Ive gone through a number of drafts and while the bots seem to do pretty well, the pool of cards seems really off particularly around removal and lack of common removal not showing up, like Unholy Hunger while other like Suppression Bonds seems super common.
My 3rd time trying it, I built another ridiculously synergistic deck.
I suspect the algorithm only evaluates cards for their raw power in a vaccuum and not their synergy. A human would notice an archetype being open as a signal, but the bot doesn't, so you can pull together cards you wouldn't normally get.
It also seems easy to wheel rare dual lands and rares that are low P1P1 but have synergy value. And that Overrun card goes around WAY too much in pack 2 and pack 3 if none of the bots picked GW. In another draft, I got dumped 2 copies in pack 3 and decided to just splash green FTW.
I probably shouldn't be able to draft 5 instants that read "destrot target creature" (passed a 3rd conclusion too). But none of the bots were BR so they kept coming. Also, drafted Syndicate Trafficker + Animation Module.
@Cele - at one point I had included autobuilding bot decks, but it was a bit computationally intensive for slow browser. There's a new team in charge of development now, maybe they'll think of implementing something like this
@FTW - yes it's pure power, no synergy. Some of the mathematical details might get posted on the methodology page soon
I dig the service, it helped me get a bit better at keeping track of picking up the right amount of all constituents for a deck. I have two suggestions though, as you already have a colour synergy rating, expanding that functionality a bit wouldn't hurt. Adding some sort of curve tracker into the offered ratings as another score wouldn't hurt. You get passed a pack with a two drop that's slightly worse than the four drop, but your pool is in dire need of two drops, so the rating gets bumped. I'd imagine this should be reasonably easy to do, and could be carried over into the automated deck building to avoid the top-heavy monstrosities the thing is quite prone to spit out from time to time. A more exhaustive project would be to try to emulate actual synergy in yet another rating tab, but that would be labour intensive and difficult to pull off. Plus, I imagine services like these are of most use to greenhorns like yours truly that need a crash course in how to draft, not people who need the bots to act at a super nuanced level including trying to build wacky decks with crazy interactions from seemingly janky cards. These are but suggestions, thank you so much for the page!
It seems pretty smart. It correctly predicted all but four or five of my picks when I drafted. The only thing that seemed dramatically off was dual lands. I was solidly in red/black in an SOI draft, when a pack came up with a Stone Quarry, and a bunch of mediocre off-color cards. As a human drafter, I instantly thought "This will come in handy if I need to splash white." This made it easily the best card in the pack, however, the computer's color indicator told it that that was a very bad idea. I understand why you want to keep it to two colored decks, but it is generally a draft principal that a card that is half on-color is better than a card that is all off-color, especially if that card is a land.
I wonder if the card rankings are flawed. I just drafted FOUR Renegade Freighter in pack 1, then was passed a 5th early in pack 2 and had to begrudgingly pass it. No matter what color the bots are in, all decks can play Freighter. Should they not be taking it? It's the best common.
Is the code open source?
I have some ideas how to encode synergy into picks. Curve is a little trickier without running a full deck simulator, but if you track the expected CMC of cards you think will pick, you might be able to assign a value to the cards cost independent of (power) ranking. Or you could rank cards by CMC + power (instead of pick order) and mix the function.
First, the Renegade Freighter rating has been updated (since we all know how good it is now).
@zenbitz I think that the right way to implement deck composition is to implement an empirical "curve bonus." For example, the first three-drop in the deck gets a +1.0 bonus, the fourth gets a +1.0 bonus, and the seventh gets a -0.5 bonus. This will guide decks towards 5-6 three drops.
To enforce this during the draft, an "expected number" of three drops can be introduced (i.e. if a player has 2 three drops halfway through the draft, they are on pace for picking up a total of 4).
First, the Renegade Freighter rating has been updated (since we all know how good it is now).
@zenbitz I think that the right way to implement deck composition is to implement an empirical "curve bonus." For example, the first three-drop in the deck gets a +1.0 bonus, the fourth gets a +1.0 bonus, and the seventh gets a -0.5 bonus. This will guide decks towards 5-6 three drops.
To enforce this during the draft, an "expected number" of three drops can be introduced (i.e. if a player has 2 three drops halfway through the draft, they are on pace for picking up a total of 4).
That sounds like a pretty good way to do it, and it seems like it shouldn't be too hard to implement.
