COMAS: A Multi-agent System for Performing Consensus Processes

COMAS: A Multi-agent System for Performing Consensus Processes Iván Palomares, Pedro J. Sánchez, Francisco J. Quesada, Francisco Mata and Luis Martínez Computer Science Department, University of Jaén, Spain {ivanp,pedroj,fqreal,fmata,luis.martinez}@ujaen.es

Abstract. The need for achieving consensus in group decision making problems is a common and sometimes necessary task in a myriad of social and business environments. Different consensus reaching processes have been proposed in the literature to achieve agreement among a group of experts. Initially, such processes were guided by a human moderator, but afterwards, some proposals to facilitate such a process arose by automating the moderator tasks. However, not many consensus support systems have been developed so far, due to the difficulty to manage intelligent tasks and cope with the negotiation process involved in consensus. This paper aims to present an initial prototype of an automatic consensus support system, developed by using the multi-agent paradigm that provides intelligent tools and capacities to tackle the inherent complexity found in this problem. To do so, we focus on the consensus model considered, the multi-agent architecture designed to develop such a system, and the ontology used for reasoning and communication tasks.

1

Introduction

In group decision making problems (GDM), two or more experts try to reach a common solution about a decision problem. Traditionally, these problems have been solved performing a selection process where the best alternative/s is/are chosen as the solution, without taking into account any previous agreement among experts [9]. This often leads to situations where some experts may consider that their individual opinion has not been taken into account, and therefore they disagree with the achieved solution [19]. To avoid such situations, the idea of carrying out a consensus process prior to the selection process emerged, so that experts express and discuss their preferences to make them closer to each other with the aim of reaching a high level of agreement before making the decision [15]. The consensus process has typically been coordinated by a human moderator, who is

2

responsible for process supervision and evaluation of the level of agreement achieved in each consensus round. Regarding the automation of consensus reaching processes (in order to perform them without human supervision), some models have been proposed [11,16], but not so many have been finally developed. Intelligent agents are software entities capable of carrying out actions in an autonomous way to achieve one or more aims, reasoning about acquired knowledge and exchanging information with other agents and/or the environment. In most contexts and problems, different agents, each one with its specific role and behaviours, must be identified, thus establishing a multi-agent system (MAS). Applications for MAS include areas such as planning [8], industry [1] and, more recently, works related to web services [21]. By considering the aforementioned problem and statements, this contribution aims to present a multi-agent system to support consensus processes (COMAS). This system automates and guides consensus processes by means of a set of different intelligent agents that manage, supervise and control them. This paper is organized as follows: in the next Section, we briefly review consensus GDM problems in general and show the theoretical consensus model used by COMAS in particular. In Section 3, we then describe the MAS architecture of our system. Section 4 briefly describes the ontology employed for agent communication. Section 5 shows an example of the system’s performance, and finally, in Section 6 some conclusions and future works are drawn.

2 Consensus Model Description Group decision-making (GDM) problems may be defined as decision situations where [16]: 1. There is a decision problem to be solved, where a solution must be chosen among a set of alternatives X = {x1,x2,…,xn}. 2. Two or more experts E = {e1,e2,…,em} participate, having each one his own opinions about the set of alternatives X that the problem considers. 3. Experts try to achieve a common solution. The process for reaching a solution in GDM problems consists of two steps [18]:  Aggregation phase, that combines the experts’ preferences.  Exploitation phase, to obtain an alternative or subset of alternatives as the problem’s solution. The above selection process does not necessarily imply any agreement among the experts, therefore the solution might not be accepted by some of them if they consider that their preferences have not been taken into account in the process [2]. In order to increase the agreement among them, an extra phase is introduced, the consensus phase.

3

A consensus reaching process consists in a discussion process where a group of experts try to achieve an agreement, supported by a human moderator. Many theoretical consensus models have been proposed in the literature [3,4,5,6,12,13]. Figure 1 shows the consensus model implemented by COMAS, based on the main ideas of the models presented in [10,11,16].

Fig. 1. Consensus reaching model employed in COMAS

Consensus reaching process is seen as an iterative process, where experts provide their preferences about the set of alternatives when each discussion round begins. A brief description of each phase is shown below: 1. Gathering information. Experts provide their opinions to the moderator, by means of structures called preference relations, which consist of one matrix by expert Pi, where each element pilk represents the degree of preference on the alternative xl over xk according to the expert ei. 2. Computing consensus degree. The moderator computes the level of agreement once gathered all the experts’ preferences, by means of a similarity measurement. The agreement level is expressed as a value in the interval [0,1], where a value of 1 means total agreement. 3. Consensus control. The degree obtained in the previous phase is checked. If it is greater than a consensus threshold given, then the desired agreement has been reached and the consensus process finishes. Otherwise, the process needs more discussion. In addition, a maximum number of rounds is set before beginning the process, so that if it is reached, then the consensus process fails. 4. Advice generation. The model suggests how experts should change some of their preferences in order to increase the level of agreement in the following rounds [5,16]. A set of suggestions regarding appropriate changes is generated and delivered to experts. A further detailed description of the consensus reaching process developed in COMAS can be found in [17].

