CONCLUSION

This work presents a participative multi-agent architecture to develop knowledge production systems. Multi-agent interaction approaches and protocols are designed according to top-down or bottom-up approaches. The architecture presented in this chapter is a bottom-up approach to the design of collaborative multi-agent systems, where every mart holds responsibilities on some domain-level knowledge, while coordination-level knowledge interfaces to other domains are well-defined . This structuring of knowledge marts can help to reduce inconsistencies between agent territories .

The participative approach presented in this work has been successfully applied to the development of learning objects, but is also applicable to other knowledge production tasks (Dodero et al., 2002). Results obtained from single-mart and two-mart evaluation scenarios have been contrasted, with the result that the coordination protocol improves conflict-solving and coordination during the shared development process. Moreover, the absence of participation of some agent does not delay the overall process. Nevertheless, in order to test the multilevel architecture, these results need to be confirmed in more complex scenarios, consisting of two or more groups of participative agents working in different knowledge marts. We are also conducting tests of the impact of the number of agents on the overall effectiveness of the model. Further validation is also needed to assess the usefulness of the approach in different application scenarios, such as software development, especially in the analysis and design stages.

The structuring of heterogeneous knowledge domains into marts presents a number of issues: What would happen if an agent changes the kind of knowledge that it is producing, and is this better classified in another mart? As time progresses, will knowledge that is being produced in a mart be biased towards a different category? As a future work, it seems reasonable to dynamically establish the membership of agents into the marts, such that an agent can change its membership to some other mart if the knowledge produced by the agent affects interaction processes carried out in that mart. Then, division and/or fusion of marts may be needed to better reflect the knowledge-directed proposed structure. In that case, clustering techniques can be readily applied to solve those issues. As well, it will be helpful that mart generation and affiliation of agents to marts be dependent on agents' ontology-based interests.