Chance-constrained genetic algorithms
Daniel H. Loughlin, S. Ranji Ranjithan · 1999
The applicability of classical mathematical optimization approaches is often limited by the necessity for strict mathematical formulations. As a result, powerful tools, such as chance-constrained optimization, have seen limited use. Genetic algorithms (GAs) are much more flexible than traditional mathematical programming, allowing many problems that were previously deemed intractable to be revisited. In this paper, we explore an approach called a chance-constrained GA (CCGA) that uses a Monte Carlo (MC) simulation in the GA fitness function. Previous applications of CCGA in the literature have required several orders of magnitude of increased computational demands over conventional GAs. To reduce these demands, we test and compare alternative MC sampling approaches and identify an approach that works efficiently with little loss in solution quality. This approach, which uses Latin Hypercube Sampling to generate a new set of only 10 to 50 realizations each generation, is demonstrated for solving an air quality management problem.