Dominion Strategy Forum

Please login or register.

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

Author Topic: New online version of Dominion (with neural-network based AI) from Temple Gates  (Read 9246 times)

0 Members and 2 Guests are viewing this topic.

dsplaisted

  • Herbalist
  • **
  • Offline Offline
  • Posts: 5
  • Respect: +11
    • View Profile
+7

Hi folks,

Has anyone heard about this new version of Dominion that's supposed to be coming out?  And what it means for https://dominion.games?

https://www.polygon.com/22440924/dominion-app-neural-network-ai-release-date-price
https://store.steampowered.com/app/1131620/Dominion/

Thanks,
Daniel
Logged

Titandrake

  • Mountebank
  • *****
  • Offline Offline
  • Posts: 2210
  • Respect: +2854
    • View Profile
0

Someone who knows more should confirm, but my understanding is that dominion.games is supposed to co-exist with the new version. ShuffleIT promised some single-player and mobile parts that never came out. Rio Grande Games reached out to another company to make a mobile app and I believe this is that mobile app, except it'll have some cross-platform support as well according to the article.

I suspect that the competitive community won't move off ShuffleIT that quickly, but more casual players will find the app to be a better experience.

Edit: Given that ShuffleIT still lets you buy subscriptions 2 years out into the future, I assume the site is still legally allowed to operate at least that long.
« Last Edit: May 18, 2021, 02:17:56 pm by Titandrake »
Logged
I have a blog! It's called Sorta Insightful. Check it out?

blueblimp

  • Margrave
  • *****
  • Offline Offline
  • Posts: 2849
  • Respect: +1559
    • View Profile
+3

Thank goodness that the mobile license is finally in the hands of a developer that has a track record of mobile boardgame adaptations. (I haven’t tried this company’s apps, but they get good ratings in the app store.) It’s crazy to me that in the 8+ years since Iso was shuttered, this is the first time that has happened.

Adapting Dominion to a small screen sounds insanely hard, so I wonder how that’ll work. I’m also curious about the AI, which is also a hard problem.
Logged

blueblimp

  • Margrave
  • *****
  • Offline Offline
  • Posts: 2849
  • Respect: +1559
    • View Profile
Logged

LastFootnote

  • Adventurer
  • ******
  • Offline Offline
  • Posts: 7495
  • Shuffle iT Username: LastFootnote
  • Respect: +10721
    • View Profile
+2

I’m also curious about the AI, which is also a hard problem.

I'm also intensely curious about the AI. There's a little about it in the Polygon article linked in the OP, but honestly I hope there's something lost in translation there. The claims they're making about the AI seem almost laughably naive. And even if they're right and it's an amazing Dominion AI that can beat even seasoned players in most games, that sounds pretty bad for casual players! Who wants to play against a computer that beats you 90% of the time? Some people will, sure, but I bet that'll get old fast for a lot of players.

EDIT: But I mean, check this out, from the linked article:

Quote
"Instead of having a concept of a card in our neural network, we have a concept of each of the components of a card,” said Duringer. “The card has a cost. The card has a victory point value. The card might give you an extra buy or an extra action. If we can just understand these components, we don’t need to know the ‘value’ of a card. We can surmise it on the spot by understanding those components."

It's simultaneously hilarious and sad. The quote is from the Temple Gates CEO, so I can only hope this information is either misleading or outdated.
« Last Edit: May 18, 2021, 03:42:43 pm by LastFootnote »
Logged

J Reggie

  • Jester
  • *****
  • Offline Offline
  • Posts: 842
  • Shuffle iT Username: J Reggie
  • Respect: +1487
    • View Profile
    • Jeff Rosenthal Music
+4

Looks like it might be time to revisit my old thread!

And even if they're right and it's an amazing Dominion AI that can beat even seasoned players in most games, that sounds pretty bad for casual players! Who wants to play against a computer that beats you 90% of the time? Some people will, sure, but I bet that'll get old fast for a lot of players.

