Artificial player for collectable card games using deep machine learning and psychological misdirection

2020 Fiscal Year Final Research Report

• Project/Area Number: 19K20429
• Research Category: Grant-in-Aid for Early-Career Scientists
• Allocation Type: Multi-year Fund
• Review Section: Basic Section 62040:Entertainment and game informatics-related
• Research Institution: Sasebo National College of Technology
• Principal Investigator: Sato Naoyuki 佐世保工業高等専門学校, 電子制御工学科, 講師 (30826889)
• Project Period (FY): 2019-04-01 – 2021-03-31
• Keywords: ゲーム情報学 / 探索

Outline of Final Research Achievements

[Research aim] This work aims to develop a competitive artificial game player which can win against expert human players in genres of imperfect information game. Two approaches are proposed, which are "using deep convolutional neural network" and "using tree search with misleading human expectation techniques." This research regards 'collective card game' genre as the final goal to implement strong artificial players.
[Progress] Proposed tree search player was implemented and evaluated its performance through matches against other artificial players in the game of "Geister."

Free Research Field

ゲーム情報学　探索

Academic Significance and Societal Importance of the Research Achievements

本研究は人間の『だまし・ブラフ』を再現したゲーム内着手を行う手法を提案した．従来型の手法は，相手となるプレイヤ（主な想定として，人間）に関しての挙動のデータを多く必要とする．しかし本研究の提案手法は，性能の上限が多少低くなる事と引き換えに，相手に関するデータを必要としない．
この研究による提案手法は，学術的には『木探索のみによる新規な“だまし”着手の生成方法の提案』という意義を持ち，社会的には『相手のデータを必要とせず，気軽に利用できる，だましを行うコンピュータゲームプレイヤ』の実現方法の提供，という有用性を持つ．