Modeling of an active multi-agent environment for the design of a multi-criteria group decision support system

Abstract

In the present study, the research problem concerns business intelligence, more precisely collaborative decision-making. The authors propose a complete modeling of a multi-agent active environment for the design of a multicriteria group decision support system dedicated to the spatial problem of localization in territory planning. The proposed model is called ActiveGDSS (Active Group Decision Support System) which uses a coupling between a geographic information system and a multi agents system and is endowed by a new negotiation protocol based on the concession allowing reaching to a consensus which satisfies the territorial actors. The main purpose is to integrate the principle of contextual activation in the modeling of the system which makes the environment an active entity. The main advantages of contextual activation are efficiency gain in terms of execution, better flexibility and reuse of agent behaviors.

Keywords

Territory planning (TP)negotiation protocol group decision support system (GDSS)multi agent system (MAS)active environment contextual activation

1. Introduction

The research problem of this study concerns business intelligence, more precisely collaborative decision-making. Collaborative decision-making, group decision or multi-decision-maker decision treats the processes in which several decision-makers are involved, having divergent interests, even conflictual, and taking part in a more or less direct way to the final decision.

Territorial problems, by their spatial complex nature, require the identification of several criteria and involve several decision-makers with conflicting interests whose different points of view must be taken into account for the group decision. This decision framework (multiple decision-makers and multiple criteria) constitutes what is known in the specialized literature under the denomination of multi-criteria group decision constituting a realistic framework for formulating and resolving complex decisional problems.

The main contribution proposed in this study is in the field of collaborative decision-making support. In this paper, we are interested in the development of a new model of a multi-criteria group decision support system based on the EASS (Environment as Active Support for Simulation) model [2] and refers to the system SIGMAS [10], dedicated to the problematic of locating a surface that best meets certain criteria for a given construction, for example the location of an infrastructure, such as an administrative building, a factory, a school, etc. The proposed model, called Active Group Decision Support System (ActiveGDSS), addresses a multi-criteria multi-decision-maker decision problem, the multi-criteria aspect is ensured by the performance matrix and the preferences of the decision-makers and the multi-decision-maker aspect is ensured by the weights. Active GDSS goes through two main steps: modelling step using the principle of contextual activation and the Property-based Coordination (PbC) principle. The second step is the negotiation which ensured by the negotiation protocol to arrive at a group decision that satisfies all decision makers. The Property-based Coordination (PbC) principle [21] represents the components of the multi-agent system using symbolic descriptions which organize these descriptions and manage their processing to meet coordination needs. The originality of our study is to integrate the principle of contextual activation [2] in the modeling of the system which allows to specialize the activation of each agent according to their needs and consists in reifying the link between the activation context of an agent and the behavior of the agent in reaction to this context, making the environment as an active entity of the system. The main advantages of contextual activation are efficiency gain in terms of execution, and better flexibility and reuse of agent behaviors.

The major contributions of this study are:

1)
The representation of the multiplicity and the diversity of the involved decision makers.
2)
The representation of the multiplicity and the diversity of the identified criteria.
3)
The proposition of a complete modeling of the system, by taking into account on one hand the scheduling of tasks of the agents and on the other one the link between the agents and the global scheduler of the simulation.
4)
The proposition of a new protocol of negotiation.

The paper describes in Section 2 the main research work in group decision support system. All the concepts and technics used in our work, an overview and the functional description of the proposed group decision support system as well as the characteristics of the negotiation protocol adapted by our model and the modeling of the module MAS (Multi Agent System) are detailed in Section 3. The simulation process is described in detail in Section 4 and Section 5 presents the process of ActiveGDSS. A case study is discussed in Section 6, and Section 7 is devoted to comparison with current research work. Finally we conclude our proposal in Section 8, by giving some perspectives.
2. Literature review

Traditional decisional models adapted to the single decision maker case are not consistent with the organizational reality. However the collective decision support treats processes in which multiple decision makers are involved. According to Smoliar and Sprague [17], decision processes in organizations, usually, involve several actors interacting with each other.

Indeed, many contributions in the area of group decision support and multi-agent systems have been published, we can quote in a non-exhaustive way the works of Cao and Burstein [3] and Liu et al. [12], which have integrated the intelligent aspect into the development of group decision support system.

The work of Hamdadou and Libourel [8], concern the development of a multi-criteria group decision support system for industrial diagnostics (DIAG-GDS) which is a collective decision-making tool for choosing the most relevant diagnostic method and a multilateral negotiation protocol, combined with the multi-criteria method ELECTRE III. The main limitation of this work is that the negotiation protocol is not deterministic; in [9], the authors proposed a methodological approach for an interactive decision-making tool multi decision-makers based on multilateral negotiation, multi-criteria analysis and game theory by combining geographic information systems and multi-agent systems. The proposed approach uses spatial information and its negotiation protocol is deterministic; the authors in [10] propose a spatial group decision support system multi-criteria (SGDSS) to carry out the process of spatial localization in territory planning. In this work, the proposed system does not address the distribution aspect and it is not web based. All these three systems integrate in various levels multi-criteria analysis tools coupled with GIS (Geographic Information System).

The work of Abdelhadi et al. [1], the authors establish a communication platform for a group decision support system based on web services, incorporating a multi-criteria analysis methods and a negotiation protocol. The agents of this system are simple and responsive; in [15], who propose a new group decision making framework based on three areas of high importance: multi-agent systems, geographic information system and multiple-criteria decision analysis. The proposed decision support system is enriched with a protocol of negotiation based on argumentation approach; the author in [13] design a group decision support system endowed by a negotiation protocol based on the argumentation and also propose a communication model between the different agents named ARG_GDSS (Argumentative Multi Agents System). In this study, the intelligent aspect is supported and the agents are cognitive with knowledge base.

Various works exist in the literature concerning the modeling of the system of decision support group, we quote mainly: the work of Sharma and Virmani [16], who proposed a decision support system for the detection of renal diseases using Grey Level Co-occurrence Matrix (GLCM) statistical features and a Support-Vector Machine (SVM) classifier from ultrasound images; In [7], the authors present a new conceptual framework for GDSS based on multi-agent systems using an application server constituting the communication center of the whole system and a data server that stores and manages all models, data, and decision-making methods. The authors studied and analyzed the communication mechanism and define a set of basic behaviors that represent the messages exchanged between the agents. The intelligent aspect is supported in this work; in [20], the author’s contribution concerns the elaboration of a model for group decision making with hybrid intuitionistic fuzzy information. A TOPSIS-based (Technique for Order of Preference by Similarity to Ideal Solution) method is used in group decision making. The model deals with hybrid intuitionistic fuzzy information and has not considered the weight of the decision makers; the work presented in [18], presents a new model for group decision-making within organizations is presented. It is based on a consensus relationship, allowing identification of the most promising alternatives to be selected in terms of partial pre-orders given by decision-makers. This model is devoted to the selection of alternatives by the board of any organization and it does not support the intelligent aspect of agents; in [4], the authors proposed a set of behavior styles that can be used to model agents that can represent decision-maker’s intentions in the context of group decision-making. They also proposed a communication model that simulates the dialogues made by decision-makers in face-to-face meetings. The authors did not define a scheduling policy of the simulation and the scheduling of the agents’ behaviors.

Our purpose is to elaborate a new model of a multi-criteria group decision support system based support system based on the protocol of monotonic concession. Our work refers to the SIGMAS system proposed in [10] using the principle of contextual activation [2] for modeling the system which makes the environment an active entity. The author in [2], propose a multi-agent simulation model, EASS (Environment as Active Support for Simulation) which aims to integrate the simulation process in the modeling of the system, outsource the evaluation of the local context to each agent and the selection of the behavior based on this evaluation in a central entity which is the environment, while keeping the properties of autonomy, reactivity and proactivity specific to the multi-agent paradigm.

3. Overview of the proposed model active group decision support system (ActiveGDSS)

Before presenting the proposed model we will present the fundamental concepts relating to our contribution.

