Generative Adversarial Nets with Reinforcement Learning

Generative adversarial nets (GANs) have been well studied and applied to various applications of data generation since 2014. Nonetheless, most of the studied data generation tasks focus on continuous data such as image and speech, while the discrete (sequence) data such as text is far less studied due to the difficulty of directly taking gradient on the discrete tokens. As such, reinforcement learning (RL) approaches start to be considered as the solution to generate the discrete data. By modeling the generator as a stochastic RL policy and training it via policy gradient methods, it is promising to bypass the discrete data gradient problem and generate high-quality discrete data. In this talk, I will introduce the fundamental SeqGAN model for discrete (sequence) data generation, and then the LeakGAN model with information leaking to further improve the learning effectiveness in the interplay between the generator and discriminator. Finally, I will extend the discussion of GANs to multi-agent RL and show some demos of our recent work.

BIO: Weinan is now a tenure-track assistant professor in Department of Computer Science, Shanghai Jiao Tong University. His research interests include machine learning and big data mining, particularly, deep learning and (multi-agent) reinforcement learning architectures, mechanisms, training algorithms and their applications in real-world data mining scenarios including computational advertising, recommender systems, text mining, web search and knowledge graphs.