Solving large-scale planning and deep learning problems

Abstract

Game theory has been researched and applied in many scenarios. However, the state, action space and time of most games are set as discrete to find the optimal strategy. Hence, the primary focus of the research will be on solving problems with large-scale action space as the direct usage of existing small or discrete solutions limits the solution quality and brings less resemblance to the increasingly complex real-life situations. In particular, we approach planning: student counselling problem with large discrete action space and deep learning problem: GAN with continuous action space. Then, we propose two solutions for the counselling problem: 1) Planning Approach and 2) Learning Approach as well as two solutions for GAN: 1) Double Oracle framework for GAN (DO-GAN) and 2) Double Oracle and Neural Architecture Search for Adversarial Machine Learning (DONAS). Finally, we conduct extensive experiments to show significant improvement of our solution quality against state-of-the-art algorithms.