Eric You XU's homepage
You can simply call me Eric
Research
My favorites:
Books:
Writers:
Quotes:
Food:
Travel Desinations:
Misc:
- Research with my advisor (Nonlinear Programming)
Nonlinear programming is the
essential part of many AI problem. Our group will propose a
partitioning and resolving approach.
a). A Generalize Nonlinear
Programming Problem Instance analyzer, or a "meta solver" (It will
later go open source)
Nowadays, the widely used problem definition language used in modelling the NLP program is called AMPL. Unfortunately, modelling is only the first step on the long march. In fact as NLP is so hard to solve, usually many existing solvers aims to solve certain classes of programs. On our research, we firstly developed a parser which can parse the original problem and probe the structure of the original problem. We've noticed that there are some similar projects like Dr. AMPL. Since we need to do convex analysis, automated initial point calculation and automated problem decision, now we are trying to develop a generalize analyzer for nonlinear programming problem.
b). Adaptive Problem Partitioning Module. (Cooperate with ANL, sponsored by DoE)
As we can problem the problem structure, the goal of this research is to develop a adaptive partitioning module which can partition the whole problem into sub problems and the by using parallel processing or just sub problem decomposing, hopefully we will accelerate the total speed of the solve and thus make the intractable large scale NLP problem solvable. However, there are two key issues:
1. How to partitioning the problem so that we can handle the global constraints easily.
2. How to make the suitable partition size so that our algorithm will get the result in the fastest manner. The goal of this research is to solve this two issues and find some theoretical or empirical result in designing partitioning strategy.
c). Basic Solver Interface (Hook other solvers like SNOPT/IPOPT, etc.)
d). Penalty Control Module in General Nonlinear Programming Solver
e). (Later) NEOS NLP solver with automated partition and solve feature.
Nowadays, the widely used problem definition language used in modelling the NLP program is called AMPL. Unfortunately, modelling is only the first step on the long march. In fact as NLP is so hard to solve, usually many existing solvers aims to solve certain classes of programs. On our research, we firstly developed a parser which can parse the original problem and probe the structure of the original problem. We've noticed that there are some similar projects like Dr. AMPL. Since we need to do convex analysis, automated initial point calculation and automated problem decision, now we are trying to develop a generalize analyzer for nonlinear programming problem.
b). Adaptive Problem Partitioning Module. (Cooperate with ANL, sponsored by DoE)
As we can problem the problem structure, the goal of this research is to develop a adaptive partitioning module which can partition the whole problem into sub problems and the by using parallel processing or just sub problem decomposing, hopefully we will accelerate the total speed of the solve and thus make the intractable large scale NLP problem solvable. However, there are two key issues:
1. How to partitioning the problem so that we can handle the global constraints easily.
2. How to make the suitable partition size so that our algorithm will get the result in the fastest manner. The goal of this research is to solve this two issues and find some theoretical or empirical result in designing partitioning strategy.
c). Basic Solver Interface (Hook other solvers like SNOPT/IPOPT, etc.)
d). Penalty Control Module in General Nonlinear Programming Solver
e). (Later) NEOS NLP solver with automated partition and solve feature.
- Research of my own (incubator)
a). Generalize Parallel NLP
solver framework based on the Functional Programming model (MapReduce
Model)
b). Data Mining in large scale data and Vertical web search. (I am now trying to build a cluster for Apache lucene, hadoop and nutch. I will use up to 10 notes in building this demonstrating system. (commodity hardware.)
b). Data Mining in large scale data and Vertical web search. (I am now trying to build a cluster for Apache lucene, hadoop and nutch. I will use up to 10 notes in building this demonstrating system. (commodity hardware.)