Nested-Wasserstein Self-Imitation Learning for Sequence Generation

Ruiyi Zhang, Changyou Chen, Zhe Gan, Zheng Wen, Wenlin Wang, Lawrence Carin
; Proceedings of the Twenty Third International Conference on Artificial Intelligence and Statistics, PMLR 108:422-433, 2020.

Abstract

Reinforcement learning (RL) has been widely studied for improving sequence-generation models. However, the conventional rewards used for RL training typically cannot capture sufficient semantic information and therefore render model bias. Further, the sparse and delayed rewards make RL exploration inefficient. To alleviate these issues, we propose the concept of nested-Wasserstein distance for distributional semantic matching. To further exploit it, a novel nested-Wasserstein self-imitation learning framework is developed, encouraging the model to exploit historical high-rewarded sequences for enhanced exploration and better semantic matching. Our solution can be understood as approximately executing proximal policy optimization with Wasserstein trust-regions. Experiments on a variety of unconditional and conditional sequence-generation tasks demonstrate the proposed approach consistently leads to improved performance.

Cite this Paper


BibTeX
@InProceedings{pmlr-v108-zhang20b, title = {Nested-Wasserstein Self-Imitation Learning for Sequence Generation}, author = {Zhang, Ruiyi and Chen, Changyou and Gan, Zhe and Wen, Zheng and Wang, Wenlin and Carin, Lawrence}, pages = {422--433}, year = {2020}, editor = {Silvia Chiappa and Roberto Calandra}, volume = {108}, series = {Proceedings of Machine Learning Research}, address = {Online}, month = {26--28 Aug}, publisher = {PMLR}, pdf = {http://proceedings.mlr.press/v108/zhang20b/zhang20b.pdf}, url = {http://proceedings.mlr.press/v108/zhang20b.html}, abstract = {Reinforcement learning (RL) has been widely studied for improving sequence-generation models. However, the conventional rewards used for RL training typically cannot capture sufficient semantic information and therefore render model bias. Further, the sparse and delayed rewards make RL exploration inefficient. To alleviate these issues, we propose the concept of nested-Wasserstein distance for distributional semantic matching. To further exploit it, a novel nested-Wasserstein self-imitation learning framework is developed, encouraging the model to exploit historical high-rewarded sequences for enhanced exploration and better semantic matching. Our solution can be understood as approximately executing proximal policy optimization with Wasserstein trust-regions. Experiments on a variety of unconditional and conditional sequence-generation tasks demonstrate the proposed approach consistently leads to improved performance.} }
Endnote
%0 Conference Paper %T Nested-Wasserstein Self-Imitation Learning for Sequence Generation %A Ruiyi Zhang %A Changyou Chen %A Zhe Gan %A Zheng Wen %A Wenlin Wang %A Lawrence Carin %B Proceedings of the Twenty Third International Conference on Artificial Intelligence and Statistics %C Proceedings of Machine Learning Research %D 2020 %E Silvia Chiappa %E Roberto Calandra %F pmlr-v108-zhang20b %I PMLR %J Proceedings of Machine Learning Research %P 422--433 %U http://proceedings.mlr.press %V 108 %W PMLR %X Reinforcement learning (RL) has been widely studied for improving sequence-generation models. However, the conventional rewards used for RL training typically cannot capture sufficient semantic information and therefore render model bias. Further, the sparse and delayed rewards make RL exploration inefficient. To alleviate these issues, we propose the concept of nested-Wasserstein distance for distributional semantic matching. To further exploit it, a novel nested-Wasserstein self-imitation learning framework is developed, encouraging the model to exploit historical high-rewarded sequences for enhanced exploration and better semantic matching. Our solution can be understood as approximately executing proximal policy optimization with Wasserstein trust-regions. Experiments on a variety of unconditional and conditional sequence-generation tasks demonstrate the proposed approach consistently leads to improved performance.
APA
Zhang, R., Chen, C., Gan, Z., Wen, Z., Wang, W. & Carin, L.. (2020). Nested-Wasserstein Self-Imitation Learning for Sequence Generation. Proceedings of the Twenty Third International Conference on Artificial Intelligence and Statistics, in PMLR 108:422-433

Related Material