I think most neural networks have some depth setting you can turn down to make it perform worse, although it may not get worse in the same way a human would.
« Last Edit: May 18, 2021, 03:51:01 pm by J Reggie »
Logged

blueblimp

  • Margrave
  • *****
  • Offline Offline
  • Posts: 2849
  • Respect: +1559
    • View Profile
0

And even if they're right and it's an amazing Dominion AI that can beat even seasoned players in most games, that sounds pretty bad for casual players!
They're not claiming that the AI beats expert players. The quoted player self-describes his experience as
Quote
I’ve been playing Dominion casually, in person, for years. I’ve probably played 100, 200, maybe more games.
So we're talking about an AI that can beat a player who has played 200 casual games of Dominion. Depending on the sets included, even BM+X might be able to do that.

Quote
Quote
"Instead of having a concept of a card in our neural network, we have a concept of each of the components of a card,” said Duringer. “The card has a cost. The card has a victory point value. The card might give you an extra buy or an extra action. If we can just understand these components, we don’t need to know the ‘value’ of a card. We can surmise it on the spot by understanding those components."

It's simultaneously hilarious and sad. The quote is from the Temple Gates CEO, so I can only hope this information is either misleading or outdated.
This bit struck me as odd too. Breaking down cards into components like this might work alright for vanilla cards, but most cards have unique text.

Plus it seemingly solves a non-problem. The AI doesn't need to adapt on-the-fly to new cards, because they can just re-run the AI training whenever they add new cards.
Logged

LastFootnote

  • Adventurer
  • ******
  • Offline Offline
  • Posts: 7495
  • Shuffle iT Username: LastFootnote
  • Respect: +10721
    • View Profile
+1

This bit struck me as odd too. Breaking down cards into components like this might work alright for vanilla cards, but most cards have unique text.

It even falls apart for vanilla cards. [+1 Cards and +2 Actions] costs $3. [+2 Cards and +1 Action] costs $5. So [+3 Cards] should cost even more, right? Nope, it costs $4. Either Smithy is an extremely strong card or Village and Lab are awful.

Plus it seemingly solves a non-problem. The AI doesn't need to adapt on-the-fly to new cards, because they can just re-run the AI training whenever they add new cards.

An excellent point. But taking it even further, what does knowing the "value" of a card get you? The value of a card changes not only from game to game, but over the course of a single game. Province is an awful card very early in the game, but a great one at the end. One Witch is great. 9 Witches is not great.
« Last Edit: May 18, 2021, 04:05:04 pm by LastFootnote »
Logged

blueblimp

  • Margrave
  • *****
  • Offline Offline
  • Posts: 2849
  • Respect: +1559
    • View Profile
0

Looks like it might be time to revisit my old thread!
I didn't know about that discussion. Here are other previous discussions:
Logged

spineflu

  • Saboteur
  • *****
  • Offline Offline
  • Posts: 1365
  • Shuffle iT Username: spineflu
  • Head Empty, Heart Worms, Can't Lose
  • Respect: +1349
    • View Profile
    • my instagram, where i paint things
+1

With the AI running on the client end, with some of those clients being a phone, I'd be surprised if they do a full-on ANN with it; To me, sounds probably something smaller like a decision tree output from a server-based ANN.

I’m also curious about the AI, which is also a hard problem.

I'm also intensely curious about the AI. There's a little about it in the Polygon article linked in the OP, but honestly I hope there's something lost in translation there. The claims they're making about the AI seem almost laughably naive. And even if they're right and it's an amazing Dominion AI that can beat even seasoned players in most games, that sounds pretty bad for casual players! Who wants to play against a computer that beats you 90% of the time? Some people will, sure, but I bet that'll get old fast for a lot of players.

