In Hearthstone’s March 14 update, the auto-completion function in the deck builder received a major upgrade. This is how it is described in the patch notes:
Complete My Deck – Hearthstone’s better at helping you build decks than ever before.
- Helps you build established, proven decks using your collection.
- If you’re short of an established deck, the helper will help you choose the strongest stand-alone cards from what you have available.
- NOTE: This feature is currently disabled in the Wild format. Please stay tuned for more info on when you can use the smart deck builder to help construct Wild decks.
The decks will be constructed based on internal Blizzard data, so in theory this feature could provide for the best decks that have been played in the game. Of course, there are plenty of obstacles on the way, such as:
- Recency: Are the decks good in the meta right now, or were they good at some point in the past?
- Sample: How big of a sample size is needed to determine the best decks? If multiple decks are within the margin of error, how is one chosen?
- Rank: Are the best decks the same regardless of rank, or do they change depending on rank? (They probably should, as the meta is different)
- Collection: How good is the feature in completing decks from a limited collection?
- Undiscovered decks: If something has not been tried, the algorithm does not know that it exists. Or does it attempt to build new decks based on data, and if so, how?
Given this context, even a sophisticated algorithm would have a hard time providing the absolute best deck to play at any given time.
It is interesting to see what the new feature can do, though, so we took it out for a spin and in this article we show all the current fully auto-completed decks for all classes as well as results of some experiments with the feature. Let’s see just how good this thing is!
Warrior
15,640
Completing a Warrior deck from scratch created an Odd Warrior build with Carnivorous Cubes and no Dragon cards at all. This is interesting, because statistics sites and analysis providers have praised the power of the Dragon synergy cards for some time now – numerous Dragon builds come up before the highest-ranking Cube build on HSReplay regardless of filter settings, for example. In fact, this exact list cannot be found on HSReplay at all, although several slightly different variants are available there – all of them near the bottom of Odd Warrior decks in performance according to the results of users of that site.
The deck itself is decent enough, it looks like an actual deck someone could put together, but the discrepancy between community-collected data and this specific list as a recommendation is interesting and warrants further examination.
Can we get a Dragon Warrior recommendation? I put in a pair of Emberscale Drakes as the foundation and let the deck builder work its magic from there, and lo and behold, it actually built me an Odd Dragon Warrior!
I actually found this build on HSReplay as well! It is not a very common one, there have been 430 games tracked with it over the past 30 days, and many of them are between ranks 10 and 25. Performance-wise, it seems to be a decent build and it looks better than the Cube list that the deck builder recommends without any nudging.
As the most recommended deck is an Odd Warrior, another easy experiment is putting in a non-Odd card and see what the deck builder recommends. It came up with this Rush Warrior build:
Again, this specific list cannot be found on HSReplay. It closely resembles some fairly well-performing Rush Warrior builds, but it goes for many single copies of cards, probably more than almost any human deck builder would. There are eight non-Legendary cards in the list that are present in a single copy only, many of them for no apparent reason.
I had one more experiment in mind for Warrior: Mecha’thun Warrior! With the rotation closing in, and Mecha’thun Warrior looking like one of the winners for Warrior class in the rotation, it is interesting to see what the best Mecha’thun Warrior build looks like! Alas, this is where the limits of the deck builder became apparent. It built me the same Rush Warrior build when I tried to auto-complete with Mecha'thun in the deck. Fine, maybe it did not recognize the combo, or maybe it thought that Rush Warrior with one bad card was still better than Mecha’thun Warrior.
As next attempt, I put in the full combo: Mecha'thun, Malygos, The Boomship, Whirlwind, and two copies of Inner Rage. That’s a six-card combo. Well. It built me the same Rush Warrior again, just with the combo included. Mecha’thun Warrior is beyond the capabilities of the autocomplete feature. Even 25 cards into the deck, the first card it wants to add is a Frothing Berserker. Because, you know, even a Mecha’thun Warrior should be ready to hit some face early in the game, right.
