Dominion Strategy Forum

Please login or register.

Login with username, password and session length
Pages: 1 [2] 3 4  All

Author Topic: Letting a computer design Dominion cards: recurrent neural network edition  (Read 4938 times)

0 Members and 1 Guest are viewing this topic.

GendoIkari

  • Adventurer
  • ******
  • Online Online
  • Posts: 7131
  • Respect: +7962
    • View Profile
Re: Letting a computer design Dominion cards: recurrent neural network edition
« Reply #25 on: February 01, 2018, 04:41:13 pm »
0

So my very limited knowledge of neural networks tells me that each generation has some sort of criteria for determining if it's good or not. Like, MarI/O looks at how far horizontally it reached; and the farther, the better.

What sort of criteria is this using to determine how "good" a particular random card is?

To my understanding, what it does is it uses 90% of the data set as the learning set and 10% as the validation set. Then it tries to predict the values in the validation set based on the learning set, and measures the inconsistency between the predicted values and the actual values and tries to minimize that. When I'm having it generate these cards, it does it by trying to predict things from nothingness, although I could give it an initial string and have it try to predict how it continues.

It seems like after enough generations that would just end up with it creating only exact copies of existing cards; without any variance or differences at all.
Logged
Check out my F.DS extension for Chrome! Card links; Dominion icons, and maybe more! http://forum.dominionstrategy.com/index.php?topic=13363.0

Thread for Firefox version:
http://forum.dominionstrategy.com/index.php?topic=16305.0

Cuzz

  • Minion
  • *****
  • Offline Offline
  • Posts: 515
  • Shuffle iT Username: Cuzz
  • Respect: +841
    • View Profile
Re: Letting a computer design Dominion cards: recurrent neural network edition
« Reply #26 on: February 01, 2018, 04:44:48 pm »
0

Echip actually sounds like a card that an android would come up with for fun. 
Logged

King Leon

  • Conspirator
  • ****
  • Offline Offline
  • Posts: 214
  • Respect: +194
    • View Profile
Re: Letting a computer design Dominion cards: recurrent neural network edition
« Reply #27 on: February 01, 2018, 04:48:48 pm »
+6

I also tried my luck. I am still training the machine. Some first results:

Ruined Square
Types: Event
Cost: $4
Once per turn: Turn your turns, but not less than $0.

Farming
Types: Action
Cost: $0*
Gain a Spoils from the top of your deck.
Setup: Put a token on that pile.

Dame Gains
Types: Action - Attack - Knight
Cost: $5
While this is in play, you may put it on top.
Reveal the top card of your deck; or gain a Duchy, putting it on top of your next turn: +1 Coin per differently named Treasure Map from your hand that is not a Treasure card from your hand.
Each other player reveals the top 2 cards of their deck, trashes one of them costing from $3 to $6, and discards the rest. If a Knight is trashed by this, trash this.

Spice Traders
Types: Action
Cost: $5
+4 Cards
When you buy or play this, each other player gains a Province card from your hand.
« Last Edit: February 01, 2018, 04:57:19 pm by King Leon »
Logged

Awaclus

  • Adventurer
  • ******
  • Offline Offline
  • Posts: 10765
  • Shuffle iT Username: Awaclus
  • (。 ω 。`)
  • Respect: +11331
    • View Profile
Re: Letting a computer design Dominion cards: recurrent neural network edition
« Reply #28 on: February 01, 2018, 04:53:07 pm »
0

So my very limited knowledge of neural networks tells me that each generation has some sort of criteria for determining if it's good or not. Like, MarI/O looks at how far horizontally it reached; and the farther, the better.

What sort of criteria is this using to determine how "good" a particular random card is?

To my understanding, what it does is it uses 90% of the data set as the learning set and 10% as the validation set. Then it tries to predict the values in the validation set based on the learning set, and measures the inconsistency between the predicted values and the actual values and tries to minimize that. When I'm having it generate these cards, it does it by trying to predict things from nothingness, although I could give it an initial string and have it try to predict how it continues.

It seems like after enough generations that would just end up with it creating only exact copies of existing cards; without any variance or differences at all.

