Brian M. Slator

Computer Science Department

North Dakota State University

Fargo, ND 58105

email: slator@badlands.nodak.edu

Several different approaches to intelligent tutoring
agents are described in the context of implemented
virtual worlds for education.

Key Words: 1) Autonomous Agents, 2) Multimedia and Human Computer Interaction

Research in active learning environments includes implementing "live" simulations for exploration and discovery that engage learners while treating them to a plausible synthetic experience. The NDSU World Wide Web Instructional Committee (WWWIC) is currently engaged in several virtual/visual development projects: three NSF-supported, the Geology Explorer (Saini-Eidukat, Schwert, and Slator, 1998), the Virtual Cell (White, McClean and Slator, 1999) the Visual Computer Program (Juell 1999), and the ProgrammingLand MOO (Hill and Slator, 1998), the Blackwood Project, Virtual Polynesia, as well as others. These have shared and individual goals. Shared goals include the mission to teach Science structure and process: the Scientific Method, scientific problem solving, deduction, hypothesis formation and testing, and experimental design. The individual goals are to teach the content of individual scientific disciplines: Geoscience, Cell Biology, Computer Science.

The WWWIC projects are designed to capitalize on the affordances provided by virtual environments. For example, to

Specifically, the WWWIC projects each design with the following over-arching principles.

Role-based: Simulated environments enable learners to assume roles in particular contexts and have meaningful, authentic experiences. Rather than simply teaching goal-based behavior and tactical task-oriented skills and methods, the role-based approach communicates a general, strategic, manner of practice (McLuhan, 1964).

Goal-oriented: Goals are important, but within the context of roles. It is through goals that obstacles leading to problem solving are encountered.

Learn by Doing: When these experiences are structured and arranged such that playing a role in the environment illustrates the important concepts and procedures of the simulated domain, students are able to "learn by doing" (Dewey, 1900).

Spatially oriented: WWWIC simulations are spatially-oriented to leverage off the natural human propensity to towards physically plausible context.

Immersive: The combination of role-based scenarios and spatially oriented simulations create an immersive atmosphere where it is strategic thinking that the apprentice eventually learns to model

Exploratory: Exploratory simulation means enabling students to pursue their own interests and where learners are self-directed and given the freedom to structure, construct, and internalize their own experience (Duffy and Jonassen, 1992).

Game-like: The value of play in learning can hardly be over- stressed. Insofar as possible, educational software should be engaging, entertaining, attractive, interactive, and flexible: in short, game-like (Slator and Chaput, 1986).

Highly Interactive: One major challenge for science educators is to develop educational tools and methods that deliver the principles but also teach important content material in a meaningful way. At the same time, the need for computer- based education and distance learning systems has become increasingly obvious, while the value of "active" versus "passive" learning has become increasingly clear (Reid, 1994).

Multi-user/player: One challenge is to craft role-based, goal- oriented environments that promote collaboration as well as the more easily conceived competition.

The overall goal is focused on developing and employing intelligent agents within multi-user distributed simulations to help provide effective learning experiences. From the perspective of intelligent tutoring systems, the agents of interest must fundamentally support models of the knowledge of a domain expert and an instructor. However, it is desirable that the agents have a number of additional capabilities as well, including awareness and understanding of other agents in the simulation. Some of the desirable intelligent agent capabilities are as follows:

All systems of multiple software agents, including those created for pedagogical purposes, must be provided with the ability to communicate with their peers through the exchanging of messages, usually expressed in an Agent Communication Language (ACL). Typically an ACL will provide for the communication of such things as constraints, negations, rules, and quantified expressions. There are a variety of approaches to providing an ACL, and there are also dependencies on communication and interoperation standards such as CORBA and OLE (Mayfield, et al., 1995). There are key issues involving the semantics of specifying such things as preconditions, postconditions, and satisfiabilty; network transport mechanisms, security and authentication (Genesereth and Ketchpel, 1996).

There are several alternative means of designing and developing an agent architecture, and they differ in their appropriateness for pedagogical applications. One type of approach employs direct interagent communication mechanisms, and all agents handle their own coordination activities. For example, in the contract-net approach (Davis and Smith, 1983), agents distribute requests for proposals to other agents, who respond with bids to the originators, who may award contracts for services. Specification-sharing approaches involve agents advertising their capabilities and needs, which are then employed by other agents. A competing approach organizes the agents into federated systems, generalizing the concept of a mediator (Wiederhold, 1989). A federated system uses facilitators to perform intermediate brokering functions and transfer of messages, eliminating direct agent-to-agent communication.

We implement three different approaches to intelligent tutoring, based on the knowledge available to the tutoring agent.

Deductive Tutors provide assistance to players in the course of their deductive reasoning within the scientific problem solving required to accomplish their goals.

Example: intelligent tutoring agents in the NDSU Geology Explorer, an educational game for teaching the Geosciences. Planet Oit is a synthetic (or virtual) environment where students explore a planet as a Geologist would. While on the planet, students are assigned goals related to rock and mineral identification.

