Collection Mémoires et thèses électroniques
Accueil À propos Nous joindre

Chapter 10 Conclusion

Table des matières

In this thesis we proposed a unified framework for the pragmatics and the semantics of agent communication. Our framework has the advantage of being based on solid philosophical foundations and equipped with a logical formalization. The philosophical foundations are supplied by the philosophical definition of social commitments, Speech Act Theory and formal dialectics (the philosophy of arguments). The logical formalization is defined in terms of a combination of branching time logic (CTL*) and dynamic logic.

Another advantage of this framework lies in the fact that it captures both the pragmatics and semantics of agent interactions. We discuss these two aspects in this section.

Pragmatics : The interactions between autonomous agents are reflected by the actions that they perform on commitments and on their contents. These actions can be supported by arguments. The dynamics of the interactions is reflected by the creation of commitments, by the agents’ positioning on these commitments (acceptance, refusal, challenge, attack, etc.), and by the evolution of commitment states in time (satisfied, withdrawn, etc). All the commitments and arguments handled in an interaction can be represented using our commitment and argument networks (CAN). This formalism allows us to model the dynamics of conversations and offers an external representation of the conversational activity. This notion of external representation is very useful because it provides participants with a common understanding of the current state of the conversation and its advancement. The formalism also allows us to ensure conversational consistency when considering the actions performed by the agents. It relies on our approach combining commitments and arguments. This approach has the advantage of capturing both the social and public aspects of a conversation, and the reasoning aspect required in order to take part in conversations. Thus, the formalism can clearly illustrate the creation phases of new commitments and the positioning phases on these commitments, as well as the argumentation and justification phases.

Semantics : All the elements captured by the pragmatic aspects of our framework are semantically defined in a logical formalism combining temporal and dynamic logics (DCTL*CAN). The concept of social commitment, the different types of commitments and the concept of argument are defined as modal operators logic. The actions that agents apply to commitments and on their contents as well as the argumentation relations are defined using the Perform operator that reflects the performance of actions. The important link between commitments and arguments that we established in the pragmatic level is formally captured by the semantics in the form of properties using the other elements of the logical model. Our semantics offers a clear and unambiguous means to introduce the different elements and the various operations that we described in the pragmatic level of agent communication. It can also be used for verification purposes. A direct application is to check if a particular protocol (for example a negotiation or a persuasion protocol) respects the introduced specifications.

Our pragmatic approach presented in Chapter 5 is different from the social approach proposed by Singh (1998, 2000) and Colombetti (2000) in the sense that social commitments in our approach are not only public states but also deontic notions. Agents must justify and defend their commitments if necessary. Thanks to the link we established between commitments and arguments, agents can reason about their commitments and consequently can communicate in a flexible way. In addition, there are many differences between our approach and the argumentative approach proposed by Amgoud and her colleagues (2002a, 2000b). The main difference is that Amgoud et al.’s proposal is based upon dialectical systems, and the evolution of agent conversations is captured using the commitment stores that only record what is uttered during the conversation (MacKenzie, 1979). However, in our approach, the evolution is captured by the notion of commitment and commitment content states that evolve as a result of the actions that agents perform when conversing (creation, withdrawal, reactivation, violation and satisfaction). The main idea of our approach is that agent communication is considered as actions that agents perform on social commitments and arguments. Thus, different speech act types can be expressed using these actions.

The CAN formalism presented in Chapter 6 as the basis of our pragmatic approach allows us to represent the dynamics of agent communication in a formal way. This new formalism for agent communication is different from all other agent communication formalisms proposed in (Pitt and Mamdani, 2000), (FIPA-ACL, 2001), (Yolum and Singh, 2002) and (Fornara and Colombetti, 2003). Unlike these formalisms, the CAN formalism can be used as a means to help agents to participate in conversations. In addition, this formalism enables agents to reason about their communicative acts and about the current state of the conversation in order to decide about the next actions to be performed. This reasoning aspect is tied to the agents’ argumentation systems.

Semantically speaking, our logical model presented in Chapter 7 is different from the semantics defined by Singh (2000) and by Verdicchio and Colombetti (2003). Our semantics is based not only on a temporal logic, but also on a dynamic logic and it captures different commitment types, different commitment states and different actions performed on commitments. Our semantics is defined as a model-theoretic semantics that can be successfully used to capture the semantics of defeasible arguments. It is therefore different from the semantics defined in (Amgoud et al., 2002) which is based on an informal logic. Another difference is that our semantic framework can be used to express the meaning of different speech act types.