True.
Logged
Bomb, Cannon, and many of the Gunpowder cards can strongly effect gameplay, particularly in a destructive way

The Twitch channel where I stream DominionThe YouTube channel where I make musicDownload my band's albums for free

Erick648

  • Bishop
  • ****
  • Offline Offline
  • Posts: 123
  • Respect: +392
    • View Profile
Re: Letting a computer design Dominion cards: recurrent neural network edition
« Reply #29 on: February 01, 2018, 04:55:55 pm »
+1

Dame Gains
Types: Action - Attack - Knight
Cost: $5
While this is in play, you may put it on top.
Reveal the top card of your deck; or gain a Duchy, putting it on top of your next turn: +1 Coin per differently named Treasure Map from your hand that is not a Treasure card from your hand.
Each other player reveals the top 2 cards of their deck, trashes one of them costing from $3 to $6, and discards the rest. If a Knight is trashed by this, trash this.
Funny that it came up with a slang term for an actual card (Dame Natalie).  And this one is a gainer, too.
Logged
Duplicate duplicates Duplicates duplicate Duplicates duplicate.

Rene Descartes taught me to believe in myself.

aku_chi

  • Minion
  • *****
  • Offline Offline
  • Posts: 556
  • Shuffle iT Username: aku chi
  • Respect: +1217
    • View Profile
Re: Letting a computer design Dominion cards: recurrent neural network edition
« Reply #30 on: February 01, 2018, 05:07:10 pm »
+4

Spice Traders
Types: Action
Cost: $5
+4 Cards
When you buy or play this, each other player gains a Province card from your hand.

This card actually has a reasonable power level.  It would have to be wordsmithed to something like: "When you buy or play this, reveal your hand.  Each other player gains a Province card if you revealed a Province.".  It makes the penalty slightly less painful, but you're pretty much never going to buy or play this with a Province in hand anyway.
Logged

King Leon

  • Conspirator
  • ****
  • Offline Offline
  • Posts: 214
  • Respect: +194
    • View Profile
Re: Letting a computer design Dominion cards: recurrent neural network edition
« Reply #31 on: February 01, 2018, 05:10:19 pm »
+10

I am starting to get actually weird cards.

Horse
Types: Action - Duration
Cost: <8>
Now and at the start of your next turn: Gain a Treasure.
While this is in play, when another player takes his -1 Coin token, he gains a Curse and trash this card.

Counting Market
Types: Action
Cost: $4
You may reveal an Action card from your hand. If you do, discard it or trash up to 4 cards from your hand.
+$2
+1 Buy

Ill-Gotten of Mystic
Types: Treasure
Cost: $2
Reveal your hand. -1 $ per differently named card revealed. If this is the first time you played a Fool's Gold this turn, this is worth $1, otherwise it's worth $4.


Edit: The last batch of interesting cards, before I am going to bed.

Maker
Types: Action
Cost: $4
+1 Card, +2 Actions
Discard a card.
Each player may reveal a Treasure. Discard the other cards. Play that Treasure.

Lost
Types: Action
Cost: $2
+$2
When you discard this from play, gain a Curse card per Embargo token on that pile.

Hunter
Types: Action - Attack
Cost: $4
You can't buy this if you have at least 2 Coins unspent, you may discard this. If you do, +1 Card per card in your hand. Gain a card costing up to $3 more than it.
+1 Card, +2 Actions
Each other player gains a Province. You may trash this card. If you do, gain a Duchy. Put that card into your hand.
You may trash a card from your hand other than a Rats (or reveal a hand of all Rats).


I am rejecting all cards whose names conflict with existing cards. And I got so many Markets and Mines ...
« Last Edit: February 01, 2018, 05:25:49 pm by King Leon »
Logged

crj

  • Jester
  • *****
  • Offline Offline
  • Posts: 926
  • Respect: +1031
    • View Profile
Re: Letting a computer design Dominion cards: recurrent neural network edition
« Reply #32 on: February 01, 2018, 08:34:09 pm »
0

