Refining the breeding of hybrid strategies

We extend our earlier work using artificial agents to model multi-period games between competing brands in an oligopoly. We do so by developing a multiple-population genetic algorithm in order to allow customized agents for each brand. We also consider more competitors and more possible pricing actions per competitor than before, and we evaluate the robustness of our results by Monte Carlo methods. All these developments have been facilitated by writing better code and by increases in computing power since our original work. We find that: 〈 Customized agents outperform traditional genetic algorithm approaches; 〈 The addition of a niche player changes the behavior of the major brands; and 〈 An increase from 4 to 8 actions per agent results in more capable agents. In addition, we report on two surprising effects. The Holyfield-Tyson effect: whereby sophisticated agents do not perform that well against primitive agents; and the Frankenstein effect: whereby agents developed in competition with other agents exhibit different behaviors when competing with the historical actions of managers. Overall, we believe the strength of our approach results from the use of an empirically grounded fitness function with which to test our assumptions and approaches.