Evolutionary Optimization of Neural Networks For DNA Sequencing

Jim Golden, Director of Bioinformatics, GeneSys Technologies

2:30 pm Fri Dec 6 2310 CS & Stats

There are many methods available to the engineer for modeling a system such as an automated DNA sequencing instrument. Many of these use an inductive approach to modeling, including numerical approximations from ordinary-derivative models, numerical solutions from partial-derivative models, and numerical simulations using finite elements. Because these estimation procedures are inductive, they requirethe engineer to write down a mathematical model of how function outputs depend on inputs. The initial guess at the shape of a function based on a desired behavior tightly constrains the categories of functions that can be estimated.

Neural networks approximate functions using raw sample data; estimation is deductive and allows the engineer to explore a wider range of possible functions governing behavior without having to know what the functions are. Neural networks, however, suffer from two shortcomings: it is difficult to specify an effective network architecture given the specifications of a problem, and it is very easy for gradient algorithms to get stuck in local minima when learning network weights. Genetic algorithms are a robust method of near optimal search on functions. For ill-defined problems, such as finding the best neural network for modeling a set of test data, genetic algorithms offer a novel and robust method for searching the space of possible designs.