+1 Coin per differently named Treasure Map from your hand that is not a Treasure card from your hand.
Although actually simpler than it sounds, that would produce so many rules questions!
Logged

sudgy

  • Cartographer
  • *****
  • Offline Offline
  • Posts: 3308
  • Shuffle iT Username: sudgy
  • It's pronounced "SOO-jee"
  • Respect: +2540
    • View Profile
Re: Letting a computer design Dominion cards: recurrent neural network edition
« Reply #33 on: February 01, 2018, 10:02:26 pm »
+7

Counting Market
Types: Action
Cost: $4
You may reveal an Action card from your hand. If you do, discard it or trash up to 4 cards from your hand.
+$2
+1 Buy

This is actually an interesting concept.  You can only trash if you have an action in your hand.
Logged
If you're wondering what my avatar is, watch this.

Check out my logic puzzle blog!

   Quote from: sudgy on June 31, 2011, 11:47:46 pm

Jack Rudd

  • Saboteur
  • *****
  • Offline Offline
  • Posts: 1100
  • Shuffle iT Username: Jack Rudd
  • Respect: +1026
    • View Profile
Re: Letting a computer design Dominion cards: recurrent neural network edition
« Reply #34 on: February 02, 2018, 06:11:04 am »
0

Counting Market is certainly workable as an idea, although "up to 4" is probably too strong. Trashing a card from your hand would probably be about right.
Logged
Centuries later, archaeologists discover the remains of your ancient civilization.

Evidence of thriving towns, Pottery, roads, and a centralized government amaze the startled scientists.

Finally, they come upon a stone tablet, which contains but one mysterious phrase!

'ISOTROPIC WILL RETURN!'

Awaclus

  • Adventurer
  • ******
  • Offline Offline
  • Posts: 10765
  • Shuffle iT Username: Awaclus
  • (。 ω 。`)
  • Respect: +11331
    • View Profile
Re: Letting a computer design Dominion cards: recurrent neural network edition
« Reply #35 on: February 02, 2018, 07:50:39 am »
+13

This is what I got this morning (in addition to a lot of nonsense):

Acbisur toker
Action
Gain a Treasure.

Small Mastle
Victory - Ruins
Worth 1

Dimaged Lar
Action - Attack
You may trash up to 5 cards from your hand.

Noniager
Action
Gain a Curse.

Heralobles
Action
+
Reveal this onto your deck.

Rocard
Landmark
When scoring, -5 per copy card in hand.

Missint
Action - Attack
Gain a Silver. Discard the rest.
(This is not in the Supply.)

Squires
Action
Now and at the start of your next turn: +3 Cards, take 2 Copper.

Dame Trader
Action - Hooter
+3 Cards.
+1 Action.
Name a card.
___________
When you discard this from play, you may trash hand.

Soother
Action
+1 Buy.
+

Contrabitarin
Action - Attack - Knight
+1 Buy.
Each other player draws a card from your hand.

Cangler
Action
+1 Card.
+1 Action.
Discard 2 cards.
_________
When you gain this, it.

Vagp
Event
Gain a card from their hand onto your deck.

I also got cards named "Bleach" and "Furry", respectively, as well as a card with the pretty interesting type "Action-Attack-Knitchooser" but they were otherwise illegible cards.
« Last Edit: February 02, 2018, 07:54:00 am by Awaclus »
Logged
Bomb, Cannon, and many of the Gunpowder cards can strongly effect gameplay, particularly in a destructive way

The Twitch channel where I stream DominionThe YouTube channel where I make musicDownload my band's albums for free

NoMoreFun

  • Saboteur
  • *****
  • Offline Offline
  • Posts: 1365
  • Respect: +953
    • View Profile
Re: Letting a computer design Dominion cards: recurrent neural network edition
« Reply #36 on: February 02, 2018, 08:47:52 am »
0

What would the neural network pop out if you fed it all the card images (including text)
Logged

Awaclus

  • Adventurer
  • ******
  • Offline Offline
  • Posts: 10765
  • Shuffle iT Username: Awaclus
  • (。 ω 。`)
  • Respect: +11331
    • View Profile
Re: Letting a computer design Dominion cards: recurrent neural network edition
« Reply #37 on: February 02, 2018, 09:35:37 am »
0

What would the neural network pop out if you fed it all the card images (including text)

