Knowledge discovery for training intelligent agents: Methodology, tools and applications

In this paper we address a relatively young but important area of research: the intersection of agent technology and data mining. This intersection can take two forms: a) the more mundane use of intelligent agents for improved knowledge discovery and b) the use of data mining techniques for producing smarter, more efficient agents. The paper focuses on the second approach. Knowledge, hidden in voluminous data repositories routinely created and maintained by today's applications, can be extracted by data mining. The next step is to transform this knowledge into the inference mechanisms or simply the behavior of agents in multi-agent systems. We call this procedure "agent training." We define different levels of agent training and we present a software engineering methodology that combines the application of deductive logic for generating intelligence from data with a process for transferring this knowledge into agents. We introduce Agent Academy, an integrated open-source framework, which supports data mining techniques and agent development tools. We also provide several examples of multi-agent systems developed with this approach.