Use of partner agents in training systems for complex tasks

Abstract

Complex tasks are an important part of the world today, but training people to perform complex tasks is difficult. These kinds of tasks usually have variable cognitive loads on the trainee, and, once trained, different people have different capacities for performing a complex task. Research has shown that performance of a complex task asymptotically approaches some level after completion of training, but also that the asymptote and the speed at which it is reached can be altered by changing the training protocol. On the other hand, certain training protocols such as AIM Dyad, which pair two students together, can offer trainees half of the hands-on experience compared to other protocols and still offer the same level of training. This is thought to be due to one trainee "modeling" the other part of the task. However, it has also been shown that social interaction is not required, which suggests that equal benefits could be derived from using intelligent agents as automated partners. In this thesis, I will describe a new training protocol where the trainee and an intelligent agent team with each other to perform the task. The agent performs half of the task, using a target strategy in a believable way. Through a series of experiments, while the trainees have only half of the hands-on practice of some other protocols, the "partner agent" protocol improves the effectiveness of their training. I also explore the impact of simulating different levels of expertise in the partner agent.