Instructional Design

The creation of pedagogical agents is a fairly new enterprise that has emerged from previous work done in autonomous agents, intelligent tutoring systems, and educational theory. Pedagogical agents are autonomous agents that occupy computer learning environments and facilitate learning by interacting with students or other agents. Although intelligent tutoring systems have been around since the 1970s, pedagogical agents did not appear until the late 1980s. Pedagogical agents have been designed to produce a range of behaviors that include the ability to reason about multiple agents in simulated environments; act as a peer, colearner, or competitor; generate multiple, pedagogically appropriate strategies; and assist instructors and students in virtual worlds.

Animated Pedagogical Agents

A new breed of pedagogical agents has begun to appear in learning environments and on websites: animated pedagogical agents. The advent of animated pedagogical agents is the result of recent advancements in multimedia interfaces, text-to-speech software, and agent-generation technologies. Some of the more high-profile systems are described below.

• ALI is an automated laboratory instructor that monitors and guides undergraduates as they solve problems while interacting with chemistry simulations.
• ADELE (Agent for Distance Learning—Light Edition) helps students work through problem-solving exercises for courses that are delivered over the Internet. ADELE-based courses have been developed for continuing medical education and geriatric dentistry.
• Auto Tutor simulates the dialogue moves of human tutors while participating in conversations with students. Auto Tutor is currently designed to help college students learn about topics in computer literacy and conceptual physics.
• Cosmo exploits deictic behaviors to offer problem-solving advice to students learning about network routing mechanisms in the Internet Advisor learning environment.
• Herman the Bug inhabits the Design-A-Plant learning environment and helps children learn about botanical anatomy and physiology.
• PPP Persona provides online help instructions while helping users navigate through web-based materials.
• STEVE (Soar Training Expert for Virtual Environments) interacts with learners in an immersive virtual environment and has been used in naval training tasks such as operating engines on U.S. Navy surface ships.
• Vincent helps workers in shoemaking factories learn about production-line control time.

These agents exhibit lifelike behaviors and have the potential to bolster student-learning outcomes by exploiting both the auditory and visual channels of the learner. In general, animated pedagogical agents are lifelike personas, which execute behaviors that involve emotive responses, interactive communication, and effective pedagogy.

Emotive responses. Clark Elliott, Jeff Rickel, and James Lester argue in their 1999 article that animated agents displaying appropriate emotions provide a number of educational benefits to learners. First, agents that appear to care about students' progress may convince students to care about their own progress. Second, agents that are sensitive to learners' emotions (e.g., boredom or frustration) can provide feedback that prevents students from losing interest. Third, agents that convey enthusiasm for the subject matter are more likely to evoke the same enthusiasm in learners. Finally, agents that have rich and interesting personalities make learning more enjoyable for the learner.

Agents can display appropriate emotions through facial expressions, gestures, locomotion, and intonation variations. For example, Cosmo uses a recorded human voice and full-body emotive behaviors to express a wide range of pedagogically appropriate emotions. When a student experiences success in the Internet Advisor learning environment, Cosmo may applaud, point to relevant information on the screen, and provide positive feedback (e.g., "You chose the fastest subnet. Also, it has low traffic. Fabulous!"). Another system, Auto Tutor, synchronizes facial expressions and intonation variations to provide feedback that reflects the quality of students' natural language contributions. If a student provides a good answer to a question, Auto Tutor may respond simultaneously with an enthusiastic "Okay!" a fast head nod, and a smile. However, if the student's answer is only partially correct, Auto Tutor may respond with a less enthusiastic "Okay," a slower head nod, and no smile.

Interactive communication. Most educational websites and software packages are designed to be mere information delivery devices that occasionally employ unsophisticated reward systems as metrics of student understanding. Pedagogical agents, however, facilitate interaction in learning environments and force students to be active participants in the learning process. Agents and learners can collaboratively perform tasks, solve problems, and construct explanations. STEVE, the agent that teaches procedural knowledge involved in operating engines on navy ships, demonstrates for learners how to perform tasks and solve problems. A learner may choose to intervene and finish the demonstration. When this happens, STEVE monitors the learner's actions and a mixed-initiative demonstration occurs. Specifically, learners can take the initiative by asking questions or performing actions, or STEVE can mediate the interaction by providing hints, asking questions, giving feedback, or demonstrating a task. Learning sessions with ADELE are interactive in that ADELE interrupts students when "she" detects student errors and suggests alternative actions to be performed instead (e.g., "Before ordering a chest X ray, it would be helpful to listen to the condition of the lungs."). Students who reach impasses during problem solving may receive hints and ask "why" questions while interacting with ADELE and STEVE. In other systems, such as Auto Tutor, the agent and student have a conversation with each other. Throughout the conversation, Auto Tutor simulates human-tutor-dialogue moves (e.g., hints, prompts, assertions, and corrections), which allow the agent and student to jointly construct answers and explanations to deep-reasoning problems.

Effective pedagogy. In order to be considered value-added entities of learning environments, pedagogical agents must be effective teachers and, therefore, adaptive and dynamic in their teaching strategies. They must be able to adjust their teaching to fit a particular problem state or learning scenario, and they must be capable of adjusting their pedagogy to accommodate students' knowledge and ability levels. Pedagogical agents should be able to ask and answer questions, provide hints and explanations, monitor students' understanding, provide appropriate feedback, and keep track of what has been covered in the learning session. All of the pedagogical agents mentioned above are, to some extent, capable of each of these functions. Of course the litmus test for any pedagogical agent is whether it produces positive student-learning outcomes.

Learning Outcomes

It has been well documented that users prefer learning environments with animated agents over those that do not have agents. Specifically, participants assigned to learning conditions with animated agents (even ones that are not particularly expressive) perceive their learning experiences to be considerably more positive than participants assigned to learning conditions that do not include animated agents. This recurring finding is known as the persona effect. The persona effect is somewhat enigmatic in that it generally is not related to student outcome or performance measures. That is, most researchers who report evidence of the persona effect also report no differences between agent and no-agent conditions for retention and learning measures.

Several recent empirical studies, however, indicate that pedagogical agents do promote learning on both retention and transfer tasks. Robert Atkinson reported that students who received explanations from an animated agent about how to solve proportion word problems outperformed other learning conditions on both near and far transfer problems. In a study conducted by Roxana Moreno et al., college students and seventh graders attempted to learn about how to design plants that could survive in a number of different environments. One group of students interacted with a pedagogical agent, Herman the Bug, while another group of students received identical graphics and textual explanations but no pedagogical agent. The results indicated that students in the pedagogical agent condition outperformed students in the no-agent condition on transfer tests but not on retention tests. In another study, Natalie Person et al. (2001) reported that the effect size for Auto Tutor was .6 compared to the other learning conditions; human tutoring studies typically report effect sizes around .5 compared to other learning controls. Given the results of these learning-outcome studies and the fact that learners perceive their interactions with agents quite favorably, the future for pedagogical agents looks quite promising.