Devika Subramanian's research is aimed at the design and analysis of resource-bounded systems that adapt and learn from experience. With Stuart Russell of UC Berkeley, she wrote the first paper defining the area of bounded optimality, i.e., what it means for an agent to make the "best" use of scarce resources. Her work centers on several applications designed to push the science of adaptive systems. Her current projects are in three main areas:

• Computational biology: With scientists at the M.D. Anderson Cancer Center and Baylor College of Medicine, she is reverse-engineering metabolic networks from gene expression data in cancer cells, reconstructing signal transduction networks in granulocyte differentiation in AML and CML from flow cytometry data, and identifying multi-locus genetic markers from genotype-phenotype association data.
• Conflict forecasting: With support from an NSF ITR, and in collaboration with Dr. Richard Stoll of Rice University, she is developing a system for predicting outbreaks of conflict in the Middle East based on temporal analysis of newswire stories from the region.
• Adaptive compilers: With support from an NSF ITR and in collaboration with Professors Cooper and Torczon of Rice University's Scalar Compiler Group, she is designing learning algorithms that help compilers customize their optimization strategies to specific programs.

She is also very interested in curricular reform in computer science and revamping the way we train computational scientists in other fields. Here are a few recent essays on these topics.

• Talk presented at Microsoft Workshop on Computational Education for Scientists, 2007.
• Position Paper presented at Microsoft Workshop on Computational Education for Scientists, 2007.
• Draft Paper, to appear in SIGCSE 2008.

Her past projects include: designing an adaptive outdoor tour guide for the Rice campus (funded by Rice Engineering), reinforcement learning for non-stationary environments and applications to network routing (funded by Southwestern Bell), designing adaptive control systems for the Mars Bioplex (funded by NASA), designing experimentation strategies for protein crystallography (funded by NIH), adaptive compilers for power-sensitive applications (funded by Darpa and the Texas Advanced Technology Program), automating the conceptual design of opto-mechanical systems from specifications of behavior (funded by NSF), and dynamically learning models of humans acquiring a complex visualmotor task (funded by ONR).

Subramanian's expertise is in the design of statistical machine learning algorithms with probabilistic performance guarantees. Her approach is experimental; she designs new algorithms in the context of large-scale applications in science and engineering. Her work has appeared in premier conferences and journals in artificial intelligence, machine learning, computer systems,compilers, networking, computational biology, protein crystallography, robotics, mechanical engineering design, computational neuroscience, cognitive science, and political science.

Subramanian served as co-Program Chair for AAAI in 1999. She was on the IJCAI Advisory Board in 2001. She has given many invited lectures on her work --- the Lucent/CRAW Lecture at the University of Washington in 2002 on her work in learning models of conflict from political events data, the ONR Invited Lecture in 2001 on her work in tracking human learning, and an invited lecture at IJCAI-93 on her work on opto-mechanical design. She served on the Editorial Board of the Journal of AI Research from 1997 to 2001. She has won several teaching awards at Stanford (George Forsythe teaching prize), at Cornell (two Merrill Presidential Awards), and at Rice (Julia Miles Chance Prize).

Here is a recent research statement.