I've been debating if I should write a basic engine to test 2 color decks for optimal creature/spell curve and colored land distribution (re: limited). I've thrown together some basic simulations w/ colorless spells to find the best curve based on what turn you want to curve out by... but I've realized to uncover quality results a 2 color build is necessary. (significant more programming)
one sim I'd like to run is 8 & 9 lands vs 8 & 8 lands + 1 CIT dual (or more), to find out how much the CIT effect hinders a curve over the long run of course (1000s of simulations).
I'd say go for it. It'd be a good tool to use when first building decks, before play testing ever begins. It would also allow you to learn some ground rules when building these deck that you can use when building sealed pools or in limited.
There is a lot be gained here, just simply ignoring that is stupid. While it's true that we already have stats and a base of Pro knowledge, gaining access to more info is always a good thing. You could plug in deck list, go to work, and come back too see what happened. I know I would use that quite a bit.
thanks for the encouragement. that's what i like to hear. i've stumbled across on player/writer who's done a similar test. but as i've tested he only simulated colorless spells over generic lands.
my first dilemma is how to approach checking lands (specifically duals) for proper mana to cast nth spell. i'm leaning towards having all colored permutations of mana available for that land state stored persistently .... or checking for available mana on the fly given nth spell's casting cost.
Playing simulations works best for decks like burn/aggro that have a very clear curve and just attack.
With pretty much no real choices in your game.
You can fairly well produce a deck this way, which will also work well.
As soon as theres some tricky moves involved or interaction with an opponent things get ugly quickly, as simulation for this is expensive and complicated, as you more or less build an AI that wants to play magic, which isnt trivial at all.
The basic idea of what you want can allready be accomplished with Math and an Excel Sheet, so you can optimise your numbers to make the best chance to archive your goal.
Math is your best friend in this regard, theres no need for simulations or an AI to play the game, as all we want is too optimize specific draws and make them as likely as possible.
However, this means we allready need to have at least an idea of what is optimal and a deck to begin with. This wont tell us to include cards we do not have on our radar and it will not point out specific interactions of cards and totally ignore the metagame (unless we use a goal like not losing to a wrath of god or graveyard hate etc.).
Overall, Excel Sheets are your best friend for most magic related as a player, to get a real idea of how your match % are and how even a single card will influence your performance against the field you expect (as putting a sideboard card in the maindeck might give you a real edge allready).
Decks like "storm" and burn can be fairly optimized to archive the quickest and most consistent kills.
I've been debating if I should write a basic engine to test 2 color decks for optimal creature/spell curve and colored land distribution (re: limited). I've thrown together some basic simulations w/ colorless spells to find the best curve based on what turn you want to curve out by... but I've realized to uncover quality results a 2 color build is necessary. (significant more programming)
one sim I'd like to run is 8 & 9 lands vs 8 & 8 lands + 1 CIT dual (or more), to find out how much the CIT effect hinders a curve over the long run of course (1000s of simulations).
anyone know of such a thing?
I've built two of them, one for Modern Knightfall and the other for Legacy Burn. They gave some very interesting results. I didn't do much with the mana using rainbow lands but there was still a lot of data in just having different land counts. With the Knightfall deck (Burn is still a WIP as I'm trying to include a few possible opponents so it's not a pure goldfish, so I don't have complete results yet... will have to wait for summer break) I found that much higher than generally accepted land counts lead to the quickest wins on average. My system would take max/min parameters for various cards, and then try every 60-62 card permutation of that deck at 10k-250k games per permutation (depending on how many permutations were possible) and then record the results.
Don't. Statistics and pro Magic players playtesting against every Tier I deck in the format does and will know better than any code you can write.
Actually, if you give a team of 20 pro's 2 weeks to focus on a single deck non stop, assuming 15 minutes per game and 14 hours/day put into testing you will only generate 7,840 games worth of data. With the automated goldfish approach, while you don't get matchup data you can generate literally millions of games worth of data in a single day. Also, believe it or not but Magic is an extremely high variance game and there is still a very noticeable effect from variance at even 100k games.
Actually, that little insight from variance made me reconsider several aspects of deck building and put far more value on the typical pro opinion of including a card just to have the option of hitting it, at the cost of some consistency.
Don't. Statistics and pro Magic players playtesting against every Tier I deck in the format does and will know better than any code you can write.
Actually, if you give a team of 20 pro's 2 weeks to focus on a single deck non stop, assuming 15 minutes per game and 14 hours/day put into testing you will only generate 7,840 games worth of data. With the automated goldfish approach, while you don't get matchup data you can generate literally millions of games worth of data in a single day. Also, believe it or not but Magic is an extremely high variance game and there is still a very noticeable effect from variance at even 100k games.
Actually, that little insight from variance made me reconsider several aspects of deck building and put far more value on the typical pro opinion of including a card just to have the option of hitting it, at the cost of some consistency.
The "quantity" of games doesnt matter as much as your thinking behind these.
You do not have to play 100+ matches against a deck to get any more information, if you get down to the fundamental important card interactions.