4

3 COMAS multi-agent architecture In this section, the multi-agent architecture designed for COMAS is presented. The architecture was developed basing on the FIPA standard1, and using the JADE agent platform2; and is depicted in Figure 2.

Fig. 2. COMAS architecture

We can consider COMAS a cooperative MAS, where agents work together in order to reach a common objective, i.e., reaching a consensus. Several agent roles are proposed according to the different roles and tasks that can be identified in the previously reviewed consensus model. These roles are described below:  Expert Agent: It represents a human expert, and acts on behalf of him.  Moderator agent: It assumes the human moderator’s role, and is responsible for guaranteeing a right development of the overall consensus process. In addition, because of the complex set of responsibilities initially assumed by the moderator, we have decided to add some specialized agents to support moderator agent in various specific tasks in the consensus process:  Consensus Evaluator Agent: It is in charge of computing the consensus degree achieved in each consensus round, and report it to moderator agent.  Change Detector Agent: It is responsible for leading the necessary operations for the Advice generation stage in the consensus process.  Analyst Agent: This additional agent assumes the task of storing all the information concerning each consensus process in a database. Other components provided in our system’s architecture are:

1 FIPA: http://www.fipa.org 2

JADE: http://jade.tilab.com

5

 Interface Agent: Each expert agent is associated to an interface agent, which allows human experts or users to provide their initial preferences before the process begins, as well as showing them the results achieved by the consensus process.  FIPA-specific components and agents: FIPA standard (Foundation for Intelligent and Physical Agents) provides some utilities for general purpose multiagent systems, such as the Directory Facilitator(DF), AMS (Agent Management System), and MTS (Message Transport System).  Ontology: Agents communicate each other by exchanging messages. Therefore, a ontology has been defined for COMAS, so that agents share a common semantics and knowledge about the problem. A detailed description about the system’s ontology is given in Section 4.  Consensus Support System: This is the software model containing all necessary Java classes to perform all operations in the consensus process.  Database: It stores and recovers past information about consensus processes.

4 COMAS Ontology In this section, we show the ontology designed to allow communication among agents under a common language and semantics [7,20]. Since COMAS ontology is based on the idea proposed by Kacprzyk and Zadrozny in [14], its design considers both the necessary concepts for performing consensus processes and those ones used by agents for reasoning about the particular problem’s knowledge. Figure 3 shows the components used in our ontology. These components are divided into three categories:

Fig. 3. Ontology components used in COMAS

6

1. Concepts: Symbolic expressions representing objects. They consist of one or more attributes or slots. Concepts are not individually used by agents, but they appear as part of predicates or agent actions in ACL messages. Concepts designed in our system include: AID, Alternative, Consensus, ExpertAgent, OpAggregation, OpSimilarity, PreferenceRelation and Recommendation. 2. Predicates: Expressions about the state of the world, their value would be either true or false. They are commonly used in COMAS as the content of ACLInform messages (answer to a request message), and include: ProvideAssessment, ProvideConsensus and ProvideRecommendation. 3. Agent Actions: These expressions are a special case of concepts with an additional semantics, indicating actions that can be performed by agents, and they are usually the content of ACL-Request and ACL-Propose messages in our system. COMAS agent actions include: JoinConsensus, MakeAssesment, RateConsensus and RateRecommendation.

4 COMAS Performance In this section, we briefly show some experimental results obtained from a simulation we have carried out on our platform. The simulation consisted in solving a consensus problem with 4 alternatives and 100 experts, by using different consensus measures. The high scalability of COMAS makes possible performing consensus processes automatically with a high number of experts, which hasn’t generally been done previously. Consensus measures are used for measuring the level of agreement among experts, through the use of similarity measures and aggregation operators. We have considered the use of different combinations of these measures for each one of the experiments made. In addition, we consider a consensus threshold of 0.85 and a maximum number of 10 rounds permitted. Figure 4 shows the convergence towards consensus, by displaying the number of required rounds and the level of agreement achieved, in some of the tests carried out on COMAS by using a similarity measure based on Euclidean Distance and different well-known aggregation operators, some of which may be found in [22] for further detail about them. The overall results obtained from these tests and another ones carried out with more different measures, leaded us to conclude that the use of different consensus measures allows decision makers to use COMAS to solve consensus problems according to their requirements, ranging from situations where a fast agreement is required, to contexts where more discussion among experts is needed.

7

Fig. 4. Results achieved solving a consensus problem through simulation.

5 Conclusions and Future Work Even though consensus processes have been widely studied and many different models were proposed, there are barely real implementations of such a type of system. This contribution has introduced COMAS, a multi-agent architecture to support and automate consensus reaching processes with a high number of experts. We are currently tackling different improvements in COMAS. On the one hand, we are aimed at deploying a user interface based on Web Services, as well as implementing different profiles of ‘personality’ for experts. On the other hand, we intend to deploy our system in intelligent environments, thus discovering real applications in smart homes.

