Multi-task Reinforcement Learning for Video Games

An important component of modern video games is the non-player character (NPC), moving entities in-game that are not controlled by a human, which may cooperate with, oppose, or otherwise interact with the player. For an NPC to interact with the game word it must often perform complex tasks that are difficult to program explicitly. Research has explored using artificial intelligence, particularly reinforcement learning, to train NPCs to achieve the desired behavior, but prior work has often focused on training NPCs only within one game. Our goal is to investigate using multi-task reinforcement learning to train NPCs that are more robust and can easily transfer from one gaming task to another without needing to be retrained. Success could, in the long term, lead to downstream development of artificial intelligence that can transfer reliably between different use cases.

Faculty Supervisor:

Florian Shkurti

Student:

Partner:

AMD Canada

Discipline:

Computer science

Sector:

Manufacturing; Professional, scientific and technical services

University:

University of Toronto

Program: