CISA Research Highlights


Integrated Optimization and Prediction based on Adaptive Dynamic Programming (ADP) for Machine Intelligence: This research aims to advance the understanding, principles, and applications of a new learning framework that integrates optimization and prediction capabilities based on adaptive dynamic programming (ADP). Understanding brain intelligence is still one of the greatest unsolved scientific challenges. Among the many efforts toward reaching this goal, optimization and prediction are the two essential objectives. The key question is how to develop general-purpose models and algorithms that can adaptively learn and accumulate knowledge, make predictions in an uncertain and unstructured environment, and adjust actions to maximize some kind of utility function over time to achieve goals (goal-oriented behaviors). This research will advance the fundamental scientific foundations of brain-like, general-purpose intelligence and develop innovative ADP-based techniques to bring such a level of intelligence closer to reality across a wide range of critical applications, such as smart grid control, cognitive radio networks, robotics, and sensor networks. This research is funded by National Science Foundation (NSF).


Intelligent Power Grid Control, Optimization, and Learning: With the continuous significant increase of energy demand and environment issues, the development of a smart electric power grid has become a critical research topic worldwide. Among many efforts toward this objective, computational intelligence (CI) research could provide key technical innovations to help the society to accomplish the essential energy objective. We have been focused on a few aspects of the intelligent energy and power systems, including: (1) Development of adaptive and intelligent control and optimization methods for power grid based on adaptive dynamic programming (ADP), particle swarm optimization (PSO), and other CI techniques; (2) Grid-connected renewable energy systems with a focus on wind energy systems, such as DFIG wind farm control, HVDC transmission, and FACTS control; (3) Agent-based learning, optimization, and prediction for power grid, and (4) power grid data analysis and signal processing, such as power quality, smart grid communications, load prediction, and wind profile analysis.


Smart Grid Security: The security and reliability of electric power grid have drawn significantly increasing attentions from academy, industry, and government after several large-scale blackouts (e.g., North American blackout in 2003, South American blackout in 2009, and the India blackout in 2012) and cyber penetrations of the U.S. grid. The objective of this research is to advance methods of vulnerability analysis and to develop innovative responses to maintain the integrity of power grids under complex attacks (both cyber attacks and physical failures). Unlike many of the existing efforts that focuses on abstract topological structure or load-based analysis, we consider both network topology and intrinsic power flow characteristics to understand system behavior in complex power grid attacks. As power and energy systems have become one of the key technology and economic development focuses across the world, this research will have far-reaching impacts at different levels: enhancing national power system security, providing technical support to government agencies and policy-makers for the U.S. energy sustainability objective, and student education and workforce preparation in this field. This research is funded by National Science Foundation (NSF).


Intelligent Data and Information Processing:  In many of the Department of Defense (DoD) applications, information processing is failing to keep pace with the explosive increase of the collected data from various types of sensor systems to support the decision-making process. Although existing machine learning and data mining techniques have shown great success in many real-world applications, there is a critical need for unified mathematical principles to analyze various types of data collected by the DOD sensor systems. The goal of this research is under the DARPA MSEE (Mathematics of Sensing, Exploitation and Execution) umbrella to develop and advance a unifying mathematical formalism that incorporates stochasticity fundamentally for large-scale data processing. Specifically, we aim to develop the statistical learning and signal processing methods for effective data representation, information extraction, and decision making. This research is funded by Defense Advanced Research Projects Agency (DARPA).


Adaptive Learning and Fusion:  The objective of this research is to develop adaptive learning and fusion methods for stream sensor data analysis in complex and networked systems. With the over abundance of multi-sensor and multi-dimensional raw data in such data-intensive applications, it is critical to develop adaptive learning and fusion approaches to transform such vast amounts of raw data into representative and relevant knowledge. This project will develop different learning and fusion techniques with stream data, including incremental learning, sensor fusion and integration, and various applications. Emerging learning problems such as imbalanced learning, semi-supervised learning, spatial-temporal learning, etc., will also be developed in this research. This research is funded by Army Research Office (ARO).


Cognitive Radio Networks:  Cognitive radio today is based on adaptive algorithms that have been pre-programmed through a tedious process that begins with a sketch on the whiteboard of a radio research laboratory and ends in shipped product 6-18 months later and at considerable expense. The situation is worse for first responders in regional crises like a Katrina Hurricane, or another Indonesian tsunami. To bring the real cognitive level intelligence into the cognitive radio networks, a team of faculty with expertise of communications, networks, and machine learning aim to investigate an adaptive, integrative, and hybrid learning framework for cognitive radio systems. Specifically, we are interested in various key aspects including radio behavior learning, software defined radio platforms, reasoning and knowledge engine for radio, and prototype development and testing.



Hardware Design for Machine Intelligence:  While many of the existing machine intelligence research in the community focus on software and algorithms development, recent developments in deep-submicron electronics (e.g., nanoelectronics, memristor, etc.) provide the technology platform to design complex and integrated intelligent systems in massive, parallel, and scalable hardware platforms. In general, intelligent system models can be simulated in software environment or built in hardware platforms such as very large scale integration (VLSI) systems and field programmable gate array (FPGA) technologies. Software implementation may be easier compared to hardware development, however, it has its own inherent limitations. Therefore, although software-based systems can be used to test some machine learning ideas, they are not sufficient to build the highly integrated and complex intelligent systems in order to mimic certain levels of brain-alike intelligence in the long-term. From this aspect, we have been investigated and developed several approaches for hardware design for machine intelligence, including hardware-oriented machine intelligent architecture, low-power design, and FPGA and VLSI design of learning modules.


 


We gratefully acknowledge the support from: