The XCS Learning Classifier System has traditionally used roulette wheel selection within its genetic algorithm component. Recently, tournament selection has been suggested as providing a number of benefits over the original scheme, particularly a robustness to parameter settings and problem noise. This paper revisits the comparisons made between the behavior of tournament and roulette wheel selection within XCS in a number of different situations. Results indicate that roulette wheel selection is competitive in terms of performance, stability and generated solution size if the appropriate parameters are used.