References [1] Aldea, A., Bañares-Alcántara, R., Jiménez, L., Moreno, A., Martínez-Miranda, J. and Riaño, D., The scope of application of multi-agent Systems in the process industry: three case studies, Expert Systems with Applications, 26 (1), pp. 39-47, 2004. [2] Alonso, S., Herrera-Viedma, E., Chiclana, F. and Herrera, F., Individual and social strategies to deal with ignorance situations in group decision making, International Journal of Information Technology and Decision Making, 8 (2), pp. 313-333, 2009. [3] Ben-Arieh, D. and Chen, Z., Linguistic labels aggregation and consensus measure for automatic decision-making using group recommendations, IEEE Transactions on Systems, Man and Cybernetics, Part A: Systems and Humans, 36 (1), pp. 558-568, 2006.

8

[4] Ben-Arieh, D., Easton, T. and Evans, B., Minimum cost consensus with quadratic cost functions, IEEE Transactions on Systems, Man and Cybernetics, Part A: Systems and Humans, 39 (1), pp. 210-217, 2009. [5] Cabrerizo, F.J., Pérez, I.J. and Herrera-Viedma, E., Managing the consensus in Group decision making in an unbalanced fuzzy linguistic context incomplete information, Knowledge-based systems, 23, pp. 169-181, 2010. [6] Fedrizzi, M., Fedrizzi, M. and Pereira, R.A.M., Soft consensus and network dynamics in group decision-making, International Journal of Intelligent Systems, 14 (1), pp. 6377, 1999. [7] García-Sánchez, F., Valencia, R., Martínez, R. and Fernández, J.T.,An ontology, intelligent agent-based framework for the provision of semantic web services, Expert Systems with Applications, 36 (1), pp. 3167-3187, 2009. [8] Gnansounou, E., Pierre, S., Quintero, A., Dong, J. and Lahlou, A., A multi-agent approach for planing activities in decentralized electricity markets, Knowledge-based systems, 20 (4), pp.406-418, 2007. [9] Herrera, F., Herrera-Viedma, E. and Verdegay, J., A sequential selection process in group decision making with linguistic assessments, Information Sciences, 85, pp. 223239, 1995. [10] Herrera-Viedma, E., Herrera, F. and Chiclana, F., A Consensus Model for Multiperson Decision Making with different Preference Structures, IEEE Transactions on Systems, Man and Cybernetics, Part A: Systems and Humans, 32 (3), pp. 394-402, 2002. [11] Herrera-Viedma, E., Martínez, L., Mata, F. and Chiclana, F., A consensus support system model for Group decision-making problems with multigranular linguistic preferente relations, IEEE Transactions on Fuzzy Systems, 13 (5), pp. 644-658, 2005. [12] Kacprzyk, J., Nurmi, H. and Fedrizzi, M., Consensus under Fuziness, International Series in Intelligent Technologies, 10, Springer, 1997. [13] Kacprzyk, J. and Zadrozny, S., On a concept of a consensus reaching process support system based on the use of soft computing and web techniques, Computational Intelligence in Decision and Control, 1, pp. 859-864, 2008. [14] Kacprzyk, J. and Zadrozny, S., Soft computing and web intelligence for supporting consensus reaching, Soft Computing, 14 (8), pp. 833-846, 2010. [15] Martínez, L. and Montero, J., Challenges for Improving Consensus Reaching Process in Collective Decisions, New Mathematics and Natural Computation, 3 (2), pp. 203-217, 2007. [16] Mata, F., Martínez, L. and Herrera-Viedma, E., An adaptive consensus support system model for Group decision-making problems in a multigranular fuzzy linguistic context, IEEE Transactions on Fuzzy Systems, 17 (2), pp. 279-290, 2009. [17] Mata, F., Sánchez, P.J., Palomares, I., Quesada, F.J. and Martínez, L., COMAS: A Consensus Multi-Agent based System, Proceedings of the International Conference on Intelligent Systems Design and Applications (ISDA’10), pp. 457-462, 2010. [18] Roubens, M., Fuzzy Sets and decision analysis, Fuzzy sets and Systems, 90 (2), pp. 199-206, 1997. [19] Saint, S. and Lawson, J.R., Rules for Reaching Consensus. A modern approach to decision making, Jossey-Bass, 1994. [20] Staab, S. and Studer, R., Handbook on Ontologies, International Handbooks on Information Systems, Springer, 2005. [21] Wang, S., Shen, W. and Hao, Q., An agent-based web service workflow model for inter-enterprise collaboration, Expert Systems with Applications, 31 (4), pp. 787-799, 2006. [22] Yager, R.R., Applications and extensions of OWA aggregations, International Journal on Man-Machine Studies, 27 (1), pp. 103-132, 1992.