First, the Renegade Freighter rating has been updated (since we all know how good it is now).
@zenbitz I think that the right way to implement deck composition is to implement an empirical "curve bonus." For example, the first three-drop in the deck gets a +1.0 bonus, the fourth gets a +1.0 bonus, and the seventh gets a -0.5 bonus. This will guide decks towards 5-6 three drops.
To enforce this during the draft, an "expected number" of three drops can be introduced (i.e. if a player has 2 three drops halfway through the draft, they are on pace for picking up a total of 4).
Good idea. I think you should evaluate the expectation after each pack is done though (in 3rds), not halfway through.
Halfway through the draft (P2P8-ish), the next half you're about to see has lower average card quality than the first half. You will see only 1 set of early picks and 2 sets of dumps, whereas before you saw 2 sets of early picks and only 1 set of dumps. So, for example, in a format where 2-drops are aggressively drafted, if you have only 2 2-drops by P2P8, you are not "on pace" to pick up a total of 4. To get 4, you may have to prioritize 2-drops over other things in pack 3.
If you do it in 3rds instead of halves (i.e. after each pack), the math works out a little better.
I tried forcing mono-red in SOI draft and the Autobuild still added 8 islands for just a Vessel of Paramnesia and a Geistblast. It's way too biased towards a dual-colored deck. I've also noticed the suggestions almost ignores the curve of cards you've already picked.
I took this for a spin with KLD drafts and it felt pretty off base. I made some screen grabs during the second draft to use as a play by play. P1P1 was Servant of the Conduit and not particularly close. Pick 2 is where the recommended card (which I assume is the same algorithm the bots use) went off the rails. Built to Smash is not even close to Seedsculptor (I think cheap 2 power beaters are undervalued and tricks overvalued in general); Brazen Scourge actually is better in a vacuum, but not enough to justify the 1CC off color cost. Pick 3 it got right, but it really shouldn't be this close; with 2 green cards already I would not want to move into white just for Aviary Mechanic when I can just take Appetite. Pick 4 is where it really breaks down; we're taking an off color card here, but it wants me to take a mediocre five drop over solid curve fillers. Picks 2 and 3 function better in the b/g deck, and Fretwork is a big roleplayer in that deck, yet it's way at the bottom of the recommendations.
I know accounting for synergy is difficult, but looking at the recommendations and AI decks I don't think it even accounts for curve or card type (it made a deck with 4 combat tricks, a vehicle, and an equipment out of my pool). Some tweaks that I would make are knocking down the ratings for combat tricks, raising them for curve out creatures, especially 1-2 mana 2 power ones, increasing the preference for on color/colorless cards, and account for cards that get better or worse in multiples (the auto build main decked a single Self Assembler that I speculatively picked). When I looked at the AI decks one of them was clearly in white, but drafted the other four colors in roughly equal amounts and they were all taking combat tricks way too early in packs.
That's correct - right now it is not taking into account curve and synergy/card type. We are looking into implementing these in the future but hopefully that explains the current suggestions you're seeing.
The ratings are all maintained by me and based on experience, professional reviews, pick order lists, etc. Essentially what the rating represents is a "pick one pack one" decision based on sheer power level, with some consideration for color. As you move farther away from P1P1 and start building based on synergy, the engine is still only going to make recommendations based on power and your color commitment, not on synergy with what you have.
I will say about Built to Smash specifically that it is the second best red common after Welding Sparks. So part of this depends on your view of the format and the fact that WotC lately likes pushing strong combat tricks over removal, for example.
That's correct - right now it is not taking into account curve and synergy/card type. We are looking into implementing these in the future but hopefully that explains the current suggestions you're seeing.
The ratings are all maintained by me and based on experience, professional reviews, pick order lists, etc. Essentially what the rating represents is a "pick one pack one" decision based on sheer power level, with some consideration for color. As you move farther away from P1P1 and start building based on synergy, the engine is still only going to make recommendations based on power and your color commitment, not on synergy with what you have.
I will say about Built to Smash specifically that it is the second best red common after Welding Sparks. So part of this depends on your view of the format and the fact that WotC lately likes pushing strong combat tricks over removal, for example.
Just based off of the screenshots above, the base ratings seem a bit off, and perhaps a bit too guided by the initial tier lists of the format were - have you updated it for what the current evaluation of the format is?
The more obvious examples -
Curio Vendor as a 1.8 when Spireside Infiltrator is a 1.9 - Infiltrator is good curve filler in red that you're generally happy playing, when Curio Vendor is very close to unplayable.
Ghirapur Guide is a 2.4 and has the same rating as Kujar Seedsculptor, when the guide is bad curve filler and seedculptor is a premium 2 drop that you never cut.
Veldaken Blademaster having a higher rating as Spireside infiltrator, when blademaster is almost as bad as curio vendor.
That's correct - right now it is not taking into account curve and synergy/card type. We are looking into implementing these in the future but hopefully that explains the current suggestions you're seeing.
The ratings are all maintained by me and based on experience, professional reviews, pick order lists, etc. Essentially what the rating represents is a "pick one pack one" decision based on sheer power level, with some consideration for color. As you move farther away from P1P1 and start building based on synergy, the engine is still only going to make recommendations based on power and your color commitment, not on synergy with what you have.
I will say about Built to Smash specifically that it is the second best red commonH after Welding Sparks. So part of this depends on your view of the format and the fact that WotC lately likes pushing strong combat tricks over removal, for example.
Channelfireball has built to smash as the second best red common (corresponding to ~2.8) since it's critical to the aggressive red archetype.
In this deck, it's quite powerful and gets in a lot of damage. However, to a certain extent, it is situational (combat trick, requires attacking, likes artifact creatures).
Historically, quality creatures have been in the 2.0-3.0 range and tricks have been in the 1.7-2.0 range. This makes the suggested picks and auto-building lean toward stable, creature-heavy decks.
Even if we don't get into whether or not Built to Smash is the second best red common, it certainly doesn't deserve to be taken P1P2. Combat tricks are completely ineffective if you don't have a strong enough board presence to be the aggressor, and taking a trick over a solid 2-3 mana beater reduces your ability to have that presence. Combat tricks, equipment, non-removal auras, and vehicles all rely on you having creatures on board, and as I mentioned I took 6 such cards, but would never actually run them all. For the AI behavior to be realistic and competitive it needs to realize that all Built to Smashes are not created equal, and that the card hates to play with other copies of itself and really anything that isn't a creature. Good 2 drops are kind of at a premium in this format and there was already a good green card in my pool; Built was not even on my radar as a card to think about taking there. I think there need to be separate ratings for the draft and deck building phases, because a combat trick can be a strong addition to your deck (4 cannot), but it's not something you should prioritize unless you somehow manage to make it to pack 3 without taking any, or if there are strong synergies, like Armorcraft Judge + Subtle Strike, that make one trick way better than any of the others for your particular deck. I'm confident that a large majority of experienced drafters would agree with my pick, and I'm confident most who don't would take Brazen Scourge, not Built to Smash. If you want to improve the simulator then adjusting it to focus more on curving out and disfavoring cards that are bad in multiples is a good place to start.
I've never not had a combat trick that I desperately needed in my pool, but I can't say the same about 2 drops, removal, and fixing.
Hey everybody - I just wanted to let you know that I've updated the site for Aether Revolt. You can now practice both draft and sealed. Here's the link for draft: http://draftsim.com/draft.php?mode=Draft_AER
Autobuilder won't even bother trying to go 3 colors, even slotting utter trash to remain in your main colors.
Been toying around and drafted a lot of naya token (modern masters 2017) and won't include 3 colors cards, even though they should made the cut.
Aside from that, nice tool.
EDIT: Just had an izzet deck filled with 25 lands insted of going 17 and putting in 8 black cards to go Grixis (my pool had notably 2 terminate, a desecration demon and a soul manipulation). Something seems off here, at least for triple colors.
One more 27 landers generated by the autobuild feature. This time with draft pool:
Yes, unfortunately the logic for autobuild only builds two color decks right now. This assumption was fine 99% of the time for formats before this one but it will need to be rewritten for multicolored sets in the future.
Private Mod Note
():
Rollback Post to RevisionRollBack
To post a comment, please login or register a new account.
I suspect the algorithm only evaluates cards for their raw power in a vaccuum and not their synergy. A human would notice an archetype being open as a signal, but the bot doesn't, so you can pull together cards you wouldn't normally get.
