An Empirical Study with BRKGA Using Integer Random Keys

The genetic algorithm of biased random keys (BRKGA) is a meta-heuristic widely recognized in the literature. Its implementation requires only one specific component of the problem, the decoder, responsible for converting random keys into assessable solutions. In the original proposal, these keys are real numbers, leading many researchers to overlook their accuracy. This work empirically investigates the impact of key format, comparing two implementations of BRKGA: one with double-precision variables and the other with controlled-interval integers. The problem addressed is the classic traveling salesman (TSP). Both versions were tested with instances of the literature. The results show that the version with whole keys, when well configured, achieves performance equal to or superior to the double-precision version. This suggests that the accuracy of the keys is a relevant parameter and should be calibrated carefully.