Collection Mémoires et thèses électroniques
Accueil À propos Nous joindre

Chapter 1 Introduction

Table des matières

In this chapter, we introduce the context of our research which is communication between software agents in a multi-agent system. We also identify the motivations, problems, and research questions that we address in this thesis. Finally, we present our hypotheses, objectives, and methodology.

This thesis is about communication between autonomous agents. In the multi-agent domain, it is widely recognized that this communication is one of the major topics of research. All the applications of Multi-Agent Systems (MASs) (Chaib-draa, 1995), (Wooldridge and Jennings, 1995), (Moulin and Chaib-draa, 1996) (Wooldridge, 2002) ranging from digital libraries through cooperative engineering to electronic commerce, have one thing in common: the agents operating in these systems have to communicate. These systems consist of multiple agents that communicate in order to solve some problems. If a problem is particularly complex, large, or unpredictable, the only way it can reasonably be addressed is to develop a number of functionally specific and modular components (agents) which are able to solve a particular problem aspect (Sycara, 1998). This decomposition allows each agent to use the most appropriate paradigm to solve its particular problem. When interdependent problems arise, agents in the system must communicate in order to coordinate with one another to ensure that interdependencies are properly managed. Therefore, it is clear that the success of these systems need powerful communication mechanisms.

Agent communication is related to several disciplines: philosophy of language, social psychology, artificial intelligence, logic, mathematics, etc. In this domain, in order to be able to negotiate, solve conflicts of interest, cooperate, find proofs, agents need not only exchange single messages, but also take part in conversations with other agents. A conversation is defined as a coherent sequence of utterances. The term "coherent" means that the information conveyed by an utterance is related to the information conveyed by the other utterances in a conversation. For example, if p is the information conveyed by an utterance, the information conveyed by the next one can be the acceptance, the refusal, the challenge, the attack, etc. of p . Indeed, if agents communicate by exchanging isolated messages, the resulting communication is extremely poor and agents cannot participate in complex interactions such as negotiations, persuasions, deliberations, etc, which are formed by a sequence of utterances.

The language used by the agents for their exchanges is the Agent Communication Language (ACL). An ACL stems from the need to coordinate the actions of an agent with that of the other agents. A first attempt to come to a standardized ACL came from the DARPA knowledge sharing project and produced KQML (Knowledge Query and Manipulation Language) (Finin et al., 1995). Another effort to come to a standard ACL has started through the Foundation for Intelligent Physical Agents (FIPA) initiative (FIPA, 2001a, 2002). KQML and FIPA-ACL are based on speech act theory and messages are considered as communicative acts whose objective is to perform some action by virtue of being sent. To enable agents to communicate, FIPA proposed a set of communication protocols that agents can follow. FIPA contract net interaction protocol is an example of these protocols (FIPA Interaction Protocols, 2001, 2002). In the contract net interaction protocol, one agent (the initiator) takes the role of manager which wishes to have some task performed by one or more other agents (the participants) and further wishes to optimise a function that characterizes the task. This characteristic is commonly expressed as the price, in some domain specific way, but could also be soonest time to completion, fair distribution of tasks, etc. Generally, agent communication protocols describe the sequence of messages that agents can exchange for particular applications. Although these protocols can successfully be used for some simple applications, they are often too rigid to be used by autonomous agents in their conversations. The reason is that these protocols are specified in such a way that agents must follow them from beginning to end without specifying how these agents can reason about them. To solve this problem, several researchers proposed dialogue game frameworks inspired by the philosophy of argumentation (Reed, 1998), (McBurney and Parsons, 2001, 2002), (Dignum et al., 2000, 2001). Dialogue games are abstract structures that can be composed in order to reflect the whole dialogue. They are interactions between two or more players, in which each player moves by making utterances, according to a pre-defined set of rules. The rules typically define what locutions may or must be uttered in different circumstances. However, the underlying semantics and the verification of these dialogue games are aspects yet to be addressed.

To be able to communicate, agents should use a common communication mechanism (for example, a communication protocol). Because agents are autonomous, this mechanism must be flexible enough and agents must reason about their communicative acts in order to decide how they can pursue their conversations. Classical protocols, like those used in distributed systems, are not suitable in this domain because they only describe the sequence of allowed actions without any reasoning mechanism. Our first motivation is to give agents flexible means of communication. These means must be formally specified by taking into account the agents’ architecture.

