In agent-based traffic simulation, calibration is an essential stage before the models applied to reproduce the individual/group travel behaviors. While traditional methods suffer from a high computational complexity, this paper proposes an improved method to alleviate the computational burden for large-scaled simulations. Specifically, we introduce variational auto-encoder to compress the original agent state vector into a lower dimensional hidden space, where the state transfer probability is calculated fast. Then the probability is mapped into the original space through a decoder, to achieve the agent travel parameters. The dynamic calibration method is tested with other baselines in urban travel demand analysis. Experiment results demonstrate that our method brings about 19% elevation of efficiency with the same accuracy of calibration.