Implementing Mediators with Cheap Talk
In this work we study the effects of cheap-talk in games with rational agents. We begin by showing that all asynchronous interactions between $n$ players and a mediator can be $t$-bisimulated without the mediator if $n > 4t$. Intuitively, $t$-bisimulation means that for any deviation performed by an adversary controlling at most $t$ players in the scenario with the mediator or in the scenario without the mediator there exists an equivalent deviation in the other scenario (i.e., a deviation that produces the same outcome). We then use this result to show that any $(k,t)$-robust strategy in a game with $n$ players and a mediator in an asynchronous system can be implemented with cheap talk if $n > 4k + 4t$. A $(k,t)$-robust strategy is one in which no coalition of $t$ malicious players can decrease the payoff of anyone else and no coalition of $k$ players can increase their payoff even in coalition with the $t$ malicious players. We also prove that the result above can be satisfied if $n > 3k + 4t$ whenever honest players can punish players that are caught deviating. A similar result can be shown for synchronous systems, but in this case we only require that $n > 2k + 3t$. We also show that for every protocol $\vec{\pi}$ for $n$ players and all $k < n$ there exists a belief system that is consistent with $\vec{\pi}$ in which all coalitions $K$ of at most $k$ players believe that, if there is any deviation during the execution of the protocol, then it was someone sending an \emph{incorrect} message to a player in $K$. We call these beliefs $k$-paranoid. Intuitively, $k$-paranoid beliefs are such that all coalitions of size at most $K$ believe that the remaining players are being honest between them. We use these beliefs to extend the results regarding $(k,t)$-robustness by showing that all $k$-resilient sequential equilibria with a mediator can be implemented with cheap-talk if $n > 4k$ in asynchronous systems or $n > 3k$ in synchronous systems. We finish this work by proving a matching lower bound for most of our results.