I would have to look into that. The current implementation I have running only knows what to do with text.
Logged
Bomb, Cannon, and many of the Gunpowder cards can strongly effect gameplay, particularly in a destructive way

The Twitch channel where I stream DominionThe YouTube channel where I make musicDownload my band's albums for free

Chappy7

  • Explorer
  • *****
  • Offline Offline
  • Posts: 302
  • Shuffle iT Username: Chappy7
  • Respect: +375
    • View Profile
Re: Letting a computer design Dominion cards: recurrent neural network edition
« Reply #38 on: February 02, 2018, 10:52:34 am »
+2

Action-Hooter  :o  Hahaha
Logged

J Reggie

  • Minion
  • *****
  • Online Online
  • Posts: 680
  • Shuffle iT Username: J Reggie
  • Respect: +1068
    • View Profile
    • Jeff Rosenthal Music
Re: Letting a computer design Dominion cards: recurrent neural network edition
« Reply #39 on: February 02, 2018, 12:44:55 pm »
+3

Noniager too OP.

And that's quite the cost on Cangler.

Commodore Chuckles

  • Tactician
  • *****
  • Offline Offline
  • Posts: 446
  • Shuffle iT Username: Commodore Chuckles
  • Respect: +703
    • View Profile
Re: Letting a computer design Dominion cards: recurrent neural network edition
« Reply #40 on: February 02, 2018, 02:25:38 pm »
0

Action-Hooter  :o  Hahaha

That must be where Dame Trader works.
Logged

King Leon

  • Conspirator
  • ****
  • Offline Offline
  • Posts: 214
  • Respect: +194
    • View Profile
Re: Letting a computer design Dominion cards: recurrent neural network edition
« Reply #41 on: February 03, 2018, 06:37:40 am »
+3

Acbisur toker
Action
Gain a Treasure.

Small Mastle
Victory - Ruins
Worth 1

These are actually not bad. However, Small Mastle is too similar to Estate and in most cases better than Ruined Village, but still a nice Ruins. Maybe it should just give VP.
« Last Edit: February 03, 2018, 06:39:53 am by King Leon »
Logged

Awaclus

  • Adventurer
  • ******
  • Offline Offline
  • Posts: 10765
  • Shuffle iT Username: Awaclus
  • (。 ω 。`)
  • Respect: +11331
    • View Profile
Re: Letting a computer design Dominion cards: recurrent neural network edition
« Reply #42 on: February 03, 2018, 06:46:57 am »
+8

Now it managed to design a card which is actually fairly interesting and balanced (by coincidence, certainly not because it has learned how to do that now):

Fambl
Action
Draw until you have 6 cards in hand.
__________
While this is in play, cards cost less, but not less than .
Logged
Bomb, Cannon, and many of the Gunpowder cards can strongly effect gameplay, particularly in a destructive way

The Twitch channel where I stream DominionThe YouTube channel where I make musicDownload my band's albums for free

Awaclus

  • Adventurer
  • ******
  • Offline Offline
  • Posts: 10765
  • Shuffle iT Username: Awaclus
  • (。 ω 。`)
  • Respect: +11331
    • View Profile
Re: Letting a computer design Dominion cards: recurrent neural network edition
« Reply #43 on: February 03, 2018, 09:20:37 am »
0

These are actually not bad. However, Small Mastle is too similar to Estate and in most cases better than Ruined Village, but still a nice Ruins. Maybe it should just give VP.

Well, all the Ruins are better than Ruined Village, and usually it's not a trivial decision whether you want to trash an Estate or a Ruins that actually does something. It's not very exciting but it would be fine, I think.
Logged
Bomb, Cannon, and many of the Gunpowder cards can strongly effect gameplay, particularly in a destructive way

The Twitch channel where I stream DominionThe YouTube channel where I make musicDownload my band's albums for free

Schneau

  • Saboteur
  • *****
  • Offline Offline
  • Posts: 1170
  • Shuffle iT Username: Schneau
  • Respect: +1446
    • View Profile
    • Rainwave
Re: Letting a computer design Dominion cards: recurrent neural network edition
« Reply #44 on: February 03, 2018, 09:34:07 am »
+3

Squires
Action
Now and at the start of your next turn: +3 Cards, take 2 Copper.