3.1 Fundamental concepts

This section aims to define the main terminologies used in relation to our study.

3.1.1 Principle of contextual activation

The contextual activation specialize the activation of each agent according to their needs. The designer must then, during the modeling phase define the behaviors of agents and, independently, identify the contexts activation of the agent. Thereafter, he should link dynamically the behaviors and the activation contexts.

The fact of reifying the link between the context of the agent and its behavior, then to outsource it in the environment which does not modify the autonomy of the agent because the agent is in charge of managing its own activation links (add/delete) and the environment only treats the different links of agents according to the scheduling policy [2].

3.1.2 The principle of property-based coordination (PbC)

The purpose of the PbC principle [21] is to represent the components of the multi-agent system using symbolic descriptions. The goal is to organize these descriptions and manage their processing to meet coordination needs.

The coordination needs refer here to the activations in order to share and make accessible some of the information relating to the system components for evaluate the contexts that will determine the action to be performed by the agent. Two categories of symbolic descriptions are defined. The first category represents the observable description of the components of the multi-agent system (agents). Thus an agent individually has his knowledge and his internal process, but he makes available in the environment a set of information characterizing it.

The second category of symbolic descriptions corresponds to the abstract components and more specifically to the coordination components. Such a component is a set condition that the current state of the system must satisfy to trigger an action on the part of the agent. These coordination components are called logical filters that allow clarifying the contextual relations between activation and symbolic descriptions of the system components [2].

3.1.3 The environment model

The Formalization of the environmental model is using the Symbolic Data Analysis [5], for data expressivity and classification tools for treating a large amount of information.

The multi agent system is defined as follows: $\textit{MAS}=<\Omega_{r},E>$ where $\Omega_{r}$ is the set of real components of the system (agent) and $E$ corresponds to the environment that contains the descriptions of the real components from their observable properties and the logical filters.

Definition 1. Environment: The environment $E$ is a triple $<\Omega_{m}$ ; $F$ ; $IP>$ [2], where

$\Omega_{m}$ : The set of $n$ entities $\Omega_{m}=\{\omega_{m1};\omega_{m2};:::;\omega_{mn}\}$ . These descriptions correspond to the observable part of the whole components of the multi-agent system $\Omega_{r}=\{\omega_{r1};\omega_{r2};:::;\omega_{rn}\}$ . Thus, $\omega_{r1}$ is an agent, and this latter has a description $\omega_{m1}$ in the environment. $F$ : The set of $k$ logical filters $F=\{f_{1};f_{2};:::;f_{k}\}$ . $I P$ : The interval of priorities.

3.1.4 The components of the system

The system is composed of a set of agents. To each of these agents, a description is associated in the environment according to its observable properties. At a given moment, the state representing the internal dynamics of the agent and its private properties is called internal state, as opposed to the visible state which corresponds to the observable properties of the agent. In Fig. 1, the environment contains a matrix where each row is the complete description of an entity and each column is identified by a property which defines a particular characteristic of the entity.

Figure 1.

Relation between the agents and their observable properties in the environment.

A property $p_{j}$ is an application which associates a value corresponding to a description field $j$ to an entity $\omega\in\Omega_{m}$ . The domain of description $D_{j}$ of the property $p_{j}$ may be quantitative or qualitative. The description $d_{\omega}$ of an entity $\omega$ is composed of a vector of pairs $<\textit{nameproperty};\textit{Value}>$ [2].

3.1.5 The logical filters

The filters express the needs and interests of the agents in terms of coordination. A filter must on one hand express a set of constraints which condition the interaction, and on the other hand to be associated with an action depending on the interaction.

The principle is that the action associated with the filter is executed if the filter is triggered, i.e., if the conditions relating to the filter are verified [2]. $\textit{Cond}_{a}$ is a set of conditions on the agent and $\textit{Cond}_{C}$ is a set of conditions on the context, the activation filter $\textit{fact}(a;C)$ is triggered if $\textit{Cond}_{a}\wedge\textit{Cond}_{C}\wedge\textit{Cond}_{t}$ are verified.

Definition 2. Function valid: The function valid corresponds to the triggering of the filter and executes the action associated to the filter. This function takes in parameter a coordination filter and returns the action associated to the filter. The function valid executes if and only if:

•
$\textit{Cond}_{a}=$ True, i.e., all conditions on the agent are verified;
•
$\textit{Cond}_{C}=$ True, i.e., all conditions on the context are verified;
•
$\textit{Cond}_{t}=$ True, i.e., the constraint on time is respected.

When an activation filter is triggered, it activates at the level of the agent the generic action “act” which corresponds to the local scheduler of each agent. This action is implemented by all the agents of simulation. Its parameters are the simulation time and a label corresponding to the behavior associated to this context. Execution of the behavior comp of an agent $a$ triggered by an activation filter fact is written: $\textit{valid}(\textit{fact}(a;C))=\textit{comp}_{a}(C;t)$ where $t$ the simulation time during the triggering of the filter [2].

Definition 3. Label: A label is a triple (id; context; time) where id is the identifier of the label and context is the activation context. The parameter time corresponds to the activation time of the agent, i.e.,to the simulation time during which the filter was triggered [2].

Definition 4. Activation filter: An activation filter $\textit{fact}\in F$ is a tuple $<f_{a};f_{c};\textit{Name};\textit{Priority};\textit{Owner};\linebreak\textit{% PotentialAgents}>$ , where

$f_{a}$ : is the description in intension of the agent concerned by the filter such as:

$\displaystyle a\in\Omega_{m};f_{a}(a)=\wedge p_{j}\in P_{fa}[p_{j}(a)R_{pj}^{a% }V_{pj}^{a}]$

$f_{c}$ : is the description in intension of the context such as:

$\displaystyle C\in\Omega_{m},f_{C}(C)=\wedge\omega\in C\,{as}_{\omega}(\omega)% ,\text{ and }as_{\omega}(\omega)=\wedge p_{j}\in P_{as\omega}[p_{j}(\omega)R_{% pj}^{\omega}V_{pj}^{\omega}]$

Name: is the name of the filter. Priority: is the priority of the filter. Owner: is the agent who has filed the filter in the environment. He also allocates him a “name” and a “priority” in the intervall IP which is defined by the designer of the MAS. PotentialAgents: is a set of agents who adhere to this filter.

The description of the agent concerned by the filter is an assertion based on the properties $P_{fa}$ which is the set of properties on which the tests relate. For each property $p_{j}$ belonging to $P_{fa}$ , $R_{pj}^{a}$ is an operator of comparison and $V_{pj}^{\omega}$ is a value [2].

A filter is split into an assertion describing an agent concerned by the activation and a context in which the agent is activated. The assertion has the following general form:

Definition 5. Assertion as: $as(\omega)=[P_{j1}(\omega)R_{j1}V_{j1}]\wedge:::::\wedge[P_{jm}R_{jm}V_{jm}]$ For $1\leqslant j_{1};:::;j_{m}\leqslant r$ , where $R_{j}$ is the operator of comparison between the associated value to $P_{j}(\omega)$ and description value $V_{j}$ [2].

An assertion is a conjunction of evaluations whose final result must be true, so as this assertion will be verified. For example, the assertion $as_{1}(\omega)=[\textit{Role}(\omega)=$ “decision-maker” $]\wedge[\textit{weight}(\omega)\geqslant 2]\wedge[\textit{weight}(\omega)% \leqslant 4]$ describe the individuals that have the role decision-maker and their weight is including in the interval [2, 4]. By default, an empty assertion $\emptyset$ is satisfied by all the descriptions because it expresses the absence of conditions on these descriptions. If the description in intension of the agent is equal to $\emptyset$ , then all agents are concerned by the filter [2].
3.2 The proposed model ActiveGDSS

The proposed model AciveGDSS refers to the SIGMAS system [10] which simultaneously uses two models, a territory model and an agent’s model and is endowed with a monotonic concession protocol. Figure 2 shows the different components of the proposed group decision support system.

