本文へジャンプ

Not only Strong, but also Human-like, Entertaining or Educating Computer Game Players

K.IKEDA Laboratory
Associate Professor: IKEDA Kokolo

E-mail:
[Research areas]
Game Informatics, Machine Learning
[Keywords]
Computer Game Player, Entertainment, Education, Procedural Content Generation, Reinforcement Learning, Genetic Algorithm, Monte-Carlo Tree Search

Skills and background we are looking for in prospective students

  • Basic Mathematic concepts, such as set, function, probability, combination and graph.
  • Intermediate programming (100+ lines) skill on any language, such as C#, Java, Python.
  • Enthusiasm for games, further, complaints about games.

What you can expect to learn in this laboratory

  1. Expert knowledge about Game Informatics, such as MCTS or Reinforcement Learning
  2. IT skills including AI such as Machine Learning or Optimization, and programming skills
  3. General skills as intelligent workers, such as writing report, presentation, problem solving, problem finding, analysis, scheduling and self-control.

【Job category of graduates】 Mainly System Engineer or Researcher, sometimes Game Developer

Research outline

[Background]

Game is very important culture of human beings. We play many games, and some of them such as Chess have been considered as the symbol of intelligence. Recently, according to the progress of artificial intelligence (AI) technologies and computer itself, very strong computer players have been achieved in many games, such as Chess or Go. However, this fact doesn't mean the end of Game Informatics, but mean the beginning of advanced Game Informatics for entertaining human players.

[Human-like Computer Players]

We human players are different from computer players, in many points. Physically, some noise is introduced to image recognition, thinking and neurotransmission make a delay of action, etc. Psychologically, we enjoy, fear, or are surprised. To understand the human-likeness and to reproduce it as a human-like computer player are valuable topic now. Ikeda-lab intensively study this topic, and apply the technology to many games and many purposes.

[Entertaining and Educating Players]

It is well known that a strong player is not necessarily a good coach. Much more skills are needed to be a good coach than a strong player. Good coach must understand the skill and preference of student, play gently and naturally, give a chance student to win, and praise well, and point some wrong moves of student with training problems. Some good results have been achieved in coaching Go, but we want to improve and extend such coaching AI method more deeply.

[Procedural Content Generation]

Recently, game contents are created not only by hand, but also automatically. Players can enjoy randomly made dungeons, stages or puzzles. However, sometimes the contents are too easy or too difficult for human players, mainly because such contents are tested/evaluated by not-human-like AI agents. Ikeda-lab tries to understand the weakness of human players, make human-like test agent, and then generate good contents special to human players, or to a player.


Figure: six requirements for entertaining player

Key publications

  1. Tianwen Fan et.al., Position Control and Production of Various Strategies for Deep Learning Go Programs, TAAI, 2019
  2. SangGyu Nam and Kokolo Ikeda, Generating Stages in Turn-Based RPG using Reinforcement Learning, IEEE CoG, 2019
  3. Kokolo et.al., Detection and Labeling of Bad Moves for Coaching Go, IEEE-CIG, 2016

Equipment

  • PC with GPU for each
  • 10 UNIX servers with 24-32 cores for lab

Teaching policy

  • Each student should propose an idea based on his/her interest and complaint about a game.
  • A research should be proceeded by repeating research cycles: (survey-) consideration - implementation - experiment - report - elder's comment.
  • By these cycles, students will learn and improve their skills (1)(2)(3) mentioned above.
  • Finally the result should be published in a conference about games or AIs.

[Website] URL:http://www.jaist.ac.jp/is/labs/ikeda-lab/

PageTop