Abstract: A virtual agent with a dialogue management system and a method of training the dialogue management system is disclosed. The dialogue management system is trained using a deep reinforcement learning process. Training involves obtaining or simulating training dialogue data. During the training process, actions for the dialogue management system are selected using a Deep Q Network to process observations. The Deep Q Network is updated using a target function that includes a reward. The reward may be generated by considering one or more of the following metrics: task completion percentage, dialogue length, sentiment analysis of the user's response, emotional analysis of the user's state, explicit user feedback, and assessed quality of the action. The set of actions that the dialogue management system can take at any time may be limited by an action screener that predicts the subset of actions that the agent should consider for a given state of the system.