3.2.1 The territory model

The territory model exploits the Geographic Information System (GIS) and allows knowledge management of the territory. Thanks to its features [10], it is possible to: Manage and handle the geographic database and allow archiving spatial information, provide a spatial representation of the studied systems and format and visualize data. We exploit GIS capabilities to prepare the necessary inputs for decision-making.

Note: In our work, we assume that the inputs are already prepared by the use of a GIS

3.2.2 The agent’s model

The agent’s model is responsible for creating a real world representation of agents. Each of these agents represents one of the actors involved in the decision-making process, and disposes also objectives and preferences of the agent that represent it.

Table 1
Description of agent properties

Property	Description
Id	Unique identifier/id $\in{N}$
Weight	Importance of each decision-maker in the group decision
Role	Each agent plays a role facilitator or decision-maker
Time	Each agent has its own internal time/time $\in{N}$

Figure 2.

The proposed group decision support system.

In the literature, there is a multitude methodology offering a certain interest for the study of MAS from an organizational point of view. Our agent modeling is based on the Aalaadin methodology [6], which relies on the concepts: agent, group and role to represent a real organization. Each agent has a set of properties that provides in the environment (Table 1).

We put ourselves in a context where the actors are geographically dispersed. There are two types of actors: facilitator and decision-maker.

The facilitator: is responsible for the creation of all participating agents concerned by the decision in territory planning (TP), the proper unrolling of the negotiation and the final choice of the elected resource (islet). The decision-maker: Every actor is involved in territorial decisions. The system consists of at least two decision-makers. Those are the decision-makers experts who make up the group, each of these decision-makers has a weight and its own preferences concerning resources. Each decision-maker has its subjective parameters: the criteria weight, the preference threshold and the indifference threshold.

When decision-makers identify the actions (The virgin islets) and criteria, a value is assigned to each criterion using the analytical capabilities of GIS. All the actions and their notes on various criteria constitute the matrix of performances managed by the GIS. It is a matrix with two inputs in which, each row represents an action, and each column a criterion. The intersection of a row $i$ with a column $j$ represents the judgment of the action $i$ with respecting the criterion $j$ , so each action is judged with respect to each of the criteria.

3.2.2.1. Characteristics of the negotiation protocol

The negotiation phase find a common agreement that satisfies the majority of decision-makers agents, for this, a monotonic concession protocol was used, which involves a facilitator agent and the set of decision-makers.

The negotiation threshold is the percentage from which the negotiation is judged as success and stops by that fact. Generally this threshold is set at 80%. The value compared to the threshold is a percentage calculated from the number of decision-makers having accepted the proposal on the total number of decision-makers. As an example, if we have 4 decision-makers and only 3 of them have accepted the proposal, then we will have the value “3/4 $=$ 75%”, then this result is compared to the threshold, 75% $<$ 80% therefore we move on to the next round with a new proposal. The rounds continue until the threshold is reached or exceeded, and if it is not and the resources are exhausted, then the negotiation is a failure.

Note: simulation activity is used to start the negotiation process.

3.2.2.2. The primitives and strategies of the negotiation

In order to lead the negotiation process to its term, it is necessary to define specific primitives to the facilitator and other specific primitives to the decision-makers.

The messages sent by the decision-maker are solely aimed at the facilitator.

There are three primitives associated to the facilitator:

REQUEST(): The facilitator sends a message to the decision-makers to initiate the negotiation, each agent must associate to each resource in its vector of preference a rank, and the first ranked resource at the level of each decision-maker will have the highest value. This value is, each time, decremented by 1 for the following resources. PROPOSE(): The facilitator proposes a contract to the decision-makers concerning a given resource; this action comes from the method BORDA [19] that classifies actions from the best to the bad. CONFIRM(): The facilitator sends a message to all agents to inform them that the negotiation has been a success and that the resource has been found.

The decision maker primitives are defined as follows:

INFORM(): The decision-makers indicate to the facilitator that it can make for them a first proposal. ACCEPT(): The decision-makers indicate, by this message, to the facilitator that it accepts the contract. CONCEED(): The decision-makers indicate that it renounces its right to vote. REFUSE(): The decision-makers indicate to the coordinator that its propositions are refused.

The different strategies of the facilitator are cited as Strategies 1–3.

Strategy 1. Sending the performance matrix: From the GIS module, the facilitator recovers the performance matrix and sends it to all the decision-makers in order to determine their preferences. The facilitator must test the size of the performance matrix before sending it to the decision-makers:

If the size $<=$ 100: the facilitator sends the performance matrix to the decision-makers and these latter execute a multi-criteria method; If the size $>$ 100: in this case, the facilitator must execute the method ELECTRE_TRI to sort the performance matrix, we get three categories: good, average, bad. Then it sends the decision makers just the matrix of the category “good”.

Strategy 2. Proposition: When there are not enough decision-makers to accept the proposal of the facilitator, this latter must modify its contract for the next round and this by drawing inspiration from all modifications sent by decision-makers in round t, in order to find a new possibility for the contract. To do this, the facilitator associates a method of voting BORDA [19], which is a very efficient method used for scoring, it is among the best methods of voting that exist. Strategy 3. The negotiation stop: If the facilitator has exhausted all actions initially identified, then we are facing a conflict, so it sends as a proposition the action having the greatest value of utility. Therefore, this protocol never ends with a failure in the negotiation process.

There are three strategies associated to each decision-maker as Strategies 4–6.

Strategy 4. Establishing preferences: When the decision-makers receive the performance matrix from the facilitator, each of the agents will be able to express their preferences concerning the resources thanks to the multi-criteria method PROMETHE II which deals with the problematic of ranking and builds what is called a preference vector that contains resources ranked from the best to the worst. When each decision-maker has established its preference vector, it associates with each resource a row. The resource classified first will have a higher row representing the preference of the decision-maker at the first round. This ranking is, each time, decremented by 1 for the following resources. Strategy 5. Response: The negotiation can proceed in several rounds, until a compromise is found. In each new round, the decision-maker receives a new proposal. According to the classification of the proposal in its preferences vector, it sends an acceptance, a refusal or a concession:

Strategy of acceptance: A decision-maker accepts when the proposal received from the facilitator is classified in the first half of its preferences vector and its value of utility is greater than that provided by the facilitator. Strategy of concession: A decision-maker concedes when the proposal received from the facilitator is classified in the first half of its preferences vector and its value of utility is less than that provided by the facilitator. Strategy of refusal: A decision-maker refuses when the proposal received from the initiator is classified in the second half of the preferences vector.

Strategy 6. The negotiation end: In this protocol, the negotiation always ends with success. Negotiation stops since the reception of the CONFIRM message is received from the facilitator.

The negotiation unrolls in several rounds until a compromise that satisfies the majority of agents is found. The facilitator makes a proposal to the decision-makers concerning a given resource given, these latter will accept or concede or refuse the proposal. The facilitator’s strategy allows him to modify his proposal if there are not enough decision-makers to accept the proposal, while the strategy associated to the decision-makers allows them to accept the proposal of the facilitator or to refuse it or to concede. We chose the sequence diagram to represent the interactions and messages between the facilitator and the decision-makers (Fig. 3).

Figure 3.

UML sequence diagram associated with the negotiation protocol, where m is the number of agents involved in the negotiation, r is the number of agents who are unable to accept/concede or reject the proposal in a given iteration.

3.2.3 Modeling the activity of an agent

The current activity of an agent is modeled using an automaton, called the automaton of behavior. Each state of the automaton is a reference to a behavior that is a consistent sequence of actions and each agent has its own scheduler that manages the automaton of behavior to specify the action to execute. When an activation filter is triggered, the environment actives for the agent concerned, the internal scheduler of the agent by specifying the identified context and the actions associated with that context using the generic action act which corresponds to the call of the scheduler of the agent. To execute the appropriate behavior, each behavior of the agent is associated with a label that is passed as a parameter of the action act.

Indeed, the same behavior can be activated from different filters and therefore from different contexts. Conversely, the same filter can be used for several behaviors, allowing an agent to modify its reaction to the same context [2].

