Global Optimization and Evolutionary Algorithms

par/by Javad Sadri
Center for Pattern Recognition and Machine Intelligence (CENPARMI)
Université Concordia

In the first part of this seminar, a new Genetic Based Particle Swarm Optimization (GBPSO) model is presented. In the original PSO model, there are no evolutionary operations and the size of the swarm is fixed. However, in our GBPSO model, particles (for example, birds) can mate together to produce new children that collaborate their parents in searching for food (optimum point). Also in our model, particles may die and be removed from their population. Two general operations, birth and death are introduced in order to make the population very dynamic and varying in size. A history is also kept for the swarm ; the swarm remembers records of its high fitness particles, even after their deaths. Compared to Genetic Algorithms and the original PSO model, GBPSO proposes a more natural simulation of the social behavior of intelligent animals. Experimental results show that GBPSO can reach broader domains and converge faster in very high-dimensional and complex search spaces.

In the second part of this talk, I will present a modified version of the ART (Adaptive Resonance Theory) algorithm for incremental learning and online clustering of handwritten digits. Our algorithm tries to minimize the number of clusters for each digit, while trying to keep the minimum similarity among the digit shapes in each cluster at a specified level. If time allows, I will briefly explain how to use GBPSO (or PSO) algorithm for optimization and search in clustering.