents on the population. The fact that the population is small a ects the appearance of new trends.
Subsequently, a second test was performed with the results given at Table 2. With an increase in the population to 50 agents without tendency, similar results to that obtained in the rst test behavior are observed, taking into account that the increase of the population is minimal.
In a third test, population size was doubled. The results of 100 agents can be seen in Table 3. This time, the agents detrended have disappeared, hence, have been in uenced by the provider agents. Similarly, in Fig. 7, it can been seen how the agent with blue trend has grabbed almost all of the population at the end of the fourth run. It was the rst to obtain enough \supporters" and, consequently, to in uence the other agents without tendency. Also, in this test, it can be seen the rst emergence of new trends.
Finally, population was increased to 500 agents. The results can be seen in Table 4 and in Fig. 8. The behavior of agents has slight variations. But in this time, there is greater number of objectives, and acquirer's beliefs are more volatile. Some changes in trends are also noted at the rst run in population without trend. The variation of the subsequent runs has no radical changes, that is, it was stabilized, thereby achieving coexistence of virtually all color shades. Social tendencies are a complex pattern to analyze and to create predictions about them. Nonetheless, current technology and algorithms permit us to create simpli ed models with acceptable time constraints. The simulated models are useful for creating inferences and some insights might be achieved.
This is the case in this study. By creating an arti cial environment in which people (autonomous agents) live, it was possible to simulate the capabilities of in uencers (specialized agents) that are introducing a fashion trend (a color). The resultant data enable us to analyze the pattern that was needed in order to people to change its color preferences.
As the results shown, with a small population, it's harder to introduce a trend. The reasoning might be that interaction of people is not big enough in order to exchange information about their preferences. As population increase, interactions became more common, and trends started to appear. This was not only the result of the in uencers, but also of agents that already have a color preference and they share that information.
Initial tests show tendencies for a single di erent color. But another interesting pattern emerged when population was considerable greater than the initial one. Trends for a mix of two colors started to appear in the population. This was the result of similar interactions, in quantity, with di erent in uencers. And as time went by, agents created a trend that was composed of a couple of colors.
This research represents the initial phase of a work in progress that aims at understanding social tendencies. In future work, it is desired to introduce more variables to get closer to complexity of the decision-making procedure of real people.