Incorrectly classifying Tron as an iterated game

[arr28]

We spent almost all of this match of Tron (new on Tiltyard) thinking we were going to win. Then we lost at the last minute.

GreenShell had predicted a win for some time, but we hadn't seen it.

[arr28]

Ah - the logs appear to reveal that we think it's an iterated game, using the old iterated game code. (The Nash Equilibrium version detects that it isn't suitable, and it isn't enabled even for suitable games).

[arr28]

Games which we currently identify as iterated are (according to the saved game characteristics)...

• 1429735602887 (which I think is a revised version of stanford.buttonsandlights)
• 1429792947665 (which I think is a revised version of base.gt_ultimatum, despite claiming to be v0)
• base.gt_chicken
• base.gt_coordination
• base.gt_staghunt
• base.gt_tinfoil
• base.tron_10x10
• stanford.arithmetic
• stanford.bestbuttonsandlightsbig
• stanford.untwistycomplex

I'm suspicious of stanford.buttonsandlights (from 1429735602887), stanford.arithmetic, stanford.bestbuttonsandlightsbig and stanford.untwistycomplex. We know that base.tron10x10 is wrong and hurting us.

[arr28]

The Stanford games are all single-player puzzles, so it doesn't matter whether we think they're iterated or not (because the puzzle solver takes priority).

We do need to continue to recognise the base.gt_* games as iterated though. Alternatively, if it's going to be hard to avoid categorising Tron as iterated then perhaps now is the time to finish off #108 and retire the current iterated game player.

[arr28]

The only place we set mGameCharacteristics.isIteratedGame to true is in stateMachineMetaGame. A game is marked as iterated if each player has the same moves in every turn (true for Tron) - provided that we have enough time to sample (v. likely for Tron).

However, there's also a hack which says a game isn't an iterated game if it's more than 30 turns long and isn't of fixed length.

We might well be able to drop the "longer than 30 turns" part of that hack and still catch all the games that we want to catch, without catching Tron. However, there are more base.gt_* games than those listed above. I want to go over all of those first to check whether this is a reasonable approach (until #108 is sorted).

[arr28]

Working through base.gt_*...

Game	Is iterated	Uses IteratedGamePlayer	Is fixed length
Chicken	Y	Y	Y
Coordination	Y	Y	Y
Prisoner's Dilemma	Y	Y	Y
Stag Hunt	Y	Y	Y
Tinfoil Hat	Y	Y	Y
Ultimatum	Y	Y	Y
War of Attrition	Y	- (!)	Y
Centipede	Sort of	Y	- (!)
Dollar Auction	Sort of	-	N

• I don't know why we don't currently detect War of Attrition as an iterated game.
• The proposed fix would stop us from using the iterated player for Centipede.

Needs a bit more thought.

[arr28]

Sorted (but assuming variable-length games can't be iterated) - but I've no idea how it still works for Centipede!

Centipede is being identified as a fixed length game of length 17 (after hundreds of thousands of iterations). That can't possibly be right, which implies that there's something very fishy about the sampling we do in order to determine game characteristics. I'll raise another bug for that.