XML-based genetic rules for scene boundary detection in a parallel processing environment

Abstract

Genetic programming is based on Darwinian evolutionary theory that suggests that the best solution for a problem can be evolved by methods of natural selection of the fittest organisms in a population. These principles are translated into genetic programming by populating the solution space with an initial number of computer programs that can possibly solve the problem and then evolving the programs by means of mutation, reproduction and crossover until a candidate solution can be found that is close to or is the optimal solution for the problem. The computer programs are not fully formed source code but rather a derivative that is represented as a parse tree. The initial solutions are randomly generated and set to a certain population size that the system can compute efficiently. Research has shown that better solutions can be obtained if 1) the population size is increased and 2) if multiple runs are performed of each experiment. If multiple runs are initiated on many machines the probability of finding an optimal solution are increased exponentially and computed more efficiently. With the proliferation of the web and high speed bandwidth connections genetic programming can take advantage of grid computing to both increase population size and increasing the number of runs by utilising machines connected to the web. Using XML-Schema as a global referencing mechanism for defining the parameters and syntax of the evolvable computer programs all machines can synchronise ad-hoc to the ever changing environment of the solution space. Another advantage of using XML is that rules are constructed that can be transformed by XSLT or DOM tree viewers so they can be understood by the GP programmer. This allows the programmer to experiment by manipulating rules to increase the fitness of a rule and evaluate the selection of parameters used to define a solution.