This article is automatically translated.
Marltas is a system that learns to operate games posted on RPG Atsumaru, one of niconico’s services, using deep reinforcement learning. Video game AI has been a field of artificial intelligence research with various studies aimed at designing COM players and automating debugging. Traditionally, these studies have designed game-playing agents using rule-based methods. In contrast, recent approaches using machine learning, such as evolutionary computation and reinforcement learning, have gained attention. In particular, reinforcement learning has become capable of handling multidimensional states by incorporating deep learning, allowing it to handle not only old arcade games like ATARI but also complex games like StarCraft II. Marltas uses Deep Q-Learning, a method of deep reinforcement learning, to learn how to operate games. Specifically, it treats the game screen as the state and browser operations as actions, aiming to maximize the game score. By using reinforcement learning, it can acquire operation methods without relying on game-dependent knowledge such as game rules or sample operations.
