Blade & Soul

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

The development of professional-level artificial intelligence for the 1v1 arena battles in the MMORPG Blade & Soul demonstrates the successful application of reinforcement learning (RL) to complex, real-time fighting games.

• NCSOFT replaced traditional hand-written AI rules in 'Blade & Soul' with reinforcement learning (RL) to create pro-level agents capable of strategic, multi-stage combat sequences.
• To ensure robustness and generalization, the team implemented a 'self-play' method where agents were trained against a growing pool of diverse, previously trained sparring partners rather than a single fixed opponent.
• The AI was trained to exhibit specific fighting styles—such as aggressive or defensive—by adjusting reward signal weights, allowing the agents to prioritize either damage output or HP preservation without hard-coded logic.