On Analyzing Interactions in a Software Agent Marketplace

The merger of distributed object technology, agent systems and electronic commerce enables a new generation of useful information services. With the advent of interoperable information services, markets featuring software agents as information service/product providers can be created. Meta search engines that rely on a marketplace of search engines and software component marketplaces (e.g., flashline.com) are examples. The design of these new e-markets requires an understanding—in economic terms—of the interaction between the market mechanism used to coordinate the agents and the decision strategies used by these agents in market sessions. Using a stylized marketplace for machine learning services, IBIZA-ML, we present our simulation-based approach to analyze the interaction between two coordination mechanisms—a centralized scheme and a decentralized scheme—and two agent decision-making strategies—random and information intensive strategies—in markets for custom-built information product, hence though multiple seller-agents may build a product; the buyer accepts only a single information product: the one with the highest quality. Using social welfare as a metric, we find that the information intensive decision strategies dominate random strategies under either coordination scheme. Within information intensive strategies, we find that the centralized mechanism is superior to the decentralized mechanism unless the marginal costs for computing are high. Finally, we find that when the buyer us allowed to specify minimum acceptable quality thresholds, information intensive strategies used with the centralized scheme do better than their counterparts used under the decentralized scheme.