A Study of the Effectiveness of Detailed Balance in Avoiding Convergence in PBIL

摘要

One of the claimed advantages of EDAs over other evolutionary algorithms is that they involve less parameters to be chosen by the user. Although this generalisation may be true it is undeniable that there still exist decision parameters in EDAs. In particular, PBIL was shown to be strongly sensitive to such a parameters. As important as diminishing the number of parameters in an algorithm is making the algorithm robust regarding changes in its parameters. This work investigated the performance of the Detailed Balance PBIL on two different test problems and also extended its application to a non-binary encoding. On both test problems DBPBIL was shown to be more robust than the standard PBIL. The results obtained suggest that the addition of detailed balance to PBIL is indeed an efficient method for diminishing the convergence pressure imposed by the probabilistic model in PBIL. DBPBIL has not only found the global optimum a greater number of times, but has also found better solutions than PBIL for practically all values of the parameter α. Turning to other EDAs, an interesting question would be: does the convergence pressure imposed by the probabilistic model affect the performance of the more complex EDAs models?