Wargaming with AlphaZero


  • Mika Cohen
  • Farzad Kamrani
  • Fredrik Bissmarck
  • Peter Hammar

Publish date: 2021-01-19

Report number: FOI-R--5057--SE

Pages: 26

Written in: Swedish


  • Artificial intelligence
  • machine learning
  • deep reinforcement learning
  • AlphaZero
  • wargame
  • the game of Risk


Optimization algorithms playing wargames have long seemed like science fiction. Then, to the astonishment of the research community, a new optimization algorithm, AlphaGo, defeated in 2016 the world champion in Go, an ancient strategy game revered not least within wargaming communities. What could be considered feasible changed over night. AlphaGo culminated a few years later in AlphaZero, a general optimization algorithm that masters a strategy game to super-human strength through a process of trial-anderror. AlphaZero has since renewed tactics and strategy in a number of classic strategy games. This report introduces AlphaZero and its possible application to wargames. The report shows how AlphaZero, which plays abstract strategy games in the research literature, can be made to play also a (simple) wargame - the game of Risk. Tentative results indicate that the algorithm needs to be given some simple heuristics (domain knowledge) in order to reach the level of a human expert or beyond.