The tutors work from knowledge of the rocks and minerals, knowledge of the "experiments" needed to confirm or deny the identity of a rock or mineral, and the student's history. Three opportunities for deductive tutoring are:

The equipment tutor is called by the purchase verb (described in http://oit.cs.ndsu.nodak.edu /oit /usercard.html). The tutor checks whether the instrument purchased can be used to satisfy any of the player's goals. If not, the tutor may decide to remediate on that topic (i.e. buying instruments that serve no obvious purpose). The equipment tutor is also called by the exit(s) from the Equipment Locker. The tutor checks whether the student has ALL the instruments needed to satisfy their goals; i.e. if their goal is to locate Calcite, which requires an acid reactivity experiment, and they have not acquired an Acid Bottle. If not, the tutor may decide to remediate on that topic (i.e. the need to buy instruments that serve to satisfy goals)

The exploration tutor is called by the exit(s) from each of the locations (rooms) on the Planet. The tutor checks whether the student is leaving a room that might satisfy a goal; i.e. if their goal is to locate Kimberlite, and there is Kimberlite in the room they are leaving, the tutor may decide to remediate on that topic. This remediation could be done on a room-by-room basis (easiest), or it could be done on a region-by-region basis (harder); or on a hot-cold basis (i.e. if the player is moving farther away from some distant goal). The current implementation is the simplest one, on a room-by-room basis.

The science tutor is called by the report verb (described in http://oit.cs.ndsu.nodak.edu/oit/ usercard.html). For example, suppose the student is given the goal of locating and identifying graphite used in the production of steel and other materials. To confirm that a mineral deposit is indeed graphite the student must test the deposit with the "streak plate" and observe a black streak, and scratch the deposit with the "glass plate" to determine its hardness is less than 2.0 on the standard (Mohs) scale, and report that the sample as graphite

The tutor checks the player's .geology_history property and determines which of the following cases pertain:

The system encodes the necessary and sufficient experiments for each rock and mineral, as well as their expected results. The tutors check these facts against the student's .geology_history property whenever the student "guesses" a deposit's identity. Tutors remediate, as appropriate, according to the four cases listed above.

Case-based Tutors provide assistance to players by presenting them with examples of relevant experience. This is accomplished by

Example: intelligent tutoring in ORCA, the ORganizational Change Advisor (Slator and Bareiss, 1992; Bareiss and Slator, 1993; Hinrichs, Bareiss and Slator, 1993), which is designed to teach consultants how to identify problems that may face a business, and to expose them to potential solutions to those problems. ORCA presents business war stories about organizational change in response to economic and technological pressures. As the user works with a client, the client becomes a new story in the system, thus extending ORCA to serve as a corporate memory.

These agents act as sub-topic experts who have access to problem solving experiences, context sensitive help and advice, conceptual and procedural tutorials, and stories of success and failure within their particular sub-topic. The agents monitor player's progress and "visit" a player when they need their particular help. The agents will coach the players by sharing their expertise in the form of prototypical case studies, problem-solving dialogs, and pre-packaged tutorials.

Tutoring agent behavior is adaptive in two senses. Each agent is the "owner" of a small set of sub-topics and related cases. As a learner operates in the synthetic environment they are building their own case, and the relevant agent is alerted for remediation whenever a learner case becomes similar and relevant to a case under the agent's control. As learner behavior changes, so does their profile and the nature of the cases they match against. In this way, agents will "gain and lose interest" in a player according to the changes in the learner's profile.

Learner state will be preserved throughout the course of their involvement of the synthetic environment. As learners leave the game, either as successful or unsuccessful players, their state and experience is saved as a new case. These saved cases, according to their profile, will become part of the inventory assigned to one or more of the tutorial agents. As later players enter the synthetic environment, the tutorial agents will have these additional cases of success and failure to present as part of their remediation package.

In other words, case-based tutoring agents begin the game armed with prototypical case studies, but they will accumulate additional student case studies as players enter and leave the game over time.

Rule Based Tutors provide assistance by

Example: the NDSU Dollar Bay Retailing Game, simulating of a micro-economic environment. For example, a player may decide to try and maximize profits by pricing their products at ten times the wholesale price. This is a naive strategy that says, "I might not sell very many, but each sale will be very profitable".

This breaks the simple rule: don't set prices unreasonably high. The intelligent tutoring agent recognizes this as a losing strategy and knows the player is unlikely to sell anything at all. When the agent detects a strategic mistake it sends a message to the player saying, "You may be setting your prices too high". Then, the player can then decide to ignore the message or pursue it.

This unintrusive method of tutoring is implemented to be consistent with the educational game principles of leaving the player in control and letting them make their own mistakes. If the player chooses to pursue the warning, the agent who presents cases and explains the ideas of profit margin and Manufacturers Suggested Retail Price (MSRP) engages them. The special tutorial lessons include further case retrievals, conversational browsing with the agent, and possibly a canned tutorial on price setting.

The drawback is that rules of this sort are relatively rare (and relatively obvious) in a complex domain. Therefore the rule-based method does not afford as many opportunities as other approaches. Finally, breaking the rules is not always the wrong thing to do (as many experts will tell you), which can create problems in all but the most elementary teaching systems.

A common problem with simulations is that, like the real world, players can foul things up and not know why. Unlike the real world, though, all the information for the simulation is readily available, and can be used to generate explanations or warnings. Tutoring agents are based on the design and information in the model, and are triggered by user actions. When an agent is activated, the player sees a warning; they can ask for more information (possibly bringing them a "visit" from an intelligent tutoring agent), or they can ignore it and carry on at their own risk.

The idea is that intelligent tutoring agents are looking over your shoulder as you play. They should be there when you need them, but when you know what you're doing (or when you think you know), you can ignore the agent. Again, there should be no penalty for ignoring the agent's warnings, other than the inevitable failure to succeed, a penalty imposed by the simulation as a consequence of the player's failure to learn their role in the environment. In all cases it is up to the player to decide how the warnings and advice apply to them.

Our approach takes the form of tutoring scripts that shape the interaction in general terms. We are working on tutoring scripts that array different combinations of questions, examples, cases, remedial exercises, and canned presentations to engage different students at different times, depending on student behavior.

WWWIC is supported by NSF-DUE-9752548 and EAR-9809761