EDIT: But I mean, check this out, from the linked article:

Quote
"Instead of having a concept of a card in our neural network, we have a concept of each of the components of a card,” said Duringer. “The card has a cost. The card has a victory point value. The card might give you an extra buy or an extra action. If we can just understand these components, we don’t need to know the ‘value’ of a card. We can surmise it on the spot by understanding those components."

It's simultaneously hilarious and sad. The quote is from the Temple Gates CEO, so I can only hope this information is either misleading or outdated.

I'm curious how much understanding you have of programming ANNs - this is the exact sort of component analysis you'd need to do to make an ANN that "understands" dominion. He's not talking "this card costs $X, it must be better than the card that costs $Y", he's talking summing the available information into a single decision process, and that includes things like opportunity cost, cost, effects, etc. Even unique features - creating sparse dimensions in the ANN training set, but whatever - can be determined to be separable features.

anyhow you can sign up for the beta here, this was posted in the dominion fb group this morning
« Last Edit: May 18, 2021, 04:21:35 pm by spineflu »
Logged

LastFootnote

  • Adventurer
  • ******
  • Offline Offline
  • Posts: 7495
  • Shuffle iT Username: LastFootnote
  • Respect: +10721
    • View Profile
0

I'm curious how much understanding you have of programming ANNs - this is the exact sort of component analysis you'd need to do to make an ANN that "understands" dominion. He's not talking "this card costs $X, it must be better than the card that costs $Y", he's talking summing the available information into a single decision process, and that includes things like opportunity cost, cost, effects, etc. Even unique features - creating sparse dimensions in the ANN training set, but whatever - can be determined to be separable features.

Admittedly I'm not an expert, but I'm not sure I understand you? A high cost by (usually) makes a card weaker, not stronger. I expect a neural network to understand that much. The way this was described, the AI can surmise the "value" (whatever that means) of a card based on its individual components. But generally those components all interact with each other. They are more (or sometimes less) than the sum of their parts. So either the CEO's explanation was totally wrong, or there's something else I'm missing here. Either way, blueblimp's point still stands. Why would you want to be able to figure out the "value" of a new card from its parts when you could run thousands of simulations instead?
Logged

spineflu

  • Saboteur
  • *****
  • Offline Offline
  • Posts: 1365
  • Shuffle iT Username: spineflu
  • Head Empty, Heart Worms, Can't Lose
  • Respect: +1349
    • View Profile
    • my instagram, where i paint things
+1

I'm curious how much understanding you have of programming ANNs - this is the exact sort of component analysis you'd need to do to make an ANN that "understands" dominion. He's not talking "this card costs $X, it must be better than the card that costs $Y", he's talking summing the available information into a single decision process, and that includes things like opportunity cost, cost, effects, etc. Even unique features - creating sparse dimensions in the ANN training set, but whatever - can be determined to be separable features.

Admittedly I'm not an expert, but I'm not sure I understand you? A high cost by (usually) makes a card weaker, not stronger. I expect a neural network to understand that much. The way this was described, the AI can surmise the "value" (whatever that means) of a card based on its individual components. But generally those components all interact with each other. They are more (or sometimes less) than the sum of their parts. So either the CEO's explanation was totally wrong, or there's something else I'm missing here. Either way, blueblimp's point still stands. Why would you want to be able to figure out the "value" of a new card from its parts when you could run thousands of simulations instead?

That's sort of the "why use AI" question in a nutshell, isn't it?

It doesn't make a decision based on a single of those components - instead, it looks at how those components a card possesses would stack up to one another, and use more of a bigger picture - how's that card do in the context of the kingdom?
Now because of the combinatorics of kingdoms - sixty six sextillion, i think the polygon article said? - you don't want to run thousands of simulations on each of those combinations to just output a table. you want to make it more bite-sized so a phone can handle it. And that's where component breakdown is better. Is moat a better option than faithful hound? is relying on thrones to be villages viable in this kingdom? can I rely on something cheap like hamlet for +buys this game, or should I blow my first $5 hand on a market? that's the sort of questions you can get answers to without having to get into the full combinatorics.

