The anomaly detection in automated video surveillance is considered as one of the most critical tasks to be solved, in which we aim to detect a variety of real-world abnormalities. This paper introduces a novel approach for anomaly detection based on deep reinforcement learning. In recent years, deep reinforcement learning has been achieving a significant success in various applications with data of a high degree of complexity such as robotics and games, by mimicking the way humans learn from experiences. Generally, the state-of-the-art methods classify a video as normal or abnormal without pinpointing the exact location of the anomaly in the input video due to the unlabeled clip-level data in training videos. We focus on adapting the prioritized Dueling deep Q-networks to the anomaly detection problem. This model learns to evaluate the anomaly in video clips by exploiting the video-level label to obtain a better detection accuracy. Extensive experiments on 13 cases class of real-word anomaly show that our DRL agent achieved a near optimal performance with a high accuracy in the real world video surveillance system compared to the state-of-the-art approaches.