Research Project Title:
Toward Human-Like Model-Based Learning in Games
abstract:Research on reinforcement learning in video game environments has produced artificial agents that exceed human performance on many games but require dozens of hours of game-play experience to produce such behavior. Humans, on the other hand, can learn to perform well on these tasks in a matter of minutes. In this paper, we propose a model that learns to play using human-inspired abstract priors about objects, agents, physics, and events to form an explicit model of the game world and show that the model quickly reaches human performance on a set of challenging games.
"Last semester, I started working on a prototype of an Atari-learning AI which was amazingly human-like in how it learned. I'm really excited to have a leading role in the final version this year, which could really change the current paradigm of video game-playing AI. It will also be an incredible opportunity to learn more about this exciting field and accrue serious research experience."