Course details

8.3425

The Hanabi Challenge: A New Frontier for AI Research (Part 2)

 SP
SS 2020 Prof. Dr. Gordon Pipa, Prof. Dr. Michael Franke HYBRID
6h/wk
12 ECTS 		M.Sc modules:
CC-MP-SP - Study Project
CS-MP-SP - Study Project
 Thu: 16-20

After AlphaGo in 2016 and other recent successes with deep reinforcement learning, in February 2019 DeepMind published another paper where they reported that their approach was unsuccessful when applied to the card game Hanabi and proclaimed it to be "A New Frontier for AI Research". [N. Bard et al., 2019, https://arxiv.org But what makes Hanabi, a seemingly simple card game in some respects more challenging for AI than the complex game of Go? Hanabi is a cooperative card game with incomplete information which relies on efficient restricted communication between the players. In this study project, we will explore the challenges underlying the game of Hanabi, both from a human's and an AI's perspective. Using concepts from Machine Learning (e.g. Reinforcement Learning), Linguistics (e.g. pragmatics, rational speech act model), Psychology (e.g. Theory of Mind) and potentially other disciplines and methods (e.g. behavioral studies), we aim at constructing a cognitive architecture able to play Hanabi successfully. At the same time, we will investigate human strategies and conventions that evolve during game play to gain insight into the underlying cognitive processes.