A multi-agent collaborative algorithm for task-oriented dialogue systemspreprint
Аннотация: Abstract In recent years, reinforcement learning has been successfully applied to dialogue systems. However, in the face of task-oriented dialogue systems where policies are difficult to optimize, states are difficult to track, and tasks are multiple and compound, task-oriented dialogue systems based on reinforcement learning have problems such as poor collaboration, non-unique learning goals, and non-stationarity due to the lack of collaboration among agents. In this paper, we propose a multi-agent cooperative dialogue (MACD) algorithm for task-oriented dialogue systems, in which, for the information interaction between multi-agent in task-oriented dialogue systems, a deep neural network approach is used to integrate the observations of multiple single agents and obtain joint observations to achieve information sharing among single agents, to solve the non-stationarity caused by the lack of joint information of multi-agent. For multi-agent policy learning task-oriented dialogue systems, the multi-agent deep deterministic policy gradient (MADDPG) architecture is applied to the policy selection of task-oriented dialogue systems to solve the problem of lack of joint policy learning of multi-agent in task-oriented dialogue systems; the observation integration of single agents and multi-agent policy learning are effectively combined to solve the problem of poor multi-agent collaboration in task-oriented dialogue systems. By verifying and analyzing reinforcement learning algorithms such as MACD, REINFORCE, DQN, and QMIX in the MultiWOZ corpus, the experimental results show that the algorithms effectively improve the success rate of multi-agent working together to complete dialogue tasks, reduce the number of invalid dialogues in dialogue turns, and outperform common reinforcement learning algorithms in terms of agents' information interaction and joint policy learning in the composite task dialogue scenario.
Год издания: 2023
Авторы: Jingtao Sun, Jiayin Kou
Издательство: Research Square (United States)
Источник: Research Square (Research Square)
Ключевые слова: Speech and dialogue systems, Topic Modeling, Multi-Agent Systems and Negotiation
Открытый доступ: green