Towards a definition of robustness for market-style open multi-agent systems

Multi-agent systems (MAS) are an integral part of current(D)AI research. Their success and the success of agenttechnology is in the majority of the literature ascribed to threemajor advantages, one of which is that multi-agent systems arerobust systems (cf. [1, 9]). Although the literature treatsrobustness like an inherent feature of MAS (which it is not),there is hardly any discussion on what robustness actuallymeans (a rare exception is the work of Kaminka and Tambe [7]and Klein and Dellarocas [3]). Just like any other artificialcomplex system, multi-agent systems need to be specificallydesigned to be robust. Compared to conventional computerscience, the issue of robustness in MAS is different. Mostcomputer science systems are transformational systems, whichmeans they compute a function on some input. Here, techniquesfor ensuring robustness exist (e.g. cf. [6]). However, the mostinteresting multi-agent systems are open systems, which donot explicitly compute a function (e.g. looking at the mostpredominant example of an open system, namely the Internet, ascomputing a function certainly misses the point).