3.2.4 Formalization of the activity of the agents

An agent is constituted of a set of behaviors ( $\textit{behavior}_{i}$ ) and each behavior corresponds to an activation context ( $\textit{context}_{j}$ ) and a sequence of actions which represents the actions related to the behavior ( $L_{\textit{action}}$ ).

Definition 6. An agent is written:

agent $=$ (id; $\text{L}_{\text{behavior}}$ ) where id: agents identifier $\text{L}_{\text{behavior}}=(\text{behavior}_{1};\text{behavior}_{2};\ldots;% \text{behavior}_{\text{n}})$ $\text{behavior}_{\text{i}}=(\text{name};\text{L}_{\text{context}};\text{L}_{% \text{action}})$ Name: behavior name $\text{L}_{\text{context}}=(\text{context}_{1};\text{context}_{2};\ldots;\text{% context}_{\text{j}})$ $\text{context}_{\text{j}}=\wedge\text{cond}_{\text{j}}\,\,\text{j}=1\ldots% \text{p}$ $\text{L}_{\text{action}}={\{}(\text{action}_{1},\text{priority}_{1});(\text{% action}_{2},\text{priority}_{2});\ldots;(\text{action}_{\text{n}},\text{% priority}_{\text{n}}){\}}$

The facilitator behaviors are defined as follows:

•
UpdateTime: Is the default behavior that increment the internal time of the agent.
•
SizeMPTest: The facilitator must test the size of the performance matrix and send a REQUEST message containing the performance matrix to the decision-makers.
•
EstablishAproposal: The facilitator executes the Borda method and sends a PROPOSE message containing the resource with the highest score.
•
ThresholdTest: Upon the receipt of all decision-makers responses, the facilitator must test the negotiation threshold and then send either a CONFIRM/PROPOSE message to all decision-makers.

The different activation contexts of the facilitator are cited as follows:

•
VerificationMP: Verification of the performance matrix.
•
ReceptionINFORM: Reception of INFORM message.
•
ReceptionACCEPT: Reception of ACCEPT message.
•
ReceptionREFUSE: Reception of REFUSE message.
•
ReceptionCONCEED: Reception of CONCEED message.

The regrouping of contexts with the associated behaviors is represented by Fig. 4 for the facilitator.

Definition 7. The facilitator: The facilitator is defined by:

Facilitator $=$ (id; L ${}_{\text{behavior}}$ ) where id: facilitator identifier L ${}_{\text{behavior}}$ $=$ (SizeMPTest; EstablishAproposal; ThresholdTest) SizeMPTest $=$ (“SizeMPTest”; L ${}_{\text{contextTMP}}$ ; L ${}_{\text{actionTMP}}$ ) L ${}_{\text{contextTMP}}$ $=$ (VerificationMP) VerificationMP $=$ [Role (a) $=$ “Facilitator”] $\wedge$ [time (a) $<$ t ${}_{\text{E}}$ ] L ${}_{\text{actionTMP}}$ $=$ {(TestMP(), 2); (SendREQUEST(), 1)} TestMP(): Test the size of the performance matrix. SendREQUEST(): Send a REQUEST message to the decision-makers EstablishAproposal $=$ (“EstablishAproposal”; L ${}_{\text{contextEP}}$ ; L ${}_{\text{actionEP}}$ ) L ${}_{\text{contextEP}}$ $=$ (ReceptionINFORM) ReceptionINFORM $=$ [Role (a) $=$ “Facilitator”] $\wedge$ [time (a) $<$ t ${}_{\text{E}}$ ] L ${}_{\text{actionEP}}$ $=$ {(ExecuteBorda(), 2); (SendPROPOSE(), 1)} ExecuteBorda(): Execute the method of scoring Borda SendPROPOSE(): Send a PROPOSE message to the decision-makers ThresholdTest $=$ (“ThresholdTest”; L ${}_{\text{contextTS}}$ ; L ${}_{\text{actionTS}}$ ) L ${}_{\text{contextTS}}$ $=$ (ReceptionACCEPT; ReceptionREFUSE; ReceptionCONCEED) ReceptionACCEPT $=$ [Role (a) $=$ “Facilitator”] $\wedge$ [time (a) $<$ t ${}_{\text{E}}$ ] ReceptionREFUSE $=$ [Role (a) $=$ “Facilitator”] $\wedge$ [time (a) $<$ t ${}_{\text{E}}$ ] ReceptionCONCEED $=$ [Role (a) $=$ “Participant”] $\wedge$ [time (a) $<$ t ${}_{\text{E}}$ ] L ${}_{\text{actionTS}}$ $=$ {(TestNbResponse(), 3); (TestThreshold(), 2); (SendResponseF(), 1)} TestNbResponse(): Test the number of decision-makers answers. TestThreshold(): Test the negotiation threshold. SendResponseF(): Send a CONFIRM message if the negotiation threshold is reached otherwise it sends a PROPOSE message to the decision-makers.

The set of the activation filters concerning the facilitator agent are defined in Table 2.

Table 2
Description of the activation filters of the facilitator

Id $f_{a}$ $f_{c}$ Name Priority Owner Potential_agents

$f_{1}$ [Role (a) $=$ “Facilitator”] [Role (a) $=$ “Facilitator”] SizeMP test 1 Facilitator id {Facilitator id}

$\wedge$ [time (a) $<$ t ${}_{\text{E}}$ ]

$f_{2}$ [Role (a) $=$ “Facilitator”] [Role (a) $=$ “Facilitator”] Establish aproposal 2 Facilitator id {Facilitator id}

$\wedge$ [time (a) $<$ t ${}_{\text{E}}$ ]

$f_{3}$ [Role (a) $=$ “Facilitator”] [Role (a) $=$ “Facilitator”] Thershold test 3 Facilitator id {Facilitator id}

$\wedge$ [time (a) $<$ t ${}_{\text{E}}$ ]

Figure 4.
The Relation between facilitator behaviors and his activation contexts.

The current activity of an agent is modeled using an automaton, called the behaviors automaton. The arcs of the automaton can evolve according to the addition or removal of activation filters that are contained in the environment. The addition (respectively the deletion) of a filter in the environment results in a new link at the automaton level that allows to achieves the behavior associated to the filter (respectively the removal of a filter, removes the agent for a link to the related behavior). Each state of the automaton is a reference to a behavior that is a coherent sequence of actions. The transition from a state to another of the automaton corresponds to a particular context [2]. Figure 5 shows the automaton of behaviors of the facilitator.

Figure 5.
Automaton of the behaviors of the facilitator.

All the decision-makers have the following behaviors:

•
UpdateTime: The default behavior that increment the internal time of the agent.
•
Arrangeactions: Each decision-maker generates the preference vector using a multi-criteria method and sends an INFORM message containing this preference vector.
•
PropositionTest: Each decision-maker divides its vector into two categories so that he can do the test.
•
Inactive: At the end of the negotiation the decision-maker becomes inactive.
•
Exit: Each agent can leave the negotiation.

The different activation contexts of the decision-maker are defined as follows:

•
Reception REQUEST: Reception of REQUEST message.
•
Reception PROPOSE: Reception of PROPOSE message.
•
Reception CONFIRM: Reception of CONFIRM message.

The regrouping of contexts with the associated behaviors is represented by Fig. 6 for the decision-maker.

Figure 6.
The Relation between decision-maker behaviors and his activation contexts.

