In this paper, we consider the problem of creating artificial training set containing instances of a given task with known meta-feature set. We focus on the Traveling salesman problem, however, the approach we present is applicable for any other optimization task. We modify the Smith-Miles approach for spanning of meta-feature space by introducing set diversity function. We use this quality function to reduce instance generation problem to an optimization problem. We can most fully span the whole meta-feature space by maximizing the set diversity quality function. Another method we present to sample an instance set is generating certain instances one by one, making them close to specific meta-feature vectors. In this case, we use Euclidian distance between the given meta-feature vector and obtained instance meta-features as the target function for minimization. We compare these two approaches with the na¨?ve baseline random instance generation method that did not take into account diversity of the obtained instance set. We use genetic algorithm and simulated annealing method as the optimization algorithms used on the meta-level. The value of the set diversity function for the na?ve method is equal to 0.123, for the method that directly maximizes diversity by genetic algorithm: 0.468, and for the method that generates single instances by simulated annealing: 0.595.
展开▼
