Abstract: This talk is about designing an algorithm that can learn to play the game of Clash Royale (a popular multiplayer real-time strategy game from Supercell). Learning to play Clash Royale is challenging due to partial observability, presence of cyclic strategies and large state-action spaces. Previous attempts to tackle such challenging games have focused on model-free learning methods, often requiring hundreds of years of experience to produce competitive agents. In this talk, I will present a model-based planning approach to this problem. We tackle the problem of partial observability by first building an (oracle) planner that has access to the full state of the environment and then distilling the knowledge of the oracle to a (follower) policy network which is trained to play the imperfect-information game by imitating the oracle's choices. Our results show that the oracle is able to discover efficient playing strategies and the follower policy successfully learns to implement them by training on a handful of battles. An arXiv pre-print of the work is available here: https://arxiv.org/abs/2012.12186

Speakers:  Rinu Boney

Affiliation: Aalto University

Place of Seminar:  Zoom (Available afterwards on Youtube)

Meeting ID: 668 2513 3920

Passcode: 577473