Zero-shot transfer learning for multi-domain dialogue state tracking can\nallow us to handle new domains without incurring the high cost of data\nacquisition. This paper proposes new zero-short transfer learning technique for\ndialogue state tracking where the in-domain training data are all synthesized\nfrom an abstract dialogue model and the ontology of the domain. We show that\ndata augmentation through synthesized data can improve the accuracy of\nzero-shot learning for both the TRADE model and the BERT-based SUMBT model on\nthe MultiWOZ 2.1 dataset. We show training with only synthesized in-domain data\non the SUMBT model can reach about 2/3 of the accuracy obtained with the full\ntraining dataset. We improve the zero-shot learning state of the art on average\nacross domains by 21%.\n