The objective of video anomaly detection is to distinguish events within videos that deviate from expected normal behavior. An effective anomaly detection model requires strong spatio-temporal feature extraction capabilities to capture both appearance and motion information from the video. We utilize a generative adversarial networks model to perform anomaly detection by predicting future frame. The model is trained with multi-objective loss function for appearance and motion constraints. We introduce multi-objective optimization algorithm to guarantee the convergence of training objectives, including intensity loss, gradient loss for appearance constraints, optical flow loss for motion constraints and adversarial loss for adversarial training. By ensuring corresponding generative outcome for both normal events to confirm to expectation and anomaly events not to do so, we obtain an anomaly detection model with satisfying experimental result on video anomaly detection datasets, showing the success of the proposed training strategy.