Using ILP-systems for verification and validation of multi-agent systems

Abstract

Most applications of inductive logic programming focus on prediction or the discovery of new knowledge. We describe a less common application of ILP namely veri cation and validation of knowledge based systems and multi-agent systems. Using inductive logic programming, partial declarative speci cations of the software can be induced from the behavior of the software. These rules can be readily interpreted by the designers or users of the software, and can in turn result in changes to the software. The approach outlined was tested in the domain of multi-agent systems, more in particular the RoboCup domain.

Top-down induction of first-order logical decision trees

“…This was done using a separate test set of 100,000 examples, which was the same for all the hypotheses. For the RoboCup data set no test set was constructed, because interpretability of the hypotheses by domain experts is the main evaluation criterion for this application (these theories are used for veri cation of the behavior of agents, see (Jacobs et al, 1998)).…”

“…This use of 7.4. EXPERIMENTS 155 the data set re ects the learning tasks considered until now by the people who are using it, see (Jacobs et al, 1998). It does not in uence the validity of our results for relational languages, because the propositions are de ned by the background knowledge and their truth values are computed at runtime, so the query that is really executed is relational.…”

Untitled

“…While this data set would allow rather complicated theories to be constructed, for our experiments the language bias was very simple and consisted of a propositional language (only high-level commands are learned). This use of the data set re ects the learning tasks considered up till now by the people who are using it, see (Jacobs et al, 1998). This does not in uence the validity of our results for relational languages, because the propositions are de ned by the background knowledge and their truth values are computed at runtime, so the query that is really executed is relational.…”

“…For the RoboCup data set interpretability of the hypotheses by domain experts is the main evaluation criterion (because these theories are used for veri cation of the behavior of agents, see (Jacobs et al, 1998) ).…”

Genetic Programming Method of Evolving the Robotic Soccer Player Strategies with Ant Intelligence

“…Another research has been reported, which uses ILP systems for verifying and validating multi-agents for RoboCup [14]. This concentrates on verification and validation of knowledge based system, not but prediction or discovery of new knowledge.…”