It also seems easy to wheel rare dual lands and rares that are low P1P1 but have synergy value. And that Overrun card goes around WAY too much in pack 2 and pack 3 if none of the bots picked GW. In another draft, I got dumped 2 copies in pack 3 and decided to just splash green FTW.
1 Night Market Lookout
3 Fretwork Colony
1 Smuggler's Copter
3 Prakhata Pillar-Bug
2 Workshop Assistant
1 Bomat Bazaar Barge
2 Ovalchase Daredevil
1 Bastion Mastodon
1 Multiform Wonder
//Spells: 7
1 Key to the City
1 Prophetic Prism
1 Morbid Curiosity
2 Essence Extraction
2 Tidy Conclusion
1 Dovin Baan
//Lands: 17
10 Swamp
1 Concealed Courtyard
1 Aether Hub
3 Island
2 Plains
1 Night Market Lookout
1 Morbid Curiosity
1 Whirlermaker
1 Rush of Vitality
1 Select for Inspection
1 Lawless Broker
1 Gearseeker Servant
1 Contraband Kingpin
So much recursion.
Key + Ovalchase.
Workshop Assistant loops.
Ovalchase + Morbid Curiosity.
Nightmarket + Vehicles.
Fretwork + 4 lifelink creatures + 4 removal that gain life
Splash 2 colors for P3P1 planeswalker.
So much removal.
3 Reckless Fireweaver
1 Embraal Bruiser
1 Eager Construct
1 Syndicate Trafficker
1 Cultivator's Caravan
1 Filigree Familiar
1 Foundry Inspector
1 Foundry Screecher
1 Salivating Gremlins
1 Weaponcraft Enthusiast
1 Maulfist Squad
1 Snare Thopter
1 Animation Module
1 Key to the City
1 Metalspinner's Puzzleknot
1 Harnessed Lightning
1 Underhanded Designs
3 Unlicensed Disintegration
2 Tidy Conclusion
8 Mountain
8 Swamp
I probably shouldn't be able to draft 5 instants that read "destrot target creature" (passed a 3rd conclusion too). But none of the bots were BR so they kept coming. Also, drafted Syndicate Trafficker + Animation Module.
@FTW - yes it's pure power, no synergy. Some of the mathematical details might get posted on the methodology page soon
Low-power cube enthusiast!
My 1570 card cube (no longer updated)
My 415 Peasant+ Artifact and Enchantment Cube
Ever-Expanding "Just throw it in" cube.
I wonder if the card rankings are flawed. I just drafted FOUR Renegade Freighter in pack 1, then was passed a 5th early in pack 2 and had to begrudgingly pass it. No matter what color the bots are in, all decks can play Freighter. Should they not be taking it? It's the best common.
By pack 2 pick 5, this was my deck:
2 Eager Construct
1 Speedway Fanatic
1 Foundry Screecher
4 Renegade Freighter
1 Ovalchase Daredevil
1 Dukhara Scavenger
1 Subtle Strike
1 Die Young
2 Rush of Vitality
1 Fateful Showdown
I can even give Freighter haste or indestructible and bring one back from the grave. I didn't bother to finish the draft as it seemed too unrealistic.
I have some ideas how to encode synergy into picks. Curve is a little trickier without running a full deck simulator, but if you track the expected CMC of cards you think will pick, you might be able to assign a value to the cards cost independent of (power) ranking. Or you could rank cards by CMC + power (instead of pick order) and mix the function.
@zenbitz I think that the right way to implement deck composition is to implement an empirical "curve bonus." For example, the first three-drop in the deck gets a +1.0 bonus, the fourth gets a +1.0 bonus, and the seventh gets a -0.5 bonus. This will guide decks towards 5-6 three drops.
To enforce this during the draft, an "expected number" of three drops can be introduced (i.e. if a player has 2 three drops halfway through the draft, they are on pace for picking up a total of 4).
That sounds like a pretty good way to do it, and it seems like it shouldn't be too hard to implement.
Good idea. I think you should evaluate the expectation after each pack is done though (in 3rds), not halfway through.
Halfway through the draft (P2P8-ish), the next half you're about to see has lower average card quality than the first half. You will see only 1 set of early picks and 2 sets of dumps, whereas before you saw 2 sets of early picks and only 1 set of dumps. So, for example, in a format where 2-drops are aggressively drafted, if you have only 2 2-drops by P2P8, you are not "on pace" to pick up a total of 4. To get 4, you may have to prioritize 2-drops over other things in pack 3.