Most of the time a matchup will have something that is especially important to resolve, and you will win on that cards backpack alone. For example you play control, and you have to have a Languish to beat any aggro opponent that floods the board with tokens. But if you find that, you win fairly easily. That fact alone will have you mulligan for that card more aggressive and you do not play hands that you simply "know" wont do the trick in the matchup.
The problem with simulations that truly try every permutation is that its not real magic, you wouldnt play "any" hand, you mulligan them away. And your decision if a hand is a mulligan or not is greatly influenced by the matchup. However, thats only true for decks that have such a relevant specific card for matchups (allmost any controll deck will be this category) ; while you can much easier test decks that allmost dont care at all for the opponent, burn just hits for 20 damage, if you can do that by turn 4, you are pretty mcuh done with the deck, doesnt really matter what your opponent is doing, as you allmost cannot do anything to change that (but at the current level of burn decks, even they have specific cards that depend on the metagame, as a card like Eidolon of the Great Revel either do a ton or nothing, and cards like Skullcrack will give you a respons to lifegain, so the sequence of cards isnt as automatic, as you have to take your opponents actions into account to optimise your 20 damage goal).
Affinity, Infect and storm combo decks all do fairly well in simulations, as they have a very clear goal in mind. Even the Eldrazi deck was fairly good to test with simulations, as you want to optimise your turn ~5 kills at first ; and if you have these, you can start doing Simulations of your decks against each other to identify a "winner" of these top decks (but most of the time you can do a very well approximation by just a handfull of games with each match, as you only really want to see if a deck is overall more likely to win than another, the exact % isnt really relevant for that).
If your simulation just plays every hand no matter what, than adding MORE lands will usually increase consistency, as it will have more hands that actually have 2+ lands. However, if you properly mulligan hands, this wrong assumption will be removed.
Simply playing more lands will reduce the amount of games you are mana-screwd, as long as the deck has a strong late-game, adding a bunch of extra lands will at first look like the best option, but it fails as soon as you face decks that require you to have early action (like you need a specific 1-2 mana removal or you are quite dead, adding more lands will reduce the chance for that, which again makes your deck perform worse).
That means, for you observation that your decks become "better" by adding lands, thats because of the flaw in your simulation (most likely). Running it more often wont change that problem until you recognize it.
If your simulation just plays every hand no matter what, than adding MORE lands will usually increase consistency, as it will have more hands that actually have 2+ lands. However, if you properly mulligan hands, this wrong assumption will be removed.
Simply playing more lands will reduce the amount of games you are mana-screwd, as long as the deck has a strong late-game, adding a bunch of extra lands will at first look like the best option, but it fails as soon as you face decks that require you to have early action (like you need a specific 1-2 mana removal or you are quite dead, adding more lands will reduce the chance for that, which again makes your deck perform worse).
That means, for you observation that your decks become "better" by adding lands, thats because of the flaw in your simulation (most likely). Running it more often wont change that problem until you recognize it.
Actually, I included a good bit of mulligan logic. Too little mana, too much mana, not enough threats, and so on. Over time, the database gets built up too and you can start to see correlations between different starting hands and win rates. The information you're trying to get from such a system isn't the same information you're getting from playtesting. Learning that Languish is good is one thing, but it's quite another to take the Knightfall deck for example and know what the balance you want to strike between all your creatures is. How does adding another Retreat to Coralhelm change your average win rate, what's your chance to combo, how often will you mulligan, and so on.
Private Mod Note
():
Rollback Post to RevisionRollBack
To post a comment, please login or register a new account.
I've been debating if I should write a basic engine to test 2 color decks for optimal creature/spell curve and colored land distribution (re: limited). I've thrown together some basic simulations w/ colorless spells to find the best curve based on what turn you want to curve out by... but I've realized to uncover quality results a 2 color build is necessary. (significant more programming)
one sim I'd like to run is 8 & 9 lands vs 8 & 8 lands + 1 CIT dual (or more), to find out how much the CIT effect hinders a curve over the long run of course (1000s of simulations).
anyone know of such a thing?
There is a lot be gained here, just simply ignoring that is stupid. While it's true that we already have stats and a base of Pro knowledge, gaining access to more info is always a good thing. You could plug in deck list, go to work, and come back too see what happened. I know I would use that quite a bit.
Cheeri0sXWU
Reid Duke's Level One
Who's the Beatdown
Alt+0198=Æ
my first dilemma is how to approach checking lands (specifically duals) for proper mana to cast nth spell. i'm leaning towards having all colored permutations of mana available for that land state stored persistently .... or checking for available mana on the fly given nth spell's casting cost.
eh i'll figure it out as i dive in. more fun.
With pretty much no real choices in your game.
You can fairly well produce a deck this way, which will also work well.
As soon as theres some tricky moves involved or interaction with an opponent things get ugly quickly, as simulation for this is expensive and complicated, as you more or less build an AI that wants to play magic, which isnt trivial at all.
The basic idea of what you want can allready be accomplished with Math and an Excel Sheet, so you can optimise your numbers to make the best chance to archive your goal.
