Bayesian optimization with approximate set kernels