If you do it in 3rds instead of halves (i.e. after each pack), the math works out a little better.
Well its pretty easy in the default midrange case, but its deck dependent
The other issue is weighting curve vs power.
I know accounting for synergy is difficult, but looking at the recommendations and AI decks I don't think it even accounts for curve or card type (it made a deck with 4 combat tricks, a vehicle, and an equipment out of my pool). Some tweaks that I would make are knocking down the ratings for combat tricks, raising them for curve out creatures, especially 1-2 mana 2 power ones, increasing the preference for on color/colorless cards, and account for cards that get better or worse in multiples (the auto build main decked a single Self Assembler that I speculatively picked). When I looked at the AI decks one of them was clearly in white, but drafted the other four colors in roughly equal amounts and they were all taking combat tricks way too early in packs.
That's correct - right now it is not taking into account curve and synergy/card type. We are looking into implementing these in the future but hopefully that explains the current suggestions you're seeing.
The ratings are all maintained by me and based on experience, professional reviews, pick order lists, etc. Essentially what the rating represents is a "pick one pack one" decision based on sheer power level, with some consideration for color. As you move farther away from P1P1 and start building based on synergy, the engine is still only going to make recommendations based on power and your color commitment, not on synergy with what you have.
I will say about Built to Smash specifically that it is the second best red common after Welding Sparks. So part of this depends on your view of the format and the fact that WotC lately likes pushing strong combat tricks over removal, for example.
Just based off of the screenshots above, the base ratings seem a bit off, and perhaps a bit too guided by the initial tier lists of the format were - have you updated it for what the current evaluation of the format is?
The more obvious examples -
Curio Vendor as a 1.8 when Spireside Infiltrator is a 1.9 - Infiltrator is good curve filler in red that you're generally happy playing, when Curio Vendor is very close to unplayable.
Ghirapur Guide is a 2.4 and has the same rating as Kujar Seedsculptor, when the guide is bad curve filler and seedculptor is a premium 2 drop that you never cut.
Veldaken Blademaster having a higher rating as Spireside infiltrator, when blademaster is almost as bad as curio vendor.
Channelfireball has built to smash as the second best red common (corresponding to ~2.8) since it's critical to the aggressive red archetype.
In this deck, it's quite powerful and gets in a lot of damage. However, to a certain extent, it is situational (combat trick, requires attacking, likes artifact creatures).
Historically, quality creatures have been in the 2.0-3.0 range and tricks have been in the 1.7-2.0 range. This makes the suggested picks and auto-building lean toward stable, creature-heavy decks.
I've never not had a combat trick that I desperately needed in my pool, but I can't say the same about 2 drops, removal, and fixing.
Been toying around and drafted a lot of naya token (modern masters 2017) and won't include 3 colors cards, even though they should made the cut.
Aside from that, nice tool.
EDIT: Just had an izzet deck filled with 25 lands insted of going 17 and putting in 8 black cards to go Grixis (my pool had notably 2 terminate, a desecration demon and a soul manipulation). Something seems off here, at least for triple colors.
//Deck from draftsim.com
1 Goblin Guide
1 Ulvenwald Tracker
2 Hellrider
1 Dynacharge
2 Dragon Fodder
1 Fists of Ironwood
1 Madcap Skills
1 Magma Jet
1 Blood Moon
1 Call of the Herd
1 Slime Molding
1 Gruul Guildgate
13 Mountain
13 Forest
1 Arid Mesa
1 Gideon's Lawkeeper
1 Soul Warden
1 Attended Knight
1 Inquisition of Kozilek
2 Intangible Virtue
1 Rootborn Defenses
2 Zur the Enchanter
1 Bronzebeak Moa
2 Momentary Blink
2 Misty Rainforest
1 Banishing Stroke
1 Eyes in the Skies
1 Sunhome Guildmage
1 Boros Guildgate
1 Sprouting Thrinax
1 Mystic Genesis
2 Mystical Teachings
1 Graceful Reprieve
1 Woolly Thoctar
2 Thundersong Trumpeter
1 Master Splicer
1 Familiar's Ruse
1 Aethermage's Touch
1 Kraken Hatchling
Yes, unfortunately the logic for autobuild only builds two color decks right now. This assumption was fine 99% of the time for formats before this one but it will need to be rewritten for multicolored sets in the future.