Math is your best friend in this regard, theres no need for simulations or an AI to play the game, as all we want is too optimize specific draws and make them as likely as possible.
However, this means we allready need to have at least an idea of what is optimal and a deck to begin with. This wont tell us to include cards we do not have on our radar and it will not point out specific interactions of cards and totally ignore the metagame (unless we use a goal like not losing to a wrath of god or graveyard hate etc.).
Overall, Excel Sheets are your best friend for most magic related as a player, to get a real idea of how your match % are and how even a single card will influence your performance against the field you expect (as putting a sideboard card in the maindeck might give you a real edge allready).
Decks like "storm" and burn can be fairly optimized to archive the quickest and most consistent kills.
WUBRG#BlackLotusMatterWUBRG
👮👮👮 #BlueLivesMatter 👮👮👮
I've built two of them, one for Modern Knightfall and the other for Legacy Burn. They gave some very interesting results. I didn't do much with the mana using rainbow lands but there was still a lot of data in just having different land counts. With the Knightfall deck (Burn is still a WIP as I'm trying to include a few possible opponents so it's not a pure goldfish, so I don't have complete results yet... will have to wait for summer break) I found that much higher than generally accepted land counts lead to the quickest wins on average. My system would take max/min parameters for various cards, and then try every 60-62 card permutation of that deck at 10k-250k games per permutation (depending on how many permutations were possible) and then record the results.
Actually, if you give a team of 20 pro's 2 weeks to focus on a single deck non stop, assuming 15 minutes per game and 14 hours/day put into testing you will only generate 7,840 games worth of data. With the automated goldfish approach, while you don't get matchup data you can generate literally millions of games worth of data in a single day. Also, believe it or not but Magic is an extremely high variance game and there is still a very noticeable effect from variance at even 100k games.
Actually, that little insight from variance made me reconsider several aspects of deck building and put far more value on the typical pro opinion of including a card just to have the option of hitting it, at the cost of some consistency.
The "quantity" of games doesnt matter as much as your thinking behind these.
You do not have to play 100+ matches against a deck to get any more information, if you get down to the fundamental important card interactions.
Most of the time a matchup will have something that is especially important to resolve, and you will win on that cards backpack alone. For example you play control, and you have to have a Languish to beat any aggro opponent that floods the board with tokens. But if you find that, you win fairly easily. That fact alone will have you mulligan for that card more aggressive and you do not play hands that you simply "know" wont do the trick in the matchup.
The problem with simulations that truly try every permutation is that its not real magic, you wouldnt play "any" hand, you mulligan them away. And your decision if a hand is a mulligan or not is greatly influenced by the matchup. However, thats only true for decks that have such a relevant specific card for matchups (allmost any controll deck will be this category) ; while you can much easier test decks that allmost dont care at all for the opponent, burn just hits for 20 damage, if you can do that by turn 4, you are pretty mcuh done with the deck, doesnt really matter what your opponent is doing, as you allmost cannot do anything to change that (but at the current level of burn decks, even they have specific cards that depend on the metagame, as a card like Eidolon of the Great Revel either do a ton or nothing, and cards like Skullcrack will give you a respons to lifegain, so the sequence of cards isnt as automatic, as you have to take your opponents actions into account to optimise your 20 damage goal).
Affinity, Infect and storm combo decks all do fairly well in simulations, as they have a very clear goal in mind. Even the Eldrazi deck was fairly good to test with simulations, as you want to optimise your turn ~5 kills at first ; and if you have these, you can start doing Simulations of your decks against each other to identify a "winner" of these top decks (but most of the time you can do a very well approximation by just a handfull of games with each match, as you only really want to see if a deck is overall more likely to win than another, the exact % isnt really relevant for that).
If your simulation just plays every hand no matter what, than adding MORE lands will usually increase consistency, as it will have more hands that actually have 2+ lands. However, if you properly mulligan hands, this wrong assumption will be removed.
Simply playing more lands will reduce the amount of games you are mana-screwd, as long as the deck has a strong late-game, adding a bunch of extra lands will at first look like the best option, but it fails as soon as you face decks that require you to have early action (like you need a specific 1-2 mana removal or you are quite dead, adding more lands will reduce the chance for that, which again makes your deck perform worse).
That means, for you observation that your decks become "better" by adding lands, thats because of the flaw in your simulation (most likely). Running it more often wont change that problem until you recognize it.
WUBRG#BlackLotusMatterWUBRG
👮👮👮 #BlueLivesMatter 👮👮👮
Actually, I included a good bit of mulligan logic. Too little mana, too much mana, not enough threats, and so on. Over time, the database gets built up too and you can start to see correlations between different starting hands and win rates. The information you're trying to get from such a system isn't the same information you're getting from playtesting. Learning that Languish is good is one thing, but it's quite another to take the Knightfall deck for example and know what the balance you want to strike between all your creatures is. How does adding another Retreat to Coralhelm change your average win rate, what's your chance to combo, how often will you mulligan, and so on.