A new breakthrough in artificial intelligence in the game industry

What the talk is about

Serhii Protsenko presents a concept for a service in which reinforcement-learning bots test gameplay: they assess level difficulty, balance and the variety of viable strategies. He explains the difference between supervised, unsupervised and reinforcement learning, mentions well-known game-playing breakthroughs, and notes that few companies in the industry use this approach. The bots learn from raw screen pixels, and the data scientist's key task is defining the reward function properly. The product aims to return a level assessment within hours (about half an hour for match-3) instead of days of testing. He explains why a studio would struggle to build this alone, outlines integration options (an SDK, Android Debug Bridge, client-server) and invites studios to become design partners. Useful for game designers, analysts and producers of casual and card games.