Like I'm not trying to be the full "AI is going to save us and we're going to go to mars because AI-as-ideology" type internet reply guy with this, but this is - a game with complex components that can be broken down into overlapping features - an exact perfect case for the use of an ANN.
« Last Edit: May 18, 2021, 04:53:24 pm by spineflu »
Logged

LastFootnote

  • Adventurer
  • ******
  • Offline Offline
  • Posts: 7495
  • Shuffle iT Username: LastFootnote
  • Respect: +10721
    • View Profile
+1

Well I hope you're right. Though I also hope there will be an option for a less-ruthless AI for starting players who don't want to lose repeatedly while learning the game.
Logged

spineflu

  • Saboteur
  • *****
  • Offline Offline
  • Posts: 1365
  • Shuffle iT Username: spineflu
  • Head Empty, Heart Worms, Can't Lose
  • Respect: +1349
    • View Profile
    • my instagram, where i paint things
+1

Well I hope you're right. Though I also hope there will be an option for a less-ruthless AI for starting players who don't want to lose repeatedly while learning the game.
ha, yeah, hopefully they'll port in Lord Rattington.
Logged

Kirian

  • Adventurer
  • ******
  • Offline Offline
  • Posts: 7096
  • Shuffle iT Username: Kirian
  • An Unbalanced Equation
  • Respect: +9411
    • View Profile
0

Quote
I’ve been playing Dominion casually, in person, for years. I’ve probably played 100, 200, maybe more games.
So we're talking about an AI that can beat a player who has played 200 casual games of Dominion. Depending on the sets included, even BM+X might be able to do that.

I came very close to actually laughing out loud at that paragraph. Like, my dude, have you tested the AI against players who've played 1000 times more games than that?

That having been said, you also don't want the AI to destroy everyone every game, but I'm certain they can tune down the settings.

"200, maybe more, games." Hilarious.
Logged
Kirian's Law of f.DS jokes:  Any sufficiently unexplained joke is indistinguishable from serious conversation.

Titandrake

  • Mountebank
  • *****
  • Offline Offline
  • Posts: 2210
  • Respect: +2854
    • View Profile
+4

The devs (more specifically keldon) are in the Dominion Discord and have talked about the AI a bit there. It's all spread out (because Discord is bad at concentrating discussion) but here is the summary:

* It's inspired by AlphaZero
* There's one NN that predicts value and another policy head that picks actions.
* NN architecture is Transformer-based, right now only takes in current game state (have tried conditioning on previous buys but did not see noticeable improvement)
* At playing time it does a search rollout. Generally with their computation budget the search tree gets to a few turns out for the deepest paths (i.e. on first turn the search tree usually rolls out to the end of 1st reshuffle).
* It's all self play, no human data.
* They started with Base-only and have been slowly introducing more expansions over time (restarting bot training from scratch every time due to needing to change NN layout + not wanting the bot to get stuck on strategies that only worked in the simplified game). This might be what the CEO is talking about? That the same method is able to handle new cards as long as you let it train against itself for long enough.
* The bot is generally very good at big money + a few action card strategies, which is how I'd describe the MF bots. (But worth noting this indicates the self-play learned BM + X on its own which is a good sign.)
* For difficulty levels, they'll probably tune the amount of search it does.
Logged
I have a blog! It's called Sorta Insightful. Check it out?

Kirian

  • Adventurer
  • ******
  • Offline Offline
  • Posts: 7096
  • Shuffle iT Username: Kirian
  • An Unbalanced Equation
  • Respect: +9411
    • View Profile
0

Also... I could swear that somewhere on this very forum, DXV said that if you try to judge values of cards based on their pieces, you're going to fail, since the game was designed at least partly to not let that happen. Anyone have a link? (Likely not, forum searches are really difficult.)
Logged
Kirian's Law of f.DS jokes:  Any sufficiently unexplained joke is indistinguishable from serious conversation.

