Print

Please login or register.
1 Hour 1 Day 1 Week 1 Month Forever
Login with username, password and session length

[ConMan]

Here's how I would use Councilroom:
<snip>

You could do something vaguely similar to that using a neural network approach - have one input for each Kingdom card, and for a training set of games like you suggest as a set of high-level versus games, define the "interesting" flag to output as being, say, 0 when there is a high similarity between the cards each player buys, and 1 when there is a disparity. Once the network has been trained, it should be able to give you an "interestingness" score between 0 and 1 for any given Kingdom, which you can use in a variety of ways - e.g. generate 10 Kingdoms at a time, but play with the most interesting one.

[jotheonah]

When I play IRL with my friends, we draft, basically.  So say there are 4 people playing. We deal out 8 cards to everyone. Each person chooses one card to include and one to veto. Then they pass one to the left. Do the same. Then for the last two cards, deal two at random from the remaining 12 nonvetoed cards.  Or something along those lines.  People get to pick a few cards they want from a limited set, a few cards are randomly chosen, but those cards are at least guaranteed to be cards no one really hates.

[popsofctown]

Here's how I would use Councilroom:
<snip>

You could do something vaguely similar to that using a neural network approach - have one input for each Kingdom card, and for a training set of games like you suggest as a set of high-level versus games, define the "interesting" flag to output as being, say, 0 when there is a high similarity between the cards each player buys, and 1 when there is a disparity. Once the network has been trained, it should be able to give you an "interestingness" score between 0 and 1 for any given Kingdom, which you can use in a variety of ways - e.g. generate 10 Kingdoms at a time, but play with the most interesting one.

Doesn't that take a complex algorithm to identify cards and pairs of cards that cause any particular kingdom to be interesting?

I'm not sure I'm familiar with neural network approach.

[Jimmmmm]

Another idea would be to pick some number less than 10 of random cards then fill in the blanks with whatever you think would be the best fit. I played a  2-player Base/Seaside/Dark Ages game where we decided to have 6 random DA cards and choose a Base and a Seaside card each. The two cards my opponent chose were Market and Tactician, and the only village was Bandit Village and there was no decent trashing, so I added Moneylender and Bazaar to enable a double-Tactician deck.
I then proceeded to demolish his deck with Knights (eventually playing 4 every turn), leaving him with 9 cards in his deck at the end of the game, but that's not really the point...

[ConMan]

Here's how I would use Councilroom:
<snip>

You could do something vaguely similar to that using a neural network approach - have one input for each Kingdom card, and for a training set of games like you suggest as a set of high-level versus games, define the "interesting" flag to output as being, say, 0 when there is a high similarity between the cards each player buys, and 1 when there is a disparity. Once the network has been trained, it should be able to give you an "interestingness" score between 0 and 1 for any given Kingdom, which you can use in a variety of ways - e.g. generate 10 Kingdoms at a time, but play with the most interesting one.

Doesn't that take a complex algorithm to identify cards and pairs of cards that cause any particular kingdom to be interesting?

I'm not sure I'm familiar with neural network approach.

The neural network essentially *is* a complex algorithm, albeit one that adjusts itself based on the input data. Very roughly, it's a way of creating a system that approximates a function (usually either a real-valued one, or a probability, or a categorical identifier) of arbitrary complexity based on a set of "training data". So for this kind of situation, we'd say that the input is a set of binary flags identifying which cards are in the Kingdom, and the "interestingness" score would effectively be the "probability that the Kingdom is interesting", where "interesting" is defined as per however the training data is coded. If we use, say, a three-layer network, then we're allowing for some reasonably complex (i.e. non-linear, up to third order) interactions between the cards which should identify some neat combinations.

[T-hawk]

Also, there's a problem with selecting against high gain% cards.  You can't just keep them out of the pool, so there will probably be at least 1 on the table.  Since the other cards will generally be of lower power level, the high gain% card is even more likely to take over the game.

If you have Gain% data and want to build a kingdom chooser around it, the answer is to select cards of the SAME gain% strata.  That's how you avoid the one-dominant-card syndrome.  Get a kingdom of cards all with similar gain% and you'll create real choices.  At the upper end, you might see KC, Bridge, Mountebank, Fishing Village, Chapel all together, and then you do have real choices even here on _which_ of the strong cards to get first and build around.  Or a set centered around lower gain% could have stuff like Apprentice, Smugglers, Workshop, Jester, Cutpurse together... cards that get overshadowed by the bullies, but could shine when playing in their own weight class.

Of course there's a feedback loop of gain% data, such that the big cards will decrease in gain% over time as they compete against other strong cards, and middling cards will pick up some more attention.  So we narrow the spread in gain% such that all cards get picked similarly often?  To that I say Mission Accomplished.

I'm very much not sold on the idea of picking always at least one attack, one village, one trasher, one alt-VP, etc.  Dominion has a richer space of possible games when sometimes you can build unmolested to your heart's content, or terminal slots are precious and must be carefully chosen, or you must deal with the starting coppers/estates, and so on.