According the the paper, they used Freeciv, because it is open source and they could modify the source code. Their source code is available here:
I'm not sure the baseline AI of civ is that hard to beat, so its hard to evaluate how good the headline stat of 78% is? Anyone?
The other thing that gives me pause, is that they performed their rollouts against the other AI player. That seems kind of unfair - might be concerned they might be specialising to beat the build in AI, rather than to play the game.
Finally, did they perform their rollouts during a training phase of their model, and then use just the fitted model during evaluation play, or did they continue to use the rollouts during evaluation? I presume it was the former, but I only skimmed the paper, and couldnt find it explicitly?
They were training online during gameplay, so each round started fresh with an untrained system.
Yes; but, like, Samuel's learning checker playing programs in the 50s and 60s did that.
>They were training online during gameplay, so each round started fresh with an untrained system.
I'm not really sure how that is considered training. Like Tim_Benham's comment, on a first read of the paper, I would be concerned that the AI is beating to inbuilt AI due to what could be termed 'save-game cheating'. And there's still the issue of whether it is overspecialised to just exploit the inbuilt AI in some way.