2017 TGDF Notes: Game AI and Level Difficulty

Speaker | Professor, National Taiwan University of Science and Technology | Wen-Kai Tai

Wen-Kai Tai is a professor at National Taiwan University of Science and Technology and a member of the GAME Research Center and GAME Lab. His work covers Procedural Content Generation, game design automation, game AI, player modeling, multimedia applications, industry-academia R&D, consulting, and government research projects.

─ Excerpted from TGDF official

These are personal notes and may not fully represent the original speaker’s intent.

Other articles in this series

• 2017 TGDF Notes: Creating 3D Japanese Animation Quality in a Small Indie Studio
• 2017 TGDF Notes: Game AI and Level Difficulty
• 2017 TGDF Notes: Finding a Path in a Crowded Market
• 2017 TGDF Notes: Making the Future Present Through Rez Infinite
• 2017 TGDF Notes: Lanota Development Experience
• 2017 TGDF Notes: Postmortem on What Went Right and Wrong
• 2017 TGDF Notes: The Art Direction of Detention
• 2017 TGDF Notes: Visual Design Notes from Qubot
• 2017 TGDF Notes: Survival Rules for Game Software Engineers

AI goals

• Tutorial
  • Help new players learn
  • Provide different difficulty settings
• Training
  • Experienced and competitive players often want to practice
  • Strategically practice projects of certain difficulty
• Replayability
  • More dynamic game value settings
• Robust to change
  • Properties in the game that may need to be adjusted over time
  • Maintenance is required every time it is updated, saving programmers valuable time

Game process model

• Game
  • Game = current game state + rules/logic
  • Current game state = (previous game state + previous player) + player set
• State
  • Game state = current game time plus player data such as HP, troop strength, and country
• Player
  • Player action decision = player + game status + player information
  • Player action decision = composed of multiple different actions
• Player Action (Partial Player)
  • Player action = action algorithm + game state + player information
  • Example: Are costs considered during production? Are population limits considered?
• AI Portfolio
  • AI portfolio = composed of multiple action algorithms

Game case: 8-ball

Game flow:

Decision process:

Player portfolio -> Strike action generation -> Table state evaluation -> Ball selection -> Shot execution

Player portfolio: strategy and skill settings.

Strike action generation:

Select hittable ball -> Select possible pocket -> Choose shot type -> Refine aim direction -> Search and optimize force

Table state evaluation:

Hit difficulty evaluation -> Ball evaluation -> Ball group evaluation -> Shot evaluation

Select target ball:

• Random numbers
• Rule resource size
• Monte Carlo tree search (MCTS) thinking time length

AI portfolio:

• Shot generation
• Table state evaluation
• Shot evaluation
• Select target ball

Control AI through profiles:

Table state estimation - combat behavior:

Table state estimation - core process architecture:

Table state estimation - experimental results:

References

• [GDC 2017 Sharing Session] AI in Games

Attribution

Please credit ARKAI Studio and link back to this article when quoting or reposting.