Period:
Effective tactical decision-making requires a quick and accurate appreciation of the current and near-future situation. In today’s highly interconnected world, data relevant to situational awareness is plentiful. A decision-maker has access to many rich sources. Some represent sensory or processed data that describe the current world state and effect future world states. Many others are archival in nature. These may include large online databases of past episodes. Others are repositories of relevant general, abstract, or qualitative knowledge and procedures. But many will be ad hoc resources containing specialized data not specifically intended for use by the decision-maker. Nonetheless, if the decision-maker can find the relevance of such sources, decision-making can be greatly improved.
While the abundance of data sources affords unparalleled opportunities for improved decision making, it also poses significant risks and challenges. Hand-in-hand with opportunities for improvement go opportunities to degrade decision-making. Data can be noisy, incorrect, or misleading. A lot of the data is unstructured, mostly text, which is qualitative in nature and difficult to interpret and utilize in decision-making situations. In a large, diverse, and interconnected system it is impossible to assure accuracy or even coherence among the data sources. Indeed there are likely to be inconsistencies within many of the individual data resources themselves.
The objective of this project is to study a unified inference framework that can take as input models that correspond to disparate sources and modalities. The intention is to study both how to learn good models from different sources with different kinds of associated uncertainty, and how to combine these into a coherent decision, taking into account characteristics of the data as well as of its source. Of specific importance to us is to incorporate—along with statistical models of sensory data—tabulated historical data, declarative information sources in the form of expert advice, qualitative textual information, and/or domain and task-specific constraints expressed in high-level language.
The information fusion process is complex and in our view requires machine learning methods and inference processes that themselves must be adaptive. This requires methods that exceed current technologies. Our research will advance the state-of-the-art research in Machine Learning, Natural Language Processing, and Artificial Intelligence Reasoning in the course of studying and developing technologies directed toward robust Information Fusion.