Unsupervised Domain Adaptation (UDA) aims to learn a predictor model for an\nunlabeled domain by transferring knowledge from a separate labeled source\ndomain. However, most of these conventional UDA approaches make the strong\nassumption of having access to the source data during training, which may not\nbe very practical due to privacy, security and storage concerns. A recent line\nof work addressed this problem and proposed an algorithm that transfers\nknowledge to the unlabeled target domain from a single source model without\nrequiring access to the source data. However, for adaptation purposes, if there\nare multiple trained source models available to choose from, this method has to\ngo through adapting each and every model individually, to check for the best\nsource. Thus, we ask the question: can we find the optimal combination of\nsource models, with no source data and without target labels, whose performance\nis no worse than the single best source? To answer this, we propose a novel and\nefficient algorithm which automatically combines the source models with\nsuitable weights in such a way that it performs at least as good as the best\nsource model. We provide intuitive theoretical insights to justify our claim.\nFurthermore, extensive experiments are conducted on several benchmark datasets\nto show the effectiveness of our algorithm, where in most cases, our method not\nonly reaches best source accuracy but also outperforms it.\n