Human agents and intelligent agents: an experiment on the Internet

Abstract

To understand the strengths and limitations of software agents is very important. Many issues and limitations have been discussed in the literature, for example, software agents have limited learning capability. However, if we could see such limitations in a controlled environment and measure how such limitations affect the agents' performance, such research can be extremely helpful to design better software agents. We developed an artificial market with networked computers to simulate the environment of the Internet. This artificial market allows players, such as, humans or software agents, to compete in a simulated dynamic market. We conducted two sets of experiments to study the limitations of software agents. The first set of five experiments were used to identify the knowledge or strategies that human subjects applied in competition. We classified the knowledge or strategies into four levels: perception, uses of simple model and heuristics, application of long-term strategies and psychological tactics, and prediction or forecasting. The second set of five experiments studied how much humans can transfer their knowledge to the software agents. In all the experiments, subjects received monetay rewards From the statistical analysis of the decisions of humans and software agents, we have identified several limitations to the software agents. For example, some subjects used veiy complicated strategies and psychological tactics in earning high profits. However, they could not transfer such knowledge to the software agents. From interviews and questionnaires we confirmed such findings. Based on these Jindings, we conclude that software agents are suitable for wellstructured and repeating tasks and not suitable for critical and high-risk task, such as, investment.