11.10 The AI Apprentice

In the true spirit of AI, the EVS should have the capabilities of incremental learning, incorporating genetic algorithms and the ability to gain insight and knowledge using machine-learning algorithms. This visionary goal requires the development of a learning apprentice system that can use a variety of knowledge and data sources and an interactive interface to acquire human expertise. The EVS should be able to integrate and comprehend direct natural language instruction and advice from criminal and counter-intelligence experts; it should also be able to learn solutions from sample cases.

Such a visionary system would combine an ability to mine large amounts of data, as well as to learn from its interaction with experienced intelligence analysts and law enforcement investigators. For example, an analyst could provide an initial description of a profile of a "smuggler entity" using a language that allows him or her to specify an initial set of defining objects and relations. Using this profile and an existing domain-general knowledge base, the EVS would search through a variety of databases to identify individuals that satisfied this profile and retrieve a set of matching smuggler entities.

The analyst would next examine these entities and label some of them as correct or incorrect, possibly providing some distinguishing properties or relations to help explain the errors. Using these matched and mismatched samples and the analyst's comments, the EVS would revise its knowledge base, which could be a set of hundreds of IF/THEN rules. This interactive process would repeat for several cycles, resulting in a significantly improved knowledge base and reduced error rates and could be used for future searches by the EVS in profiling potential smugglers.

In order to construct such an EVS, techniques are required for organizing knowledge, such as abstraction hierarchies that allow the efficient use of prior knowledge during learning. Already such knowledge bases are being developed under DARPA's High Performance Knowledge Bases Program. Domain-specific knowledge bases of concepts and theories for particular applications, such as narcosmugglers, money laundering, or bioterrorist activities, need to be developed and deployed. They would be a key component of the envisioned EVS.

Interactive theory refinement algorithms need to be developed that handle incremental addition of prior knowledge and exploit directed examples and advice from the human experts. Learning algorithms that exploit forms of prior knowledge other than declarative domain theories need to be developed. This includes methods that automatically extract knowledge using data and text mining agents. The algorithms need to utilize procedural prior knowledge and additional relevant unsupervised data to guide the EVS in learning.