In addition, in Chapter 8, we proposed a new model checking algorithm for the verification of dialogue game protocols whose complexity matches that of the best existing algorithms. Our model-checking technique allows us not only to verify if the dialogue game protocol satisfies a given property expressed in our DCTL*CAN, but also if this protocol respects a simplified version of the tableau semantics of the communicative acts. To our knowledge, this model-checking technique is the first proposal in the domain of dialogue game verification.

Finally, there are many differences between our dialogue game protocol presented in Chapter 9 and the other dialogue game protocols discussed in Chapter 3. The main differences are:

1- Our proposal is based on a social and argument approach. Consequently, agents can reason about their actions in order to decide about the dialogue game to be played.

2- The decision making process is based not only on the agents’ argumentation systems, but also on the agents’ trustworthiness.

In addition, we provided a termination proof of our protocol, and we discussed its computational complexity.

The main contributions of this thesis are:

1- A formal pragmatic approach capturing the conversations’ public elements and the agents’ reasoning mechanisms using their private states for modeling agent communication. This approach was published in (Bentahar et al., 2003).

2- A formalism called Commitment and Argument Network representing the dynamics of agent communication and helping agents to participate in conversations in a flexible way. This main contribution resulted in two publications: (Bentahar et al., 2004b, 2004c). Together, contributions 1 and 2 allowed us to achieve our first and second objectives stated in Chapter 1.

3- A model-theoretic semantics for the pragmatic approach defining the meaning of the different communicative acts that we use in our pragmatic approach, especially the ones commonly used in multi-agent interactions, and capturing the semantics of defeasible arguments. This semantics resulted in two publications (Bentahar et al., 2004e, 2004f). This contribution matches the third objective of this thesis.

4- A tableau-based model checking technique for the verification of dialogue game protocols specified in our framework. This contribution is published in an internal report (Bentahar and Moulin, 2004), and it is the subject of a submitted paper (Bentahar et al., 2005). This verification method is the fourth objective that we set in Chapter 1.

5- A new persuasion dialogue game protocol specified in our framework using a logical language, and implemented using an agent-oriented programming language. This contribution that matches the fifth objective of this thesis is published in (Bentahar et al., 2004d). The algorithmic specification of this protocol in the context of the CAN framework was the subject of another publication (Bentahar et al., 2004a).

Thus, all the objectives of this thesis are reached. In addition, contributions 1, 2, and 5 answer the first research question stated in Chapter 1: " How may autonomous agents participate in conversations in a flexible way? " Contribution 3 answers the second research question: " How can we unify pragmatic and semantic approaches and how can the link between pragmatics and semantics be established in such an approach? " Finally, contributions 4 and 5 answer the third research question: " How can we formally specify and verify the agent communication mechanisms? "

As future work we intend:

1- To use our unified framework to specify other sophisticated protocols according to Walton and Krabbe’s classification. Because this framework is based on a commitment and argument approach, the dialogue types in this dialectical-based classification can be supported. An important result of this work is to explain and formalize the shift between these different dialogue types during a conversation. The idea is to define a general dialogue-game protocol combining the different protocols (the combined protocol). The rules defining the dialectical shifts can be expressed in a logical language extending the one we proposed in Chapter 8. The implementation of such a protocol can be done using the same logic-programming and agent-oriented paradigm that we used for the persuasion dialogue game.

2- To define an operational and a denotational semantics for the different protocols and for the combined protocol. The operational semantics constitutes a means to formally derive the computation steps of the protocols. The denotational semantics provides a tool for specifying the compositionality of these protocols.

3- To implement and evaluate the model checking technique proposed in Chapter 8. The ABTA for DCTL*CAN will be implemented in the Concurrency WorkBench of the New Century CWB-NC verification tool (Cleaveland and Sims, 1996). The ABTA manipulation procedure will be implemented in Standard ML. This work will be done in collaboration with Rance Cleaveland from State University of New York at Stony Brook.

4- To define a model checking technique for all the logic proposed in Chapter 7. The tableau technique proposed in Chapter 8 can be improved to support the complete version of the logic.

5- To explore other argumentation models, particularly Toulmin’s model (1958) that is widely cited in the philosophy of argumentation, but still unexplored in the domain of agent communication.

© Jamal Bentahar, 2005