Maybe it's just in comparison to these other cards, but this doesn't seem like a half bad idea, if you interpret the end to mean "gain 2 Coppers".
Logged

crj

  • Jester
  • *****
  • Offline Offline
  • Posts: 926
  • Respect: +1031
    • View Profile
Re: Letting a computer design Dominion cards: recurrent neural network edition
« Reply #45 on: February 03, 2018, 10:50:59 am »
0

I pondered that.

Then I realised, even if it flooded your deck with Copper, it's still a $3-cost giving +$3 this turn and next. Buying just those and Provinces might be hard to beat!
Logged

Philipp

  • Pearl Diver
  • **
  • Offline Offline
  • Posts: 11
  • Respect: +14
    • View Profile
Re: Letting a computer design Dominion cards: recurrent neural network edition
« Reply #46 on: February 03, 2018, 12:38:06 pm »
+2

Maybe another approach to machine learning created cards could be this: the AI understands the rules of Dominion and can play against itself, having been trained to play good. It then throws random newly designed rule cards into the mix, but making sure that rule card doesn't randomly already exist in Dominion, and then plays another million rounds with those against itself. It thus now incorporated the new card into its training data. Now we check if the new random card turns out to be bought a much higher or lesser amount of time than any other, i.e. if it is completely over or underpowered. If neither of those, it may be potentially a new interesting card. Or it may end up being boring because it's not a truly new type of rule ;)
Logged

crj

  • Jester
  • *****
  • Offline Offline
  • Posts: 926
  • Respect: +1031
    • View Profile
Re: Letting a computer design Dominion cards: recurrent neural network edition
« Reply #47 on: February 03, 2018, 01:17:50 pm »
+1

I don't think that would bring much to the table in practice.

Firstly, as we've discovered, even constructing meaningful cards is difficult. Secondly, as we've discovered, making a computer play Dominion well is difficult. It's not easy to be in a position to have a computer design cards and playtest them.

But more importantly, making cards that are legal and balanced is only part of the battle. Cards also have to be different, and interesting, and fun. Ideally, they're also thematic, and introduce entirely new game mechanics.

Without those considerations, we'd be flooded with new card ideas like "Blacksmith: Action $4. +3 Cards. Discard the top card of your deck." or "Boring Village: Action $3. +1 Card. +2 Actions. Each other player may gain a Curse." They're not existing cards but, I mean, really... (-8

There are a lot of other really disruptive skills computers are going to acquire long before one can take the rules PDF for the Dominion base set, churn for an hour and spit out something as fun, innovative and high quality as the Adventures expansion. Donald X. has rather more job security than doctors and airline pilots.
Logged

Philipp

  • Pearl Diver
  • **
  • Offline Offline
  • Posts: 11
  • Respect: +14
    • View Profile
Re: Letting a computer design Dominion cards: recurrent neural network edition
« Reply #48 on: February 03, 2018, 04:46:27 pm »
+2

Secondly, as we've discovered, making a computer play Dominion well is difficult.

That's what they said about Go too, then Google DeepMind came along ;)

In all seriousness though, self-training neural networks can be extremely effective in coming up with strategies. But as you say, and as I said in my comment also, this would not necessarily translate to interesting cards. That too will one day be possible -- just simulate a human brain digitally, and you have all the ingredients to have the neural network measure every card for how stimulating it is! -- but by that time, we're happy if our AI overlords still keep us around in the first place (if how humanity treats less intelligent species is any indicator, we better learn to be cute and cuddly, as surely we want to go the way of cats and dogs instead of pigs and cows!).
Logged

J Reggie

  • Minion
  • *****
  • Online Online
  • Posts: 680
  • Shuffle iT Username: J Reggie
  • Respect: +1068
    • View Profile
    • Jeff Rosenthal Music
Re: Letting a computer design Dominion cards: recurrent neural network edition
« Reply #49 on: February 03, 2018, 05:01:50 pm »
+3

We should petition the Google DeepMind team to make the bot for Shuffleit!
Pages: 1 [2] 3 4  All
 

Page created in 0.158 seconds with 20 queries.