Reinforcement Learning in Action - Creating Arena Battle AI for Blade & Soul

Explore the NCSOFT team's journey in developing pro-level AI agents for arena 1v1 battles in Blade & Soul during this 2019 GDC session. Delve into the complexities of reinforcement learning as applied to game AI, covering topics such as skill systems, real-time response, and generalization. Learn about the team's research goals, the challenges of high complexity in the game environment, and their innovative approaches to guiding fighting styles. Discover techniques for move space reduction, skill space reduction, and feature engineering. Gain insights into the learning process, including the use of sparring partners and self-play. Examine the development of different fighting styles and the pretest phase with professional gamers. Acquire valuable knowledge on creating sophisticated AI for competitive gaming scenarios in this comprehensive 30-minute talk.

Intro
B&S (Blade and Soul)
B&S Arena Battle
Research Goal
Skill Systems of B&S
High Complexity
Real Time Response
Generalization
Guiding Fighting Style
Leaming how to walk
Additional Reward for Guiding Battle Style
Agent Environment Plot
Learning Process
Move Space Reduction - Decision Frequency
Skill Space Reduction
Feature Engineering
Sparring Partner - Built-in AI
Sparring Partner - Self play
Learning Progress
Different Fighting Styles
Pretest with professional gamers
Conclusion