Someone's been taking a basic game theory course!

Glad you noticed! Yes, I'm definitely curious how GGP players handle these type of games, which are pretty different from connect four, checkers, etc but still fit nicely within the GGP framework. These particular game rulesheets have been around for a few years now but I finally sat down and got them running on Tiltyard.

My hypothesis, before looking, is that most games will look like repeated defection (about 20 points each), or repeated cooperate-defection (100 to 0 points, approximately). I suspect that most players don't model their opponent too much. Those that do will tend to model their opponent as trying player-score-minimizing, or opponent-score-maximizing. In either case, the equilibrium strategy will emerge, where greedy players will always pick defect, since it's always worth more.

Now, to look at the results.

Not very many games played so far, and only a few players playing them. The highest min-score is a game between SPlayer and GreenShell, at 44-29. Most games have a player who scores under 20. In general, almost every game has a player who always defects, and a player who usually defects, but sometimes randomly picks cooperate. The decision to cooperate is, I think, likely due to not evaluating fully, rather than modeling the opponent as one who will cooperate.

i.e., consider the first game, www.ggp.org/view/tiltyard/matches/8ade3a322ad92ddd7c4e653a8b00e89fc91b4bb2/. kazk always defects (it's the first-order rational thing to do). QFWFQ1 defects 10 times in a row, then decides to start cooperating, then gives up on that. If QFWQ1 was trying to create a cooperative game, they should have cooperated in the first couple moves. I'd be interested in what motivates ggp players to pick cooperate, and whether their authors consider those reasons correct behavior or bugs.


I expect it will be several months, at least, until there's even a few 60-60 games, or even near 60-60 games. 'Good' players will tend to defect. In order to justify cooperation, players will need to model their opponents as not only opponent-score-maximizing, but model their opponents as believing that the player is score maximizing. i.e., a player needs to recognize that cooperate-cooperate is better than defect-defect, expect that their opponents recognize it, and expect that their opponents model the player as recognizing it. Then, their opponent needs to actually behave this way.

on a related note, it'd be nice to see a 'high score' list for each role in each game, showing who has the high, consistent skill ratings for that role.

Cooperation can be rational even for self-regarding agents (i.e. robots trying to climb Tiltyard's ladder). Only it is more complex to calculate sustained cooperation strategies.

Current GGPlayers are heavily biased towards zero-sum Go-like strategies. UCT was born in Computer Go and almost every GPPlayer uses some variation of it.

Someday I'll come up with an improvement to MCTS to cater for hybrid competition-cooperation games. But I'm trying to finish my PropNet first. I suspect sustained cooperation will require analyzing a much bigger part of the game tree and a fast state machine will come in handy.