151
Simulation / Another approach: metric on games
« on: September 13, 2011, 10:50:40 pm »
New to the forum, but I was wondering if anyone had explored the following possibility in addition to simulations.
We have huge amounts of played games on the council room server. I don't know the exact amount, but I would assume a few millions.
Now there are many many more combinations of kingdoms than that. That being said, many kingdoms look alike: say, you group together most of the villages into a class, or disregard weak cards in a pool where a stronger strategy clearly dominates.
So, if we could manage to group games together by how close they are, it would probably be possible to "simulate" the results of a game just by hashing average scores statistically from previous games that looked similar enough.
There are clearly two problems with this approach, however. First, the grouping method must be both convincing and lead to a sufficient amounts of games to allow for useful statistical evaluations.
One possibility is to just define a metric, a distance, between two games. Then, given a set of kingdom cards (possibly never played before), you could then search for the, say, 1000 closest games given that metric and use those for your statistics and sample games.
Now defining a metric between games seems a daunting task of its own, given that a lot of (somewhat arbitrary) factors could go into it. Fine-tuning the metric could take a fair bit of time, though a good human player should be able to spot if the results "make sense". Lastly, the ideal case would be to have an algorithm that could search for "the closest games" in real time.
I have a rather simplistic metric in mind that would fulfill some of those criteria, though I would first like to hear reactions and propositions before going into any more details.
The code I have is a very crude R script, but it runs nearly in real time for a small database of 100k games. The results are somewhat convincing but the metric clearly has to be tweaked.
We have huge amounts of played games on the council room server. I don't know the exact amount, but I would assume a few millions.
Now there are many many more combinations of kingdoms than that. That being said, many kingdoms look alike: say, you group together most of the villages into a class, or disregard weak cards in a pool where a stronger strategy clearly dominates.
So, if we could manage to group games together by how close they are, it would probably be possible to "simulate" the results of a game just by hashing average scores statistically from previous games that looked similar enough.
There are clearly two problems with this approach, however. First, the grouping method must be both convincing and lead to a sufficient amounts of games to allow for useful statistical evaluations.
One possibility is to just define a metric, a distance, between two games. Then, given a set of kingdom cards (possibly never played before), you could then search for the, say, 1000 closest games given that metric and use those for your statistics and sample games.
Now defining a metric between games seems a daunting task of its own, given that a lot of (somewhat arbitrary) factors could go into it. Fine-tuning the metric could take a fair bit of time, though a good human player should be able to spot if the results "make sense". Lastly, the ideal case would be to have an algorithm that could search for "the closest games" in real time.
I have a rather simplistic metric in mind that would fulfill some of those criteria, though I would first like to hear reactions and propositions before going into any more details.
The code I have is a very crude R script, but it runs nearly in real time for a small database of 100k games. The results are somewhat convincing but the metric clearly has to be tweaked.