Dialogue Management with Deep Reinforcement Learning: Balancing Exploration and Exploitation

Autor:	Anna Helena Reali Costa, Bruno Eidi Nishimoto
Rok vydání:	2019
Předmět:	Training set Computer science Human–computer interaction Softmax function Supervised learning Reinforcement learning Dialog box Dialogue management Transfer of learning
Zdroj:	BRACIS
DOI:	10.1109/bracis.2019.00085
Popis:	Although there have been many advances in Natural Language Processing in recent years, the reason why dialogue systems do not take off is due to the fragility of Dialog Managers (DM). One approach that has demonstrated recent success for the development of DMs is Deep Reinforcement Learning (DRL). The main advantages are that DRL can handle stochastic environments (dialogue with different users) and does not require as much training data as in supervised learning. This paper proposes a DM system with a DRL algorithm and uses a softmax strategy combined with the transfer learning of the designer's knowledge to balance exploitation and exploration in DRL. The experimental results indicate that our proposal is quite promising.
Databáze:	OpenAIRE
Externí odkaz:	https://explore.openaire.eu/search/publication?articleId=doi_________::62b5246c8f1cb89b27a64ea10d6b5f82 https://doi.org/10.1109/bracis.2019.00085 Zobrazit plný text záznamu