Continuing beyond NFL: dissecting real world problems
Oliver Sharpe · 1999
In this paper it is argued that the classification of search spaces by the performance of different algorithms on them is an essential part of a principled methodology for tackling new search problems. It is further argued that the success of a given algorithm on a given problem can often be attributed to particular aspects of the generation or selection mechanisms used in the algorithm. This hypothesis is examined by removing different parts of the algorithm and observing the resulting changes in its performance. It is shown how dissecting the algorithm in this way can give rise to greater insight into how the algorithm is working on a given search space and at the same time help classify the nature of that space.