Case-based argumentation in agent societies

Resumen

Nowadays large systems are viewed in terms of the services that they offer and the entities that interact to provide or consume these services. Open multi-agent systems, where agents can enter or leave the system, interact and dynamically form groups (e.g. agents' coalitions or organisations) to solve problems, seems a suitable technology to implement this new computing paradigm of service-oriented systems. However, the high dynamism of open multi-agent systems requires agents to have a way of harmonising the conflicts that come out when they have to collaborate or coordinate their activities. In those situations, agents need a mechanism to argue (persuade other agents to accept their points of view, negotiating the terms of a contract, etc.) and reach agreements. Argumentation provides a fruitful means of dealing with conflicts and knowledge inconsistencies. Agents can reach agreements by engaging in argumentation dialogues with their opponents in a discussion. In addition, agents in open multi-agent systems can form societies that link them via dependency relations. These relations can emerge from agents' interactions or be predefined by the system. In addition, agents can have individual and social values, inherited from the groups that they belong to, which they want to promote. The dependencies between agents and the group(s) that they belong and the individual and social values define the agents' social context. This context has an important influence in the way agents can reach agreements with other agents. Therefore, agents' social context should have a decisive influence in the computational representation of arguments and in the argument management process. The main contribution of this PhD work is the proposal of an argumentation framework for agent societies, based on the case-based reasoning methodology. Reasoning with cases is especially suitable when there is a weak (or even unknown) domain theory, but acquiring examples encountered in practice is easy. Most argumentation systems produce arguments by applying a set of inference rules. In multi-agent systems, the domain is highly dynamic and the set of rules that model it is difficult to specify in advance. Thus, reasoning with a predefined set of rules can be difficult while tracking the arguments that agents put forward in argumentation dialogues could be relatively simple. The framework proposed allows agents to computationally represent arguments and reason about them, taking into account the agents' social context in the way agents can argue. Thus, social dependencies between agents and their individual and social values are also considered. Also, agents that comply with this framework are able to engage in argumentation dialogues following different dialogue strategies. With these strategies, agents can select the most appropriate argument to bring about their desired outcome of the dialogue. In addition, the framework proposed provides agents with individual knowledge resources to generate their positions and arguments. On the one hand, agents have a domain-cases case-base, with domain-cases that represent previous problems and their solutions. On the other hand, agents have an argument-cases case-base, with argument-cases that represent previous argumentation experiences and their final outcome. In addition, agents can accede to a set of argumentation schemes, which represent stereotyped patterns of common reasoning in the application domain where the framework is implemented. All these resources are represented by using ontologies. Thus, we have developed the case-based argumentation ontology that acts as a representation language for the knowledge resources proposed. The reasoning process that agents of our framework can use to generate positions and arguments is also defined. To allow agents to interact and control the argumentation process between them, we have designed a dialogue game protocol. Furthermore, we have proposed a set of dialogue strategies that agents can use to improve the performance of their dialogues. Finally, we have tested our proposals with two study cases. On the one hand, the formal specification of the framework has been applied to a water-right transfer domain where agents engage in argumentation dialogues to reach agreements over the allocation of water resources. This is a theoretic example where the semantic properties of the framework have been validated. On the other hand, the framework has been implemented to develop an application on the customer support domain. Here, we consider a call centre where several operators must reach an agreement to solve the incidences received by the centre.