Privacy Loss in Classical Multiagent Planning

Privacy is often cited as the main reason to adopt a multiagent approach for a certain problem. This also holds true for multiagent planning. Still, papers on multiagent planning hardly ever make explicit in what ways their systems protect their users´ privacy, nor do they give a quantitative analysis. The reason for this is that a theory of privacy loss in multiagent planning is virtually non-existent so far. This paper proposes a measure for privacy loss based on Shannon´s theory of information. To illustrate our approach, we apply this metric to an existing multiagent planning system to assess its merits when it comes to privacy on two domains. For this, we compare its plans to centrally generated solutions (by a trusted third party) for the same problems. The results clearly establish the need for such an analysis: even though the multiagent planner seemingly exchanges little information, its overall performance with respect to privacy is less than that of the centralised system (not taking into account the privacy loss with respect to the central planner, of course).