A1061
Title: Lifting the veil: Unlocking the power of depth in Q-learning
Authors: Shao-Bo Lin - Xi'an Jiaotong University (China) [presenting]
Abstract: With the help of massive data and rich computational resources, deep Q-learning has been widely used and received great success in numerous applications, including recommender systems, games and robotic manipulation. Compared with avid research activities in practice, there is a lack of solid theoretical verifications and interpretability for the success of deep Q-learning, making it a little bit mystery. The aim is to discuss the power of depth in deep Q-learning. In the learning theory framework, it is rigorously proven that deep Q-learning outperforms the traditional one by showing its good generalization error bound. The results show that the main reason for the success of deep Q-learning is due to the excellent performance of deep neural networks (deep nets) in capturing special rewards properties, such as the spatially sparse and piecewise constant, rather than due to their large capacities. In particular, answers are provided to questions about why and when deep Q-learning performs better than traditional ones and how the generalization capability of deep Q-learning.