In addition, in the domain of agent communication, semantics is one of the most important aspects, particularly in the current context of open and interoperable MASs. Semantics lays down the foundation for a concise and unambiguous meaning of agent messages. When agents interact to achieve a goal, the mutual understanding of the exchanged messages depends on the semantics given to communicative actions. Although some significant research work was done in this field (Singh, 2000) (Guerin and Pitt, 2001) (Amgoud et al., 2002), (McBurney, 2002), (Verdicchio and Colombetti, 2003), (Flores et al., 2004) the definition of a clear and global semantics is an objective yet to be reached. Agent communication pragmatics is another important aspect to be addressed in this domain. While semantics is interested in the meaning of communication acts, pragmatics deals with the way of using these acts. Pragmatics is related to the dynamics of agent communication. Because agents do not exchange isolated messages, but participate in complete conversations, the semantics must take into account the chaining of the communicative actions. Therefore, the semantics must be defined in a pragmatic perspective. Our second motivation is to contribute to the advance of research in this domain by developing a unified formal framework establishing the link between the pragmatics and semantics of agent communication. This motivation is related to the first one in the sense that the formal specification of the communication mechanism should allow us to verify whether or not agents respect the defined semantics when conversing.

The first problem that we address in this thesis is the lack of flexibility in most current agent communication protocols (FIPA Interaction Protocols, 2001, 2002). These protocols are static and agents must execute them from beginning to end in order to communicate. In addition, these protocols do not specify how agents can manage exceptions (messages not specified or not supported by the protocol), and how they can choose a communicative act among others. To address this problem, several researchers proposed dialogue game frameworks. These frameworks attempt to support more complex conversations by combining different atomic dialogues. Agents participating in a dialogue game framework must agree on all the rules of the framework. However, several proposals in this domain do not specify how agents can reason about these rules and participate in conversations in a flexible way. If the decision making process belongs to the agent architecture, the link between this architecture and the communication model must be specified. In addition, these frameworks do not take into account the link between the private mental states and the reasoning abilities of agents. Thus, our initial research question is: " How may autonomous agents participate in conversations in a flexible way? "

In the literature, three main approaches have been proposed for modeling agent communication: the mental approach, the social approach, and the argumentative approach. The mental approach focuses on the agents’ private mental states like beliefs, desires and intentions. In this approach, the semantics of the communicative acts is defined using these mental states. The social approach highlights the public and observable elements like social commitments that agents exchange when conversing. A social semantics is defined using the notion of social commitment. The argumentative approach is based on the agents’ reasoning capabilities. The meaning of communicative acts in this approach is defined in terms of arguments in favor or against the content of these acts. These approaches reflect only a partial view of agent communication. When participating in conversations, agents should use their mental states, exchange observable elements, and reason about these states and elements. Therefore, these approaches should be combined in a unified approach. The question that we explore here is: " How can we unify these approaches? " Another related question is: " How can the link between pragmatics and semantics be established in such an approach? "

The third problem that we explore in this thesis is the verification of agent communication. Two fundamental aspects need to be verified when specifying and developing agent communication mechanisms: the agents’ compliance to the ACL semantics, and the correctness of the specification in the sense that the mechanism satisfies some given properties. Although this verification is extremely important in open environment and in complex and interoperable systems, the different protocols proposed in the literature (classical or based on dialogue games) do not address it. Verifying these protocols is not an easy task when considering the different states of agents and their reasoning capabilities. The question is: " How can we formally specify and verify the agent communication mechanisms? " In addition, the termination of agent conversations and the complexity analysis of the corresponding reasoning algorithms must be addressed.

To take part in flexible conversations (persuasions, argumentative negotiations, deliberations, etc.), software agents must have a suitable communication model. Agents must build their conversation dynamically while it advances. Thus, our first research hypothesis is that in their conversations, agents do not have to follow pre-established and fixed protocols. Instead, they need to reason about all utterances that have been uttered during the conversation in order to decide about what is necessary to utter next. In flexible conversations, protocols are only interesting as long as agents can use them as stereotypes which can help them in their conversations and not as means imposing what agents must do. Protocols only specify the allowed communicative acts, and do not indicate how agents can choose between these acts. In other words, protocols do not specify the underlying decision making process which is fundamental for conversing agents.

