An Experiment on Learning with Limited Information: Nonconvergence, Experimentation Cascades, and the Advantage of Being Slow

We present the results of an experiment on learning in a continuous-time low-information setting. For a Cournot oligopoly with differentiated products, a dominance solvable game, we find little evidence of convergence to the Nash equilibrium. In an asynchronous setting, play tends toward the Stackelberg outcome. Convergence is significantly more robust for a Serial Cost Sharing game, which satisfies a stronger solution concept of overwhelmed solvability. However, as the number of players grows, this improved convergence tends to diminish. This seems to be driven by high and correlated experimentation or noise and demonstrates that even when play converges, the convergence times may be too long to be of practical significance.