The Santa Fe Bar Problem: a Study in Multiagent Learning

We study the process of multiagent learning in the context of the \Santa Fe Bar Problem" (Arthur 1994). We imagine a system of bounded rational agents taking decisions within a repeated game. Each agent has a nite number of decision procedures with which to reason, and access to the entire history of outcomes of the game. We rst allow the agents to use only deterministic decision procedures. We show that an eecient solution emerges, without central coordination or explicit coordination, through the dynamics and diversity of the agents. We then introduce randomized decision procedures and demonstrate that good coordination emerges more quickly, and that the long term equilibrium is more stable and less bursty.