![[personal profile]](https://www.dreamwidth.org/img/silk/identity/user.png){kind=small}
(Oops, it seems that my reduction might not be so safe after all. I'll say why in the main text, but oh well!)
-
In my previous post, I mentioned a way of solving semi-realtime games by the use of minimax, by just splitting the timeline into a series of "virtual turns", one for each instant of the clock, where each player may pass or do something. I then said that this would be horribly expensive because each minimum time instant would be a move and the game tree would get really deep really fast.
Here's an idea regarding how one may solve it. It may hold in unorthodox chess, and if it does, in realtime games in general.
( Read more... )
So there you are. Perhaps the less expensive AI is possible. Perhaps it wouldn't be smart enough. I don't know... I can argue both in favor and against.
But the revealing insight, really, was the idea that real-time games could be coerced to be turn-based by adding lots of virtual turns. In a real game, this might even produce desirable side effects: if each virtual turn is, say, one second, that means no player can go faster than 60 APM. No more silly AIs that only win because they're inhuman at micromanagement.
(The stock Starcraft II AI can reach 2000 APM. The human record is somewhere around 500 for longer times, 800 for instants. At least in my opinion, an AI is much more interesting if it can play better than humans while otherwise sharing the handicaps - reaction time, etc., of humans.)
And after writing all of the above, I found that my minimax-generalization idea had already occurred to others. If you want a rigorous paper, go there. It investigates the consequences of the AI having to actually think while time passes, too.
Finally, again, I really do wonder what unorthodox chess would play like. I could see lots of time strategies: the mental time-on-target I mentioned above, the tension between waiting (so as not to reveal intelligence) and moving (so as not to lose on time)...
-
In my previous post, I mentioned a way of solving semi-realtime games by the use of minimax, by just splitting the timeline into a series of "virtual turns", one for each instant of the clock, where each player may pass or do something. I then said that this would be horribly expensive because each minimum time instant would be a move and the game tree would get really deep really fast.
Here's an idea regarding how one may solve it. It may hold in unorthodox chess, and if it does, in realtime games in general.
( Read more... )
So there you are. Perhaps the less expensive AI is possible. Perhaps it wouldn't be smart enough. I don't know... I can argue both in favor and against.
But the revealing insight, really, was the idea that real-time games could be coerced to be turn-based by adding lots of virtual turns. In a real game, this might even produce desirable side effects: if each virtual turn is, say, one second, that means no player can go faster than 60 APM. No more silly AIs that only win because they're inhuman at micromanagement.
(The stock Starcraft II AI can reach 2000 APM. The human record is somewhere around 500 for longer times, 800 for instants. At least in my opinion, an AI is much more interesting if it can play better than humans while otherwise sharing the handicaps - reaction time, etc., of humans.)
And after writing all of the above, I found that my minimax-generalization idea had already occurred to others. If you want a rigorous paper, go there. It investigates the consequences of the AI having to actually think while time passes, too.
Finally, again, I really do wonder what unorthodox chess would play like. I could see lots of time strategies: the mental time-on-target I mentioned above, the tension between waiting (so as not to reveal intelligence) and moving (so as not to lose on time)...