A multi-agent collaborative algorithm for task-oriented dialogue systems

In recent years, reinforcement learning has found successful applications in dialogue systems. However, when confronted with task-oriented dialogue systems, optimizing the strategy becomes challenging due to difficulties in state tracking and the complexity of multitasking. Task-oriented dialogue systems based on reinforcement learning encounter issues such as poor collaboration, non-unique learning goals, and non-staticity arising from the absence of agent cooperation. To address these challenges, this paper introduces a novel multi-agent cooperative dialogue (MACD) algorithm for task-oriented dialogue systems. In MACD, a deep neural network framework is employed to facilitate information exchange among multiple agents within task-oriented dialogue systems. This integration enables the consolidation of observations from individual agents, leading to the derivation of joint observations and fostering information sharing among the agents. Consequently, MACD aims to mitigate the problem of non-stationarity resulting from the lack of shared information among multiple agents. In the context of multi-agent strategy learning within task-oriented dialogue systems, we employ the MADDPG architecture to address the challenge of inadequate joint strategy learning among multiple agents. By integrating single-agent observations and multi-agent strategy learning, we aim to alleviate the collaborative deficiencies inherent in task-oriented dialogue systems involving multiple agents. Through experimentation with reinforcement learning algorithms such as MACD, DQN, OPPA, JOIE, and QMIX on the MultiWOZ 2.0 corpus, our results demonstrate significant enhancements. Specifically, the proposed algorithm effectively elevates the success rate of multi-agent collaboration in accomplishing dialogue tasks in composite task scenarios. Furthermore, it mitigates the occurrence of ineffective dialogues during the dialogue rounds. Comparative analysis reveals that our approach surpasses conventional reinforcement learning algorithms in facilitating agent information interaction and joint strategy learning within the task-oriented dialogue context.