Definition 8. The decision-maker: The decision-maker is defined by:

Decision-maker $=$ (id; L ${}_{\text{behavior}}$ ) avec: id: Decision-maker identifier L ${}_{\text{behavior}}$ $=$ (ArrangeActions; PropositionTest; Inactive; Exit) ArrangeActions $=$ (“ArrangeActions”; L ${}_{\text{contextRA}}$ ; L ${}_{\text{actionRA}}$ ) L ${}_{\text{contextRA}}$ $=$ (ReceptionREQUEST) ReceptionREQUEST $=$ [Role (a) $=$ “Participant”] $\wedge$ [time (a) $<$ t ${}_{\text{E}}$ ] L ${}_{\text{actionRA}}$ $=$ {(ExecuteMMC(), 2); (SendINFORM(), 1)} ExecuteMMC(): Execute the multi-criteria method PROMETHE II SendINFORM(): Send a INFORM message to the facilitator PropositionTest $=$ (“PropositionTest”; L ${}_{\text{contextTP}}$ ; L ${}_{\text{actionTP}}$ ) L ${}_{\text{contextTP}}$ $=$ (ReceptionPROPOSE) ReceptionPROPOSE $=$ [Role (a) $=$ “Participant”] $\wedge$ [time (a) $<$ t ${}_{\text{E}}$ ] L ${}_{\text{actionTP}}$ $=$ {(Test(), 2); (SendResponseP(), 1)} Test(): Test the proposal. SendResponseP(): Send a response (ACCEPT, CONCEED or REFUSE) to the facilitator. Inactive $=$ (“Inactive”; L ${}_{\text{contextinactive}}$ ; L ${}_{\text{actioninactive}}$ ) L ${}_{\text{contextinactive}}$ $=$ (ReceptionCONFIRM) ReceptionCONFIRM $=$ [time (a) $<$ t ${}_{\text{E}}$ ] L ${}_{\text{actioninactive}}$ $=$ { } Exit $=$ (‘Exit’; L ${}_{\text{contextExit}}$ ; L ${}_{\text{actionExit}}$ ) L ${}_{\text{contextExit}}$ $=$ (ReceptionREQUEST; ReceptionPROPOSE) ReceptionREQUEST $=$ [time (a) $<$ t ${}_{\text{E}}$ ] ReceptionPROPOSE $=$ [time (a) $<$ t ${}_{\text{E}}$ ] L ${}_{\text{actionExit}}$ $=$ { }

The set of the activation filters concerning the decision-makers are defined in Table 3.

Table 3
Description of the activation filters of the decision makers

id $f_{a}$ $f_{c}$ Name Priority Owner Potential_agents

$f_{1}$ [Role (a) $=$ “Participant”] [Role (a) $=$ “Participant”] Arrange actions 1 Facilitator id {Decision-makers id}

$\wedge$ [time (a) $<$ t ${}_{\text{E}}$ ]

$f_{2}$ [Role (a) $=$ “Participant”] [Role (a) $=$ “Participant”] Proposition test 2 Facilitator id {Decision-makers id}

$\wedge$ [time (a) $<$ t ${}_{\text{E}}$ ]

$f_{3}$ [Role (a) $=$ “Participant”] [time (a) $<$ t ${}_{\text{E}}$ ] Inactive 3 Facilitator id {Decision-makers id}

$f_{4}$ [Role (a) $=$ “Participant”] [time (a) $<$ t ${}_{\text{E}}$ ] Exit 4 Facilitator id {Decision-makers id}

Figure 7.
Automaton of behaviors of a decision-maker.

Figure 7 shows the automaton of behaviors of the decision makers.
3.3 Dynamic management of activation filters

Id	$f_{a}$	$f_{c}$	Name	Priority	Owner	Potential_agents
$f_{1}$	[Role (a) $=$ “Facilitator”]	[Role (a) $=$ “Facilitator”]	SizeMP test	1	Facilitator id	{Facilitator id}
		$\wedge$ [time (a) $<$ t ${}_{\text{E}}$ ]
$f_{2}$	[Role (a) $=$ “Facilitator”]	[Role (a) $=$ “Facilitator”]	Establish aproposal	2	Facilitator id	{Facilitator id}
		$\wedge$ [time (a) $<$ t ${}_{\text{E}}$ ]
$f_{3}$	[Role (a) $=$ “Facilitator”]	[Role (a) $=$ “Facilitator”]	Thershold test	3	Facilitator id	{Facilitator id}
		$\wedge$ [time (a) $<$ t ${}_{\text{E}}$ ]

id	$f_{a}$	$f_{c}$	Name	Priority	Owner	Potential_agents
$f_{1}$	[Role (a) $=$ “Participant”]	[Role (a) $=$ “Participant”]	Arrange actions	1	Facilitator id	{Decision-makers id}
		$\wedge$ [time (a) $<$ t ${}_{\text{E}}$ ]
$f_{2}$	[Role (a) $=$ “Participant”]	[Role (a) $=$ “Participant”]	Proposition test	2	Facilitator id	{Decision-makers id}
		$\wedge$ [time (a) $<$ t ${}_{\text{E}}$ ]
$f_{3}$	[Role (a) $=$ “Participant”]	[time (a) $<$ t ${}_{\text{E}}$ ]	Inactive	3	Facilitator id	{Decision-makers id}
$f_{4}$	[Role (a) $=$ “Participant”]	[time (a) $<$ t ${}_{\text{E}}$ ]	Exit	4	Facilitator id	{Decision-makers id}

A filter may concern a set of agents, it is necessary to discuss the consequences of the addition and removal of an activation filter by an agent according to its impact on the other agents of the system.

3.3.1 The addition of new activation filters in the environment

In the model that we propose, the addition of the filters in the environment is managed by the agent “Facilitator”. To manage the agents that do not wish to belong to the list of agents potentially concerned by the filter, the author in [2], added a tuple that defines a filter, an argument Potential_agents that, during filter addition processing, lists all the agents that adhere to the filter. Thus, only the agents belonging to this list will be tested to validate the conditions on the component $f_{a}$ of the filter. This argument is a set of identifiers where each identifier characterizes an agent. If the list Potential_agents is empty then the filter will not be enabled for any of the agents. In this case the filter is not added to the environment. Three situations are then identified.

Firstly, the Addition of the filter only affects the facilitator agent (owner of the filter). This situation then requires the condition $id(\omega)=\textit{idOwner}$ which verifies that the identifier of the entity tested by the filter corresponds to the identifier of the owner of the filter, thus the list of potential agents is limited only to the owner agent, which is written (f.Potential_agents) $\leftarrow$ {Owner} in the algorithm.

The second situation concerns a condition relating to an identifier of another agent $(id(\omega)=v,v\in N)$ . Then adding the filter to the environment requires the agreement of the agent identified by the value $v$ . This validation of the addition of the filter is carried out by the environment. In the same way as previously, the list of potential agents contains the only agent determined by the constraint on the identifier after its agreement.

The last situation identifies the facilitator agent that wishes to deposit in the environment a filter with a more general scope and which can thus potentially concern a set of agents. When such a filter wants to be added, the environment executes the method agreement() for all agents who have properties necessary for context evaluation. The designer must have previously implemented this functionality in the agents to have their agreement. If the agents adhere to the filter, they are added to the list of potential agents of the filter. For example, to ask if an agent $a$ wants to adhere to a filter $f$ , the method agreement() of the agent with the filter $f$ in parameter, is called and written to $\textit{agreement}_{a}(f)$ . Thus, each agent wishing to adhere to the filter is added to the list of potential agents.

Algorithm 1 has as initial data the set of new activation filters $N_{\textit{filters}}$ to be added in the environment and returns the set of filters $R_{\textit{filters}}$ that are not added because they do not correspond to any agent. The initial condition is that the filter does not exist in the environment. This condition is verified by the environment when the addition of filter is requested. In this case, the request is not treated.

