On the Use of Intelligent Agents as Partners in Training Systems for Complex Tasks1
Thomas R. Ioerger, Joseph Michael Sims, Richard A. Volz, Judson W. Workman, Wayne L. Shebilske · eScholarship (California Digital Library) · 2003
Training protocols that involve working with a human partner have been shown to be beneficial for learning complex tasks.In this paper, we explore emulating the function of the partner with an intelligent agent.Given a cognitive task analysis, the task can be decomposed into cognitive components, and these behaviors can be independently automated using agent-programming techniques.Then a trainee and the agent can work together to solve practice problems, each taking responsibility for a different function.We argue that it is desirable not only for the agent to produce correct and consistent behavior (e.g.demonstrating the optimal strategy), but also to appear realistic (human-like, including errors), and we show how this can be achieved by introducing randomness in an agent's decisions.We implemented a Partner Agent for Space Fortress, a laboratory task designed to be representative of complex tasks, and found that trainees who swapped roles with this agent during training achieved significantly higher performance scores asymptotically than those who trained using a standard (whole-task) training protocol.We also simulated 3 different levels of expertise and found that trainees who worked with an "expert-level" agent received the most benefit.