Efficient Discretization Scheduling In Multiple Dimensions

There is a tradeoff between speed and accuracy in fitness evaluations when various discretization sizes are used to estimate the fitness. This paper introduces discretization scheduling, which varies the size of the discretization within the GA, and compares this method to using a constant discretization. It will be shown that when scheduling the discretization, less computation time is used without sacrificing solution quality. Fitness functions whose cost and accuracy vary because of discretization errors from numerical integration are considered, and the speedup achieved from using efficient discretizations is predicted and shown empirically.