A modified deterministic/stochastic genetic algorithm (DS-GA) method is proposed for the determination of the global minimum of atomic clusters described by pairwise analytic interaction potentials. Our modification of the standard GA method involves a coarse local minimization of each member of the population at every generation, as well as including the gradient into the fitness function. For Lennard-Jones (Ar)(n) clusters with n<30, the DS-GA converges far more quickly to the global minimum than either conventional GA methods or random search procedures. An application of this DS-GA is made to heterogeneous clusters of B(P-2) with multiple Ar atoms. The interaction potential is given by the lowest state of a 3X3 electronic Hamiltonian. The Ar-Ar potential and the lower energy (IT state) B-Ar potential are very similar. In contrast, the higher energy (Sigma state) B-Ar interaction is essentially repulsive. Consequently, the B atom is nearly always found to substitute for one of the atoms in the corresponding (Ar)(n+1) cluster with the fewest number of nearest neighbors.