Algorithm 1: Algorithm of adding activation filters in the environment
Input: $N_{\textit{filters}}$ : Set of new filters to add
Output: $R_{\textit{filters}}$ : Set of non-add filters
Begin
$R_{\textit{filters}}\leftarrow\emptyset$
For each $f\in N_{\textit{filters}}$ do
add $=$ False
If $\exists$ $\textit{cond}_{id}\in(f.f_{a})$ Such as $\textit{cond}_{id}=[id=\textit{f.Owner}]$ Then
(f.Potential_agents) $\leftarrow$ {f.Owner}
$F\leftarrow F\cup\{\bar{f}\}$
Else
If $\exists$ $\textit{cond}_{id}\in(f.f_{a})$ and $v\in N$ Such as $\textit{cond}_{id}=[id=v]$ Then
If $\textit{agreement}_{\{a\|id(a)=v\}}(f)=$ True Then
(f.Potential_agents) $\leftarrow$ { $v$ }
$F\leftarrow F\cup\{\bar{f}\}$
Else
$R_{\textit{filters}}\leftarrow R_{\textit{filters}}\cup{\{}f{\}}$
Else
For each $a\in\textit{Cat}(f.f_{a})$ do // $\textit{Cat}(f.f_{a})$ : Regroups the set of agents that possess the
// properties necessary for the evaluation of the context
If $\textit{agreement}_{a}(f)=$ TrueThen
(f.Potential_agents) $\leftarrow$ (f.Potential_agents) $\cup\{a.id\}$
add $=$ True
If add $=$ TrueThen
$F\leftarrow F\cup\{f\}$
Else
$R_{\textit{filters}}\leftarrow R_{\textit{filters}}\cup\{f\}$
Return $R_{\textit{filters}}$
End

The complexity of Algorithm 1 is $O((n_{a}*\textit{agreement}())*n_{f})$ where $n_{a}$ is the cardinality of the category of agents of the simulation defined by $f_{a}$ and $n_{f}$ is the number of activation filters to be added in the environment. The function agreement() is configurable by the designer and specifies the cost for an agent to accept the association with a filter.

3.3.2 Removal of activation filters from the environment

Removing an activation filter for an agent involves removing the link between an action of the agent and an associated context in the environment [2]. Removing filters for agents, Algorithm 2 takes as a parameter the set $L_{id-n}$ of the couples (id, name) where id is the identifier of the agent wishing to remove its adhesion to the filter with the name name. It consists in removing from the list of potential agents of the filter $f$ , the identifier corresponding to the agent (couple.id) which is expressed (f.Potential_agents) $\leftarrow$ (f.Potential_agents) $\backslash$ {couple. id}.

The filters to be tested are identified by the condition f.name $=$ couple.name. The filter is effectively deleted when the list of potential agents is empty, which implies that the filter will never be active for any agent.

Algorithm 2: Algorithm for removing activation filters from environment
Input: $L_{id-n}$ : Set of couples (id, name)
Begin
For each $\textit{couple}\in L_{id-n}$ do
For each $f\in F$ do
If f.name $=$ couple.namethen
If couple.id $\in$ (f.Potential_agents) Then
(f.Potential_agents) $\leftarrow$ (f.Potential_agents) $\backslash$ {couple.id}
If (f.Potential_agents) $=\emptyset$ Then
$F\leftarrow F\backslash\{f\}$
End

The complexity of Algorithm 2 is $O(n_{f}*n_{s})$ where $n_{s}$ is the number of links to be removed and $n_{f}$ is the cardinality of the set of activation filters contained in the environment.

3.4 Functional descriptions of ActiveGDSS

This sub-section details the proposed model by describing each of its components: the environment component and the agent component.

3.4.1 The multi agent’s environment

Figure 8 schematizes the architecture of the environment that meets the specifications of our model. The execution module is related to the management of the execution automaton, which includes the management of the time of the simulation and the transition between the states of the execution automaton. For each category of filters there corresponds a module in which the corresponding filters are stored, namely an activation module and a dynamic module.

The dynamic module contains the two filters $f_{\textit{time}}$ which manages the updating of the time allowing to pass from time $t$ to $t+1$ when all the agents to be activated have been activated, and $f_{\textit{state}}$ which corresponds to the processing of the change of state of the simulation (See Section 4). The activation module contains all the filters related to the activation contexts of the agents.

The description module contains the symbolic descriptions of all agents and has an interaction interface that accepts the three basic actions on these descriptions which are the addition, removal or modification of a symbolic description. The activation and dynamics modules have access to the description module because the filters stored in the execution automaton module are triggered according to the descriptions and the current state of the simulation.

Figure 8.

The architecture of the environment.

The environment is continually interacting with the agents and vice versa agents perform actions on the environment. The interaction module treats the actions of the agents by directing their actions to the appropriate module. To do this, every action taken towards the environment by agents must be typed [2]. Three families of actions are identified: adding, updating or removing. Each action also relates to an object that is either a filter or a description.

The last module of the environment component is the perception module which allows agents to retrieve their local context represented by a subset of descriptions contained in the description module in the environment [2].

3.4.2 Functional description of an agent

Figure 9 defines the internal structure of an agent. The agent model is composed of three modules, the behavior module, the perception module and the interaction module, and a knowledge base.

Figure 9.

Functional schema of the agent.

The behavior module is composed of two submodules, the scheduler submodule and the behaviors library submodule. The scheduler submodule manages the agent’s behaviors automaton and allows the agent, depending on the activation of the environment, to trigger the behavior associated with the activation (information contained in the label) or to enrich his activation process with his knowledge and his perception that may have of the environment state. The scheduler submodule is activated by the environment using the action “act” that is executed when an activation filter is triggered. The behaviors library sub-module is a control structure that lists the actions that make up the behaviors and the contexts associated with them. The contexts are stored in their knowledge base [2].

The interaction module allows the agent to interact with the environment in terms of either filters additions/removals, actions to update the agent description, or additions/removals of new descriptions. The role of the knowledge base is to store the activation contexts, knowing that these contexts are used for the construction of the activation filters [2].

4. Simulation process

The purpose of this section is to describe how the simulation process is handled by describing the scheduling policy.

4.1 Global time management of the simulation

The environment is the time reference for all agents of the simulation and each agent has its own internal time. The time of the environment is discrete, and corresponds to the global time of the simulation [2].

At a given time cycle, the simulator ensures that the agents which are ready to act are activated and are activated only once per simulation cycle by comparing the global time of the simulation with the internal time of the agent at the level of each activation filter, which implies the addition of an observable property time to the description of the agent. If the value of the observable property time is greater than the time of the environment then the agent is not activated. This control presupposes that the internal time of the agent is automatically updated when it is activated [2].

To ensure the default activation of agents, the filter $f_{\textit{default}}$ is systematically added to the environment, and its triggering depends only on the comparison between the time of the simulation $(t_{E})$ and the internal time of the agent ( $\textit{time}(a)<t_{E}$ ). This filter is triggered when no other activation filter has been triggered for an agent. The goal is to update the agent’s internal time and execute default behavior for the agent [2]. In this simulation case, if no context is verified for an agent and if its internal time is less than the time of the simulation, the agent is enabled to execute the default behavior which is to update the internal time of the agent (UpdateTime) for the context VerificationTime

The incrementing of the environment time $(t_{E})$ is ensured by the simulation filter $f_{\textit{time}}$ which verifies that the internal time of the agents of the simulation is greater than or equal to the global time ( $\textit{time}(a)\geqslant t_{E}$ ). The action associated with the triggering of the filter $f_{\textit{time}}$ updates the global time $t$ to $t+1$ [2].

4.2 Control and parameterization of the simulation

In this study, two levels of scheduling listed: the global scheduling that controls the execution of the simulation, and the local scheduling that controls the behavior that the agent must execute.

At the global scheduling level, the problem is that each agent must be activated in the same cycle of the simulation. We use a state automaton, called the execution automaton, to control the execution of the simulation

A state of the automaton corresponds to a group of filters that can be triggered during the same simulation time cycle. By default, the execution automaton has a state that is responsible for the technical management of the simulation. The group of filters related to this state contains at least two filters which are $f_{\textit{default}}$ that activates an agent by default and $f_{\textit{time}}$ that allows advancing the time of the simulation.

The transitions between states of the execution automaton are triggered by modifying the value of a state variable managed by an $f_{\textit{state}}$ filter. By default, $f_{\textit{state}}$ is triggered when all filters in the concerned state are evaluated.

The activity of an agent is controlled by a priority level given to each filter; the filter with the highest priority is evaluated before a filter with a lower priority.

Figure 10 illustrates the relation between the behavior automaton associated to the agents and the automaton of execution of the simulation of our model.

Figure 10.

Behavior automaton and simulation process.

Figure 11.