The second hypothesis is related to the importance of the conversation context. The conversation context is defined by the set of knowledge and beliefs that agents suppose they share during their conversations. For example, as members of the same cultural community, the participants in a conversation share knowledge of a general nature and knowledge related to the existing standards and conventions. We make the hypothesis that agents communicate in a particular context that they share.

The main objectives of this thesis are:

1- To propose a pragmatic approach for agent communication taking into account the different elements that agents use in their conversations. This pragmatic approach based on social commitments and arguments must illustrate how agents use their communicative acts when conversing. It must also represent the dynamics of agent communication. This approach, based on speech act theory and specified by a formal language, will be used to develop a formal framework allowing agents to take part in conversations in a flexible way. This framework, specified as a mathematical structure, should be able to represent the various actions that can take place in agent conversations and to model the dynamics of these conversations.

2- To develop a communication model and a corresponding agent architecture on the basis of the pragmatic approach.

3- To define a formal semantics related to the pragmatic approach. The idea is to specify a unified framework for the pragmatics and the semantics of agent communication. The meaning of the communicative acts must take into account the dynamics of agent communication.

4- To develop a verification method for dialogue game protocols specified using the unified framework.

5- To specify and to implement a flexible dialogue game protocol using the unified framework.

At the beginning of this thesis, we studied research work done in the domain of agent communication. We noticed that the classical protocols used in this field are not suitable in the context of MASs in which agents are autonomous. Particularly, we noted the absence of the reasoning aspect in these protocols. For this reason, we looked at the work done in another field: argumentation and defeasible reasoning. We had the idea to combine an approach proposed in the domain of agent communication, the social approach, which has the advantage of being semantically verifiable and an argumentative approach.

In addition, we noticed that the traditional formalisms used to model agent communication are limited. They do not make it possible to reflect the dynamics of this communication in terms of the actions which agents perform when conversing and do not help agents to take part in these conversations in a flexible way. We thus developed a formalism addressing these limits using our hybrid approach. The proposed approach and formalism only reflect the pragmatics of agent communication. To deal with the semantic aspect, we proposed a logical model for the pragmatic approach. Indeed, we developed a unified framework for the pragmatics and semantics of agent communication.

Although certain researchers recently started to emphasize the importance of verifying MASs, this aspect has yet to be explored in the field of agent communication. In this domain, only a small amount of research work addressed this complicated issue, for example (Wooldridge, 2000) (Huget and Wooldridge, 2004). For this reason, we studied more profoundly this aspect which is traditionally related to software engineering. We proposed a model-checking method in order to verify dialogue game protocols specified using our unified framework. Finally, as an application of our theoretical results, we specified and implemented a flexible dialogue game protocol. We proved its termination and we discussed its computational complexity.

This thesis is divided into two parts.

Part I is about the state of the art, and it consists of three chapters:

Chapter 2 introduces the agent communication. In this chapter we present some examples of Agent Communication Languages (ACLs), we discuss their semantics, and present their philosophical foundations.

Chapter 3 presents some dialogue game frameworks. In this chapter, we highlight their theoretical foundations, advantages, and limits. These limits will be addressed in our proposal.

Chapter 4 presents our taxonomy of the main approaches in the domain of agent communication and dialogue modeling. This chapter compares and discusses these different approaches.

Part II consists of five chapters in which we present our contributions:

Chapter 5 articulates a pragmatic approach combining the different approaches discussed in Chapter 4.

Chapter 6 proposes a formalism based on the pragmatic approach presented in Chapter 5. The purpose of this formalism is to represent the dynamics of agent communication, analyze conversations, and help agents to participate in conversations in a flexible way.

Chapter 7 defines the semantics related to the pragmatic approach as a logical model. This chapter establishes the link between the pragmatics and the semantics of our agent communication proposal.

Chapter 8 proposes a verification method for dialogue game protocols. These protocols are specified using the unified framework presented in Chapter 7. In this chapter, a tableau proof system for the logical model specified in Chapter 7 is defined. This proof system is used in the verification method.

Chapter 9 presents a persuasion dialogue game protocol based on our approach. This chapter discusses the formal specification, implementation, and complexity analysis of this protocol.

We conclude this thesis by summarizing our contributions and identifying directions for future work.

© Jamal Bentahar, 2005