Game observations: Context matters


The Halite III competition is halfway through! We’ve compiled a few observations on gameplay so far.

In Halite III, the number of players, the specific players, the size of the map, and the amount of halite on the board all impact game outcome. As every Halite board is different, the game exposes us to the challenge of writing AI that is effective in a number of (often, new!) contexts. Within the industry there is focus on developing systems that can operate in broader domains with less exhaustive training.

We see differences in the types of strategies that are successful in 4-player games and 2-player games, and we see great variability in individual bots’ performance across different game and map types, and vs. different opponents.

4-player games seem more competitive and more interactive than 2-player games:

  • Games are closer in score: The average win margin (difference between the 1st & 2nd place bot scores) in 4-player games is nearly 60% less than it is in 2-player games.
  • More opponent ships interact: Winning bots have on average 30%+ more collisions, and 50%+ more bonus halite collected from being “inspired”, or close to another ship, than they do in 2-player games.
  • Interaction is more important: In four-player games, the average score bonus provided by “inspiration”, the state enacted when close to multiple opponents, is more than 2-times the win margin. In 2-player games, the average inspiration bonus is less than the win margin.

2-player games provide opportunities for emergent aggressive strategies:

  • Players have implemented aggressive “blocking” and “rushing” strategies.
  • This base-blocking strategy is only activated in 2-player, 32x32 games.
  • This game shows a bot intentionally colliding with opponent ships.

Outcomes vary with context:

  • Individual bots perform differently across different board sizes, opponents, as @mlomb’s community-contributed tool shows.
  • Same players, different outcomes: Two games between four of the top players, with different outcomes: here, here

Those of you actively playing continue to make improvements! 70% of those who have submitted a bot have submitted at least one update to their bot.

We’re excited to see the developments in gameplay over the next several weeks.


Is this observation an argument for having separate tournaments?

Halite III Midseason Update