Flowchart of the ActiveGDSS process.

The initialization of the simulation process consists in adding the logical filters in the different states of the execution automaton elaborated by the designer. For our model, the automaton has two states. The activation state that contains the activation filters and the state change filter $f_{\textit{state}}$ . The technical state is specific to the simulation and contains three filters: the default activation filter of an agent ( $f_{\textit{default}}$ ), the filter that manages the simulation time ( $f_{\textit{time}}$ ) and the state change filter ( $f_{\textit{state}}$ ) which allows repeating the simulation cycle.

The transition from one state to another is carried out when all the filters of the state have been evaluated and it is done by means of a variable state that is modified by the filter $f_{\textit{state}}$ which is the filter of priority the most low for each state [2]. In Fig. 10, if the variable state changes from 1 to 2, then the filters contained in the activation state are deactivated and the technical state filters become evaluable.

5. The ActiveGDSS process

In this section, we present a flowchart that illustrates the process of the ActiveGDSS. The flowchart in Fig. 11, demonstrates the sequence and approach taken by the system to conduct the collective decision support process.

6. Case study: Experimental results

This section is devoted to the presentation of a case study in territory planning to illustrate and implement our ActiveGDSS model. Indeed, our work will concentrate on the choice of the most suitable area for the construction of a habitat. The implementation of our proposed model is accompanied by an application on a test region with real data those that allows the validation of our proposal. For the development of the multi-agent module, we chose the MAS JADE platform (Java Agent DEVELOPMENT framework).

6.1 Delimitation of the area of study: Identification of resources and criteria

The area of study is situated in the canton of Vaud, to approximately 15 km in the north of Lausanne. The surface of this area is of 52.500 km ${}^{2}$ . Its geographical boundaries within the Swiss coordinate system are 532 750–532 500 (m) and 158 000–164 000 (m). For the choice of this region, we drew on the work of Joerin [11] and Hamdadou and Bouamrane [10], which mainly resulted from the large number of available spatial data. For this case study, 650 actions are proposed and seven criteria, Table 4 describes, in detail, the various criteria considered in this study.

Table 4
Description of the identified criteria

Criteria	Type	Scale	Factor (sub-criteria)	Evaluation method
Harm	Natural	[0, 1]	Air Pollution, Odors	Attribution of a note
Noise	Social	[0, 1]	Motorways, Railways	Attribution of a note
Impacts	Social	{0, …, 6}	Groundwater, sectorial plan	Procedures of space analysis
Geotechnical and Natural risks	Natural	{0, …, 6}	Constraints, Landslides, Flood, Seism, Firescriptsize	Consultation of experts
Equipment	Economic	[0, 2244]	Distance to: Gas, Electricity, Water, Roads	Balanced distances for various
				networks
Accessibility	Social	[0, 15]	The average time of trips between home and the workspace	Consultation of experts
Climate	Natural	[0, 1]	Sun, Fog, Temperature	Attribution of a note

The definition and evaluation of the criteria identified allow developing the performance matrix, shown in Table 5. This matrix is managed by the GIS component [10].

Table 5

The matrix of performances

N ${}^{\circ}$	ID_Zone	Harm	Noise	Impacts	Geotech	Equipment	Accessible	Climate
1	202	1.00	0.68	0	1	816	8	0.92
2	209	1.00	0.46	0	1	1249	9	0.91
3	210	1.00	0.69	0	1	1165	9	0.89
4	211	1.00	0.48	0	1	1518	9	0.92
5	213	1.00	0.92	0	1	1356	9	0.89
6	215	1.00	1.00	0	1	1434	8	0.75
7	216	1.00	0.97	0	1	1490	10	0.83
8	218	1.00	1.00	0	1	1556	8	0.70
9	219	1.00	1.00	0	1	1638	12	0.68
10	220	1.00	1.00	0	1	1626	8	0.68
11	221	1.00	0.95	0	1	1641	10	0.84
132	223	1.00	1.00	0	1	1697	8	0.68
13	224	1.00	0.98	0	1	1758	10	0.70
14	225	1.00	1.00	0	1	1801	8	0.67
15	226	1.00	0.91	0	1	1809	10	0.84
:	:	:	:	:	:	:	:	:
:	:	:	:	:	:	:	:	:
646	9519	0.57	0.82	6	6	1589	10	0.47
647	9525	1.00	0.98	6	6	1766	12	0.07
648	9534	1.00	0.26	6	6	1912	11	0.39
649	9548	1.00	0.14	6	6	2240	12	0.66
650	9550	0.00	0.03	6	6	2012	10	0.54

6.2 Simulation of the negotiation process

In this study, we assume that there are four decision-makers involved in this decision-making process. Each decision maker is modeled by an agent; the creation of agents is done using the JADE platform as shown in Fig. 12.

Figure 12.

Creation of agents.

Each agent corresponds to a description that is a set of properties: identifier, weight, role, and time. The set of descriptions, the added filters and the performance matrix are illustrated in Fig. 13.

At the beginning of the simulation cycle,1 the filter contained in state activation state of the execution automaton with the highest priority SIZE_MP_TEST is evaluated first. The agents activated by the latter as well as the internal time management of the agents and the global time of the simulation are shown in Fig. 14.

Figure 15 shows the end of execution of all activation and simulation filters.

The description of the actions executed of each activation filter is illustrated in Table 6.

Table 6

Description of all the actions executed of each activation filter

Actions	Description
TestMP()	Test the size of the performance matrix
SendREQUEST()	Send an REQUEST message to decision-makers
ExecuteBorda()	Executing the Borda scoring method
SendPROPOSE()	Send an PROPOSE message to decision-makers
TestNbResponse()	Test the number of decision-makers responses
TestThreshold()	Test the negotiation threshold
SendResponse()	Send an CONFIRM message if the negotiation threshold is reached otherwise send an PROPOSE
	message to the decision-makers
ExecuteMMC()	Executing the multi-criteria method PROMETHE II
SendINFORM()	Send an INFORM message to the facilitator
Test()	Test the proposal
SendResponseP()	Send a response (ACCEPT, CONCEED or REFUSE) to the facilitator

Figure 13.

Principal interface of ActiveGDSS.

Figure 14.

Execution of the first activation filter.

Figure 15.

Execution of the all activation filters.

The facilitator sends the performance matrix to the different decision-makers. Each of the decision-makers establishes his preferences vector, where he classifies resources from the best to the worst based on his subjective parameters. To achieve this objective, it uses the multi-criteria method PROMETHE II (Fig. 16).

Figure 16.

Subjective parameters and vector of preferences (ranking) expressed by each decision maker.

Before starting negotiation, it is essential to define the negotiation threshold. As soon as the participating agents receive the message Confirm(), synonymous with the end of the negotiation, the final resource has been found. The different messages exchanged during the negotiation phase are presented in Fig. 17. The message Cancel() corresponds to the primitive Conceed().

Figure 17.

Displaying exchanged messages during the negotiation process via the sniffer.

After several changes to the contract and at the end of the sixth round, decision-makers arrive at a consensus, the resource chosen is resource 537 (Fig. 18).

Figure 18.

The selected resource.

The negotiation log, in terms of executed filters (activation and simulation) is shown in Fig. 19.

Figure 19.

The journal of negotiation.

7. Comparison with related work

Research on multi-criteria group decision-making aims to develop more or less formalized models to improve the decision-making process. We can quote the work of Abdelhadi et al. [1], who proposed a model of group decision process with a new negotiation protocol based on mediation integrating a web communication platform, and the work of Oufella et al. [15], who proposed a new group decision making framework endowed with a protocol of negotiation based on argumentation approach. These two works are based on a coupling between multi agents systems and geographic information systems and use classical models in the modeling of group decision support systems.

