It is shown that on fairly weak conditions, the current solutions of a metaheuristic following the ant colony optimization paradigm, the graph-based ant system, converge with a probability that can be made arbitrarily close to unity to one element of the set of optimal solutions. The result generalizes a previous result by removing the very restrictive condition that both the optimal solution and its encoding are unique (this generalization makes the proof distinctly more difficult) and by allowing a wide class of implementation variants in the first phase of the algorithm. In this way, the range of application of the convergence result is considerably extended.