Projects

Automated Image Understanding for Maritime Threat Analysis

The Maritime Activity Analysis Workbench (MAAW) is a ground-breaking intelligent system that automatically learns to detect suspicious maritime activity near ports and harbors by analyzing surveillance video, improving the US Navy's capability to secure ports and harbors.

We are collaborating with the Navy Center for Applied Research in Artificial Intelligence (NCARAI) at the Naval Research Laboratory (NRL) to develop MAAW and new machine learning approaches for maritime image processing and situation understanding.

Advanced Question-Answering for Intelligence (AQUAINT)

Researchers the world over are exploring next-generation search techniques called question answering (Q-A) to enable users to ask questions in natural language and receive precise answers. In 2003, the Intelligence Advanced Research Projects Activity (IARPA) launched the AQUAINT Program to develop these Q-A technologies.

In support of the AQUAINT Program, Knexus Research performed two broad tasks in collaboration with the Interactive Systems Section at NRL. First, we performed an extensive human-subject study of advantages and disadvantages of Q-A systems compared to traditional keyword search systems. Second, we developed a secure and configurable Service-Oriented Architecture (SOA) called Intelligent Question-Answering (INQA) which integrates three state-of-the-art Q-A systems. INQA is a Q-A service broker and an application server that provides a rich, intuitive web-based interface for end users to perform intelligence analysis. It also supports a wide variety of performance evaluation metrics used in our human-subject study.

Transfer Learning

Can an intelligent agent be developed which learns to play chess at a higher level or more quickly because it had previously learned to play checkers? Human beings leverage other skills all the time, and this Defense Advanced Research Projects Agency (DARPA) program worked on intelligent agents capable of similar feats. Researchers developed methods, implemented as software agents, that could transfer skills learned from a "source" task to a range of "target" tasks. Agents developed in this project may someday function as adaptive opponents in military simulations.

For this project, we collaborated with the Adaptive Systems Section at NRL to investigate new techniques using case-based reinforcement learning to achieve learned performance in simulations with large decision spaces. we also built, maintained, and deployed two frameworks for evaluating intelligent agents transfer learning studies: TIELT and LIET. We extended an open-source football simulator, Rush 2008 for use in Transfer Learning research. Finally, we performed evaluations for the Transfer Learning programs using our developed tools.

Machine Translation Evaluation

Automatic translation systems are needed for a variety of tasks from intelligence analysis to mission execution. What makes a translation good or even acceptable? How do we measure translation quality? These are some of the questions we are trying answer in this project.

Although automatic measures of translation quality exist, they are not good predictors of how a human will rate a translation. We are conducting a deep linguistic analysis to examine this issue.

Coping with Information Overload

The objective of this project was to develop methods for reducing the effort and improving the decision effectiveness of radio operators, who must simultaneously monitor multiple audio channels.

We worked with the Navy Center for Applied Research in Artificial Intelligence (NCARAI) at NRL to develop techniques that learn to automatically prioritize, intelligibly speed-up, and optimally schedule radio communications. Additionally, we developed the Concurrent Audio Serialization Workbench (CASW) application to demonstrate our methods and integrate it with another DoD radio communications project.

Administrative Process Re-engineering

Administrative processes in an organization can span long periods of time and multiple company divisions, leading to poor service, errors, and unacceptably long delays. We are re-engineering and overhauling our client's administrative processes to dramatically improve service quality and process transparency.

We are supporting our client with management and execution of all aspects of this program, including vision, analysis, tool selection, implementation, and operations.

Intelligent Reference Data Mapping

Our objective was to reduce the effort and errors in upgrading massive multi-agency legacy applications. We developed techniques that learn to assist database analysts performing reference data mapping, an arduous task of identifying and maintaining links from legacy database schema elements to constantly-changing reference data schema elements.

We developed this capability for our client, CDM Technologies Inc. to be embedded in their Intelligent Mapping Toolkit (IMT) for the United States Transportation Command (USTRANSCOM). Our intelligent mapping technology is capable of suggesting correct mappings within the top five suggestions over 50% of the time. This level of mapping support is unprecedented for schema pairs that have over 10 million possible mappings. IMT continuously learns and improves its suggestions based on the mapping selections made by the datamappers. Its ability to explain its recommendations is an important factor towards IMT's acceptance by datamappers.

We have shown that the tool can save USTRANSCOM datamappers an hour per mapping; a saving of one full-time data mapper.

Advanced Techniques for Netcentric Warfare

Netcentric warfare requires rapid and accurate information access for superior situational awareness and decision making. Web services are quickly becoming the Department of Defense's information access method of choice. However, exploiting available web services that have heterogeneous semantics (WSDLs) is problematic.

In collaboration with our government partners, the Intelligent Decision Aids (IDA) Group of NRL and the NRL Stennis Space Center, we have developed new methods of request and response brokering that enable a web client to dynamically locate and access web services of a desired domain (e.g., weather information).

We showed that, using a set of approximately 60 known webservices, our "Intelligent" Broker learned to identify the relevant webservices with an accuracy of over 97% and found the relevant methods for brokering requests and responses over 97% of the time.

Our accomplishments in this project include an embedded Java library for intelligent brokering, five peer-reviewed publications and a patent application.