In this literature, there are a few proposed models that have taken into account in the modeling of systems the definition of agent behaviors and the global scheduling of the simulation. In [4], the authors propose a set of behavioral styles to model agents that represent decision-makers in a group decision-making process. An agent modelled with each of these behavior styles is able to act following the intentions of the decision-maker it represents. The authors did not define a scheduling policy of the simulation and the scheduling of the agents’ behaviors. For that, we are particularly interested in the modeling and simulation of group decision support systems as an alternative to traditional models and the negotiation and participation of various actors. Thus, the modeling of the system requires not only to model the behavior of the agent but also to model the behavior of the simulation taking into account the global scheduling that controls the execution of the simulation and the local scheduling that controls the execution of the behaviors of each agent of the simulation.

The main advantages of our approach are:

•
Efficiency gain in terms of execution, and better flexibility and reuse of agent behaviors.
•
A unified approach of activations.
•
In addition, the outsourcing of the treatment of the context of the agents within the environment makes it possible to treat the global and local scheduling at the same level.

The main disadvantages are:

•
The negotiation protocol is not deterministic.
•
The functionalities of the territory model are not exploited.
•
The perception module is not developed.

8. Conclusion

In this study, we proposed a new model for the design of a multi-criteria group decision support system, which is dedicated to the problem of localization in territory planning, with a new negotiation protocol based on the concession. This model, called Active Group Decision Support System (ActiveGDSS) is based on the Property-based Coordination principle. The originality of our work is to integrate the principle of contextual activation in the modeling of the system. The main advantages of contextual activation are efficiency gain in terms of execution, and better flexibility and reuse of agent behaviors.

The perspectives of the current study concern the improvement of this paper by:

The development of the territory model by exploiting the functionalities of geographic information systems for the generation of the performance matrix;

The enrichment of the agents model by developing the contextualization aspect and the perception module;

Improving the negotiation protocol;

Adapting the proposed model to the distributed decision support systems.

Footnotes

A simulation cycle corresponds to the composition of the elementary actions executed by each activated agent during the time step.

Acknowledgments

Authors would like to thank the Directorate General for Scientific Research and Technological Development (DGRSDT), an institution of the Algerian Ministry of Higher Education and Scientific Research, for their support on this work.

Author’s Bios

Amel Kahina Nemdili received her Master degree in Information System Technologies at University of Oran 1 Ahmed BenBella, Algeria in 2013. Currently, she is a Ph.D. student in Computer Science Department at the same University. Her research interests include Business Intelligence, Group Desion Support System, Artificial Intelligence and Multi-Agents System.

Djamila Hamdadou received her Engineering degree in Computer Science and her Master of Science degree from the Computer Science Institute in 1993 and 2000, respectively. She also obtained her doctorate in 2008. She received her PHD in 2012 from the Computer Science Department. She is specialized in Artificial Intelligence, Decision Support Systems, Multi Criteria Analysis, Collaborative and Spatio Temporel Decisional Systems and Business Intelligence. She is a Professor at the University Oran 1 in Algeria where she leads the research team “Artificial Intelligence Tools at the service of Spatio-Temporal and Medical Decision Support” at the laboratory of computer science of Oran (LIO).

References

Abdelhadi

Hamdadou

and Menni

, A communication platform for group decision support system: based web services and multi-criteria method, IJESMA 10 (2018), 19–41.

Badeig

, Un environnement actif pour la simulation multi-agents: application à la gestion de crise dans les transports, Doctoral dissertation, Paris 9, 2010.

Cao

P.P.

and Burstein

F.V.

, An asynchronous group decision support system study for intelligent multi-criteria decision making, in: Proceedings of the 32nd Annual Hawaii International Conference on Systems Sciences. 1999. HICSS-32. Abstracts and CD-ROM of Full Papers, IEEE, 1999, p. 9.

Carneiro

Saraiva

Martinho

Marreiros

and Novais

, Representing decision-makers using styles of behavior: an approach designed for group decision support systems, Cognitive Systems Research 47 (2018), 109–132.

Diday

, De la statistique des données à la statistique des connaissances: avancées récentes en Analyse des Données Symboliques, In EGC, 2005, 703.

Ferber

and Gutknecht

, A meta-model for the analysis and design of organizations in multi-agent systems, in: Proceedings International Conference on Multi Agent Systems (Cat. No. 98EX160), IEEE, 1998, pp. 128–135.

Lin

and Cao

, Research on the key technology of group decision support system based on multi-agent, Journal of Computational Information Systems 6 (2010), 4897–4904.

Hamdadou

and Libourel

, A multi-criteria group decision support system for industrial diagnosis, INFOCOMP Journal of Computer Science 10 (2011), 12–24.

Hamdadou

Oufella

and Bouamrane

, Un système d’aide à la décision spatiale de groupe: Couplage analyse multicritère et théorie des jeux satisfaisants, in: In 7th Conference of Advances in Decisional Systems, Marrakech, Morocco, 2013.

10.

Hamdadou

and Bouamrane

, A spatial group decision support system: coupling negotiation and multi-criteria approaches, Intelligent Decision Technologies 10 (2016), 129–147.

11.

Joerin

, DÉCIDER SUR LE TERRITOIRE Proposition d’une approche par utilisation de SIG et de méthodes d’analyse multicritère, Doctoral dissertation, Ecole Polytechnique Fédérale de Lausanne, 1997.

12.

Liu

Qian

and Song

, Intelligent group decision support system for cooperative works based on multi-agent system, in: 10th International Conference on Computer Supported Cooperative Work in Design, IEEE, 2006, pp. 1–5.

13.

Madouri

, Vers un système d’aide à la décision spatiale de groupe: négociation par argumentation, Doctoral dissertation, University Oran 1 Ahmed BEN BELLA, 2019.

14.

Marreiros

Ramos

and Neves

, Modeling group decision meeting decision-makers with an agent-based approach, International Journal of Engineering Intelligent Systems for Elecrical Engineering and Communications-New Series- 15 (2007), 167.

15.

Oufella

and Hamdadou

, A collaborative spatial decision support system applied to site selection problems, International Journal of Applied Management Science 10 (2018), 127–156.

16.

Sharma

and Virmani

, A decision support system for classification of normal and medical renal disease using ultrasound images: a decision support system for medical renal diseases, International Journal of Ambient Computing and Intelligence IJACI 8 (2017), 52–69.

17.

Smoliar

and Sprague

, Communication and Understanding for Decision Support, in: Proceedings of the international Conference IFIP TC8/WG8.3, Ireland, 2002, pp. 107–119.

18.

Tavares

L.V.

, An acyclic outranking model to support group decision making within organizations, Omega 40 (2012), 782–790.

19.

Van Newenhizen

, The borda method is most likely to respect the condorcet principle, Economic Theory 2 (1992), 69–83.

20.

Yue

and Jia

, A group decision making model with hybrid intuitionistic fuzzy information, Computers and Industrial Engineering 87 (2015), 202–212.

21.

Zargayouna

Trassy

J.S.

and Balbo

, Property based coordination, in: International Conference on Artificial Intelligence: Methodology, Systems, and Applications, Springer, Berlin, Heidelberg, 2006, pp. 3–12.

Modeling of an active multi-agent environment for the design of a multi-criteria group decision support system

Abstract

Keywords

1. Introduction

3. Overview of the proposed model active group decision support system (ActiveGDSS)

3.1 Fundamental concepts

3.1.1 Principle of contextual activation

3.1.2 The principle of property-based coordination (PbC)

3.1.3 The environment model

3.1.4 The components of the system

3.2.1 The territory model

3.2.2 The agent’s model

Table 1 Description of agent properties

3.2.2.1. Characteristics of the negotiation protocol

3.2.2.2. The primitives and strategies of the negotiation

3.2.4 Formalization of the activity of the agents

3.3.1 The addition of new activation filters in the environment

3.3.2 Removal of activation filters from the environment

3.4 Functional descriptions of ActiveGDSS

3.4.1 The multi agent’s environment

4.1 Global time management of the simulation

4.2 Control and parameterization of the simulation

6. Case study: Experimental results

6.1 Delimitation of the area of study: Identification of resources and criteria

Table 4 Description of the identified criteria

Footnotes

Acknowledgments

Author’s Bios

References

Table 1
Description of agent properties

Table 4
Description of the identified criteria