저널 : JICRS(Journal of Institute of Control, Robotics and Systems), Vol.24, No.9, pp.829-837, 2018. 9.
· 논문제목 : “Deep Reinforcement Learning for Optimizing Visual Questions”
· 저자 : 조영수, 황지수, 김인철
· 요약 : This paper presents a novel game environment, GuessWhat+, for visual dialogue research,and proposes an efficient deep reinforcement learning algorithm, MRRB, for optimizing visual questions. GuessWhat+ is an extended version of the existing visual dialogue environment, GuessWhat?!. In order to overcome the limitations of GuessWhat?!, it enables the participating agents to utilize immediate rewards from games. The proposed deep reinforcement learning algorithm, MRRB(Mini-Batch REINFORCE with Return Baseline) is a new policy gradient algorithm to meet both the data inefficiency problem and the unstable convergence problem. Experiments showed the usefulness of the GuessWhat+ environment and the high performance of the proposed MRRB algorithm.