AgentGraph: DM with Structured DRL

Recently, DRL and its variants have been used successfully in Dialogue Management to learn a domain specific policy for a specific task. But, they all suffer from scaling limitations i.e. not sample efficient and not able to transfer policy one domain to another. Hence, AgentGraph that is a combination of graph neural network (GNN) and DRL-based algorithm. On experimenting, the proposed method has outperform traditional reinforcement learning approaches.

This article explores on how the use of AgentGraph with a structured model of Deep Reinforcement Learning Improves the Dialogue Management over traditional approaches. Before going deep, let's introduce what is Dialogue Management. Dialogue Manager has mainly two task to perform: one is track the dialogue state and second is to respond the user using a policy.