Configuring microgenetic algorithms for solving traffic control problems: the case of number of generations

Efficient and successful use of genetic algorithms (GAs) requires careful selection of several parameter values. One such critical parameter is the processing time (or, number of generations) that is sufficient to ensure suitable convergence. Todate there is only limited guidance on this subject, and in most cases detailed knowledge of the structure and properties of the problem is necessary for such guidance to be useable. For real world problems such knowledge may not be readily available. We describe an experimental approach to establish relationships between time to convergence and problem size of microgenetic algorithms (m-GAs). A discrete time dynamical traffic control problem with different sizes and levels of complexity was used as a test bed. The results showed that upon appropriately sizing the m-GA population, the m-GA can converge to a near-optimal solution in a number of generations equal to the string length. The results also demonstrate that with the selection of appropriate number of generations, it is possible to get most of the worth of the theoretically optimal solution but with only a fraction of the computation cost. The results showed that as the size of the optimization problem grew exponentially, the time requirements of m-GA grew only linearly thus making m-GAs especially suited for optimizing large scale and combinatorial problems for online optimization