Using Prior Knowledge and Learning from Experience in Estimation of Distribution Algorithms

Estimation of distribution algorithms (EDAs) are stochastic optimization techniques that explore the space of potential solutions by building and sampling explicit probabilistic models of promising candidate solutions. One of the primary advantages of EDAs over many other stochastic optimization techniques is that after each run they leave behind a sequence of probabilistic models describing useful decompositions of the problem. This sequence of models can be seen as a roadmap of how the EDA solves the problem. While this roadmap holds a great deal of information about the problem, until recently this information has largely been ignored. My thesis is that it is possible to exploit this information to speed up problem solving in EDAs in a principled way. The main contribution of this dissertation will be to show that there are multiple ways to exploit this problem-specific knowledge. Most importantly, it can be done in a principled way such that these methods lead to substantial speedups without requiring parameter tuning or hand-inspection of models