Source
Czarnecki, Gidel, Tracey, Tuyls, Omidshafiei, Balduzzi & Jaderberg (2020), LIT-1504, read in full in NOTE-1146 from arXiv:2004.09468v2.
The thesis
Games people play for skill have a transitive axis, from weak to strong, and a non-transitive part that is thickest at middling strength and thin at both ends: a spinning top. The abstract says “We prove the existence of this geometry”; the body withdraws it: “Naturally, the above does not prove that real world games follow the Games of Skill geometry” (p. 6).
What the evidence is
THEORY-930: about a thousand sampled search agents in each of eight small OpenSpiel board games, and the AlphaStar league, with clusters taken from maximum-entropy Nash and a skew-normal curve fitted by eye. The authors say their sampler “does not guarantee uniform sampling of strategies” (p. 9), that the AlphaStar profile reflects “AlphaStar induced game geometry” (p. 18), and that minimum-entropy equilibria would give “more peaky tops” (p. 17).
What this does not say
- Anything from the theorems. Theorem 1 (THEORY-926) shows long cycles among strategies that can signal, not narrowing at the top.
- That every game has the shape. Blotto, Kuhn poker and the noisy Elo games do not.