Asynchrony and Learning in Serial and Average Cost Pricing Mechanisms: An Experimental Study

This paper reports the first experimental study of the serial and the average cost pricing mechanism under three different treatments: a complete information treatment and two treatments designed to simulate distributed systems like the Internet with extremely limited information, synchronous and asynchronous moves. Although both games are dominance-solvable and the proportion of equilibrium play is statistically indistinguishable under complete information, their performance does change dramatically in settings that resemble distributed systems: the serial mechanism performs robustly better than the average cost pricing mechanism both in terms of convergence to Nash/Stackelberg equilibrium and system efficiency. These results provide some support for Friedman and Shenker's (1997) new solution concepts for implementation on the Internet. Four payoff-based learning models are simulated in order to understand individual learning behavior in distributed systems. Under the serial mechanism the payoff-assessment learning model (Sarin and Vahid (1997)) provides the best fit to the data, followed by the experience-weighted attraction learning model (Camerer and Ho (1999)), which in turn, is followed by a simple reinforcement learning model and the responsive learning automata. Under the average cost pricing mechanism, both the experience-weighted attraction learning model and the reinforcement model track the data better than the responsive learning automata, however, other pair-wise rankings of the four models are statistically insignificant.