深層学習と強化学習の発展に伴い,コンピュータ囲碁の実力は人間のトッププレイヤを超えた.一方で,初中級者の教育やエンタテインメント方面での研究はまだ十分に行われていない.例えば初中級者と対戦して楽しませるためにはコンピュータ側に意図的に悪い手を打たせる必要があるが,評価関数にノイズを加えるなどの静的な方法,現在の勝率に応じて勝率の低い手を打たせるなどの動的な方法,それぞれに課題がある.本稿では,AlphaGo Zeroモデルに基づくプログラムLeela ZeroとELF OpenGoを採用したうえで,既存の勝率制御法を再現し,その有効性を検証する.そしてプログラムのモデルが変わったこと,より強くなったことによる新たな課題を発見し,その緩和法を提案し,評価することを試みる.:In recent year, Computer Go AI has exceeded the top-level human player by the advancement of Deep Learning and Reinforcement Learning techniques. However, the other approach for “Entertainment Go AI” or “Coaching Go AI” are still received less attention. But several approaches have been made to entertain the beginner or intermediate level of player. For example, to control strength by “static methods” that adding noise into evaluation function, or “Dynamic” approach which selects low winning ratio move according to current winning ratio, etc. However, it still appears that some task to be improved. In this paper, we present the reproduction and validation of the existing strength control method using the "AlphaGo Zero"-based program Leela Zero and ELF OpenGo. Also, it is reasonable to try to find and evaluate some novelty research approaches and ideas using newer architecture and stronger methods.