LastFootnote

  • Adventurer
  • ******
  • Offline Offline
  • Posts: 7495
  • Shuffle iT Username: LastFootnote
  • Respect: +10721
    • View Profile
0

So the AI cheats? It plays a few turns ahead and makes decisions based on that?
Logged

Titandrake

  • Mountebank
  • *****
  • Offline Offline
  • Posts: 2210
  • Respect: +2854
    • View Profile
+2

So the AI cheats? It plays a few turns ahead and makes decisions based on that?

Not really? I mean, it doesn't know exactly what shuffle it's going to get. It just simulates a bunch of potential shuffles based on what happens if it makes certain moves and then picks ones that made it win more in those simulations.
Logged
I have a blog! It's called Sorta Insightful. Check it out?

spineflu

  • Saboteur
  • *****
  • Offline Offline
  • Posts: 1365
  • Shuffle iT Username: spineflu
  • Head Empty, Heart Worms, Can't Lose
  • Respect: +1349
    • View Profile
    • my instagram, where i paint things
0

Also... I could swear that somewhere on this very forum, DXV said that if you try to judge values of cards based on their pieces, you're going to fail, since the game was designed at least partly to not let that happen. Anyone have a link? (Likely not, forum searches are really difficult.)

I'm pretty sure thats in the guide to pricing fan cards, not a principle for judging cards.
Logged

LastFootnote

  • Adventurer
  • ******
  • Offline Offline
  • Posts: 7495
  • Shuffle iT Username: LastFootnote
  • Respect: +10721
    • View Profile
0

So the AI cheats? It plays a few turns ahead and makes decisions based on that?

Not really? I mean, it doesn't know exactly what shuffle it's going to get. It just simulates a bunch of potential shuffles based on what happens if it makes certain moves and then picks ones that made it win more in those simulations.

OK, that's fine for shuffles. But say it plays Wishing Well, or decides whether it's going to play Wishing Well. What then?
Logged

seanahan

  • Pawn
  • **
  • Offline Offline
  • Posts: 2
  • Shuffle iT Username: seanahan
  • Respect: +2
    • View Profile
0

Presumably it just guesses the most likely card in its deck, the same way a person does (if they are properly deck tracking).
Logged

Wizard_Amul

  • Conspirator
  • ****
  • Offline Offline
  • Posts: 205
  • Respect: +216
    • View Profile
0

Presumably it just guesses the most likely card in its deck, the same way a person does (if they are properly deck tracking).

That's not always the optimal strategy, though--if finding a less common card helps your turn a lot more, then it may be worth it to guess that card instead of the most common card. E.g., you really need a village but the most common card is copper. I don't know how the new AI would handle it, but if it looked to the end of the turn or to the end of a couple turns, it could create a probability model out of different scenarios. I just wonder how slow/fast the AI will be--imagine if the remaining deck has multiple Wishing Wells, but Wishing Well is not the most common card remaining in the deck.
Logged

Jack Rudd

  • Saboteur
  • *****
  • Offline Offline
  • Posts: 1323
  • Shuffle iT Username: Jack Rudd
  • Respect: +1379
    • View Profile
0

There's also the scenario where you guess Estate because you already have what you need for the buy you intend to make, and so you're trying to find the least helpful card for your next turn and draw that.
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!'

ConMan

  • Saboteur
  • *****
  • Offline Offline
  • Posts: 1400
  • Respect: +1705
    • View Profile
+1

Assuming it's been made "right" (for some value of right), it would presumably build a prediction of what card it could be wishing for, how likely that card is to show up, and what it can do after that. So it wouldn't necessarily name the most likely card but instead the one that gives the best overall results when it wishes for it (which might well be balancing the pros and cons of "what happens if I wish right" and "what happens if I wish wrong" just like a good human player would do).
Logged
Pages: [1] 2  All
 

Page created in 0.054 seconds with 21 queries.