The Context Machine. The Neal Stephenson book entitled The Diamond Age – or, A Young Lady’s Illustrated Primer (1995) contains a very sophisticated piece of technology, the Primer. In the story, the character Nell reads the primer, the purpose of which is to make a proper lady out of the reader. It creates games related to the events in her life by sensing her “context” and providing her with an analogous story, one that educates her in the ways of the world. If the primer can sense her “context” then it could affect her more directly by providing warnings, guidance and help at the appropriate time and in the appropriate form. The creation of a device similar to the primer is the goal of the Context Machine project, sponsored by DARPA’s Augmented Cognition Program.

The MOVES Institute at the Naval Postgraduate School is participating in the Augmented Cognition Program by creating the Context Machine to explore the notion of “context” in a general way and to study how such a device might improve future warfighting capabilities. The user's current situation, such as their location, their objectives, and the presence of other people and objects, are inputs to the Context Machine. The machine uses this information to determine context, e.g. “Where am I?, What am I doing?”, What do I need to know to accomplish my task?”. Based upon this context, it determines the best course of action to achieve the user’s goals, which is then conveyed to the user. It is imperative that the assistance supplied by the Context Machine be appropriate to the situation, useful, and wanted. We have four goals in building the Context Machine:

1. Develop a robust, dynamic representation of knowledge which is sufficient to denote context.

2. Create a method to accurately transform sensory data (in symbolic form) from the environment into this context data structure.

3. From the current context, determine a course of action that best meets the user’s goals.

4. Accommodate shifting and divergent goals.

Technical Approach. The Context Machine’s most important property is the ability to learn from its user and its environment as it is used. This is crucial to allow such a machine to adapt to a specific user and to continue to adjust as that user’s circumstances and goals change. In addition, systems that learn are not constrained by the limits of their developer’s knowledge. This means they can produce novel results or solutions to situations that the creators did not foresee as they designed the system. Therefore, we intend to begin our investigation into this problem using an agent-based approach that employs machine learning algorithms to give the Context Machine the ability to learn as it is used.

In our agent-based approach, we anticipate that the Context Machine will be made up of two types of interdependent agents, each with a different task:

· Symbolic constructor agents (SCA) – Create an inner topography based upon input from the outer world. Doing so creates arranged combinations of gradients.

· Reactive agents (RA) – Traverse the inner world, following gradients created by SCA’s. These agents make judgments, inferences and recommendations.

Milestones. We have broken this project down into the following milestones:

· Determine the architecture to use as a base for the Context Machine. (CY-01)

· Investigate what information is necessary and sufficient to denote context for a given situation. (CY-01)

· Describe a situation’s context, given inputs from the symbolic sensor stream. This will likely be task dependent. (CY-01)

· Dynamically gather information from the symbolic sensor stream and insert into the data structure to be factored into determining context. (Future work)

· Infer the user’s implicit goals and mesh these with stated, or explicit, goals. (Future work)

· Determine appropriate actions or responses based upon the current context and the user’s goals. (Future work)