Toward Using Symbolic Discovery in Designing Controllers of Autonomous Swarm Robots

In this paper, we propose an approach which iterates a designtest- analysis cycle using symbolic data mining methods for designing controllers of autonomous swarm robots. The approach is applicable even if the onboard signal is unavailable to the designer, which is common for such kinds of robots. As the first step, we tackle a specific task in which two swarm robots try to visit as many cells as possible in a square field before a fatal collision. Quick analysis using conventional techniques, relying also on human inspection revealed interesting essentials including a desirable type of interaction between the swarm robots and possible refinements of the controllers. We consider possible usages of data mining methods including an efficient trajectory discovery method, an effective minority subset discovery method, and a robust partial classifier discovery method.