Learning Mutual Trust

ABSTRACT Multiagent learning literature has looked at iterated twoplayer games to develop me hanisms that allow agents to learn to onverge on Nash Equilibrium strategy pro les. Su h equilibrium on guration implies that there is no motivation for one player to hange its strategy if the other does not. Often, in general sum games, a higher payo an be obtained by both players if one hooses not to respond optimally to the other player. By developing mutual trust, agents an avoid iterated best responses that will lead to a lesser payo Nash Equilibrium. In this paper we onsider 1-level agents (modelers) who sele t a tions based on expe ted utility onsidering probability distributions over the a tions of the opponent(s). We show that in ertain situations, su h sto hasti ally-greedy agents an perform better (by developing mutually trusting behavior) than those that expli itly attempt to onverge to Nash Equilibrium.