draw on cloud-based cognitive services and, in turn,
can offer cloud-based learning services. Developers
can create agent-enhanced learning objects and distribute them to other providers, or make them available publicly as open educational resources (OERs).
Hosting digital ecosystems in the cloud provides
access to learner data. As learners access learning
services in the cloud, the services can capture and
analyze the learners’ responses. Machine learning
algorithms can then be applied on the captured data
both to improve agent behavior models and to develop models of common learner errors. We foresee a
transition to a data-driven approach to pedagogical
agent development, where agents act as data collection tools as much as learning tools. This double
focus will further accelerate the development and
adoption of agents and lower production costs.
Outstanding Research Questions
Although there have been significant advances in the
science and technology of pedagogical agents, there
remain important outstanding questions. Such questions are likely to guide future research in the field.
Although there is evidence of effectiveness overall,
and evidence that agents support certain types of
learners, there are also exceptions. More research is
needed relating the characteristics of agents, learners,
and domains with learning outcomes. We now rec-
ognize that there are multiple types of agents and
that these agent types differ in terms of the learners
and domains that they benefit. Classic pedagogical
agents are well suited for younger learners, while
role-playing agents can benefit adult learners as well.
Classic pedagogical agents benefit low-performing
learners, while teachable agents benefit high-per-
forming learners. By analyzing each type separately
and comparing results through meta-analyses, it will
become clearer how to use agents most effectively.
Further research should also consider cost-effec-tiveness, to determine whether the benefits of agents
justify the cost in particular applications (Schroeder
et al. 2011). In practice, there are ways to manage
costs, such as using a combination of animation and
still images, as shown in figure 5 (right). In the past,
some researchers have questioned the value of agents
altogether, arguing that when agents are found to
enhance learning, a less expensive and less distracting alternative has equal or greater benefits (Clark
and Choi 2007). We think that it depends upon the
pedagogical use of the agents. As the above examples
illustrate, pedagogical agents make possible new
types of interactive learning experiences that are difficult to match using static multimedia presentations. They engage learners emotionally and socially,
not just at a cognitive level. Meanwhile, machine
learning and data-driven approaches will continue to
drive down development costs, shifting the return on
investment (ROI) breakeven point in the direction of
more highly capable pedagogical agents.
Concluding Remarks
Back in the 1990s, when we started work on peda-
gogical agents, we saw a bright future for the tech-
nology. Much has been accomplished in the past 20
years; agents have become established as learning
tools and continue to improve. So we continue to be
excited about the new possibilities that the technol-
Articles
Figure 7. A Virtual-Reality Environment for Learning Cross-Cultural Skills.
©Alelo Inc. and Concurrent Technologies Corporation. Reprinted with permission.