Discovering Strategic Behaviors in Multi-Agent Scenarios by Ontology-Driven Mining
Davide Bacciu, Andrea Bellandi, Barbara Furletti, Valerio Grossi, Andrea Romei · InTech eBooks · 2008
We have introduced an ontology-based approach for association analysis in the context of behavior mining in multi-agent systems. Our proposal is based on the idea of the ontological description of the domain as an essential via-point for accessing the expert knowledge concealed underneath massive amounts of "flat" data. The introduction of a multi-layered and multi-relational representation of the domain allows approaching the information content from several, diverse, viewpoints. Within the multi-agent area, this approach offers considerable advantages since it allows the agents to gather personalized views of the extracted knowledge, represented by means of rule ontologies. By exploiting this "relativistic" representation, an agent can dynamically generate and selectively access, at the desired level of abstraction, the knowledge that is of higher relevance for its current decision-making activity. Besides presenting the clear advantage of offering personalized views of the world, which is totally consistent with the multi-agent model, this approach relieves the agents from the burden of acquiring, maintaining and mastering a monolithic, and encyclopedic, representation of their knowledge. The model presented in this chapter tackles the association analysis task with the standard static approach where each transaction is considered in isolation and not as part of a spatiotemporal trajectory. Sequence mining, on the other hand, studies how to approach the problem of finding frequent patterns for trajectory data. An interesting future development for our model would be to extend it to sequence mining (Agrawal & Srikant, 1995): in particular, this would be of great interest for behavioral pattern mining, since it can naturally tackle the problem of planning medium to long term strategies comprising lengthy sequences of inter-dependent actions. However, we would like to point out that our model already offers a means for processing multiple transactions in a sort of time trajectory. Through the domain ontology is possible, in fact, to describe concepts whose definition transcends the single transaction. Consider, for instance, the pass concept in the soccer ontology: since it requires to specify source and destination of the action, its instantiation needs to process multiple transactions to find all the requested information, e.g. the receiver of the pass.