Adaptive Divide-and-Conquer Using Populations
                  and Ensembles 


 
Abstract:

Many real-world problems are too large and complex for a single monolithic
system to solve. The divide-and-conquer strategy has often been used in
practice to break a large problem into tractable smaller sub-problems and
then solve them. However, useful division of a large and complex problem
often requires experienced human experts and rich prior domain knowledge,
which are usually unavailable for real-world problems. This talk describes
some of our research efforts towards an automatic approach to
divide-and-conquer. By evolving and training a team of specialists from
random initial conditions, we were able to "decompose" a large and complex
problem into simpler ones and solve them without human intervention. Two
major approaches will be described. One uses the population structure in
evolutionary algorithms, where individuals in a population are evolved
into species (i.e., specialists for solving sub-problems). The other uses
neural network ensembles in which individual neural networks learn to
differentiate from and cooperate with each other. A constructive algorithm
for designing ensembles as well as individual neural networks will be
introduced.