[2Q4-OS-27b] 強化学習の新展開

[2Q4-OS-27b-05] Top-level battle AI of The Modern Mobile Game "Gyakuten Othellonia" by Using Reinforcement Learning

〇Katsuki Ohto Ohto¹, Yuta Goto², Yu Kono¹, Ikki Tanaka¹ (1. DeNA Co., Ltd., 2. Tokyo University of Agriculture and Technology)

Keywords:Reinforcement Learning, Game AI

We developed an AI agent based on deep reinforcement learning for "Gyakuten Othellonia," an online game that adds elements of card games to the base of Othello. Our AI greatly surpasses the previous AI developed through prior research and has been evaluated as top-level by human players. Furthermore, we discuss the practical measures taken to maintain the strength of the AI in a game that continues to add new elements and change its rules on a daily basis, leading to an improvement in development efficiency.