Shaman
Completing a Shaman deck from scratch created this Elemental Shaman. OK, it has Hagatha the Witch and Shudderwock, that’s the good old grindy control deck. Except that it has a Bloodlust in it. And Electra Stormsurge, apparently just to get more oompf from that one copy of Bloodlust and from random Hagatha the Witch spells, as there are no other cards that synergize with Electra in the deck. I am not completely sure what the deck builder was going for here, but those were some next-level card choices right there.
To put the tool to the test again, I gave it Thunderhead and Doomhammer to get started, and waited for it to build an Aggro Overload Shaman for me.
It kind of did! There are many familiar elements in the deck, although there are also curiously many one-offs there, including a seemingly random Lightning Storm. Also, the top end of the curve consists of one copy of Fire Elemental and one copy of Blazecaller. Truly next-level thinking in giving the deck some additional reach, or limitations of the algorithm?
At this point, I started to think that perhaps the algorithm is not a genius that will teach us the next level of Hearthstone even when it comes to meta decks.
Rogue
Completing a Rogue deck from scratch provided this Odd Rogue build. Yet again, there is a curious amount of one-offs, this time including a seemingly random Stonehill Defender. There are no Odd Rogue builds with Stonehill Defenders tracked on HSReplay.
Overall, however, the deck seems good. One Ironbeak Owl, one Void Ripper – the algorithm’s tech card game is on point.
As the most recommended Rogue deck is an Odd deck, it was easy to manipulate the deck builder to give me something else. I put in a pair of Backstabs to prime the machine, and it came up with this old-school Miracle Rogue:
I managed to track down this build from HSReplay. It is not a very popular one, and it does not have great results either, being at the low end of Miracle Rogue decks. Nonetheless, it is an actual deck that you can play on the ladder, so progress has been made compared to the previous auto-completion function.
Paladin
Completing a Paladin deck from scratch resulted in this Even Paladin list. It actually looks pretty strong, with the only real question mark being the single copy of Wild Pyromancer even though there is no Equality in the deck. Sure, Pyromancer can still do things without its long-time partner, but it is a rare choice to include one without the other: on HSReplay, I found exactly one such Even Paladin deck, but it is not this one.
I tried to make the deck builder build me some other Paladin decks. Odd Paladin was easy to get to, as many of the odd-cost cards provide enough of a hint to the algorithm. Here’s the pure version it built based on Baku the Mooneater alone:
A fine deck, indeed!
I also tried to get the builder to build me a Mech Paladin, but I was not successful. It always built either Odd Paladin or Secret Paladin, depending on what I provided as a starting point. Hilariously enough, Kangor's Endless Army resulted in a Secret Paladin deck that did not have a single Mech in it. When it comes to off-meta decks, the builder is clueless.
Hunter
Completing a Hunter deck from scratch resulted in a Midrange Hunter deck. This one can be found on HSReplay, and while it is also a rare deck, it seems to be performing well.
The deck builder also managed a start from Rhok'delar – it built a valid Spell Hunter list:
- 1Candleshot2
- 1Hunter’s Mark1
- 1Secret Plan2
- 1Tracking2
- 2Explosive Trap2
- 2Freezing Trap1
- 2Wandering Monster2
- 3Animal Companion2
- 3Deadly Shot2
- 3Eaglehorn Bow1
- 3Kill Command2
- 3Unleash the Hounds1
- 4Flanking Strike2
- 5Baited Arrow1
- 5Lesser Emerald Spellstone2
- 6Deathstalker Rexxar1
- 6To My Side!2
- 7Rhok’delar1
- 10Zul’jin1
Likewise, Deathrattle Hunter is a recognized archetype and came out nicely from Devilsaur Egg as seed:
Druid
Completing a Druid deck from scratch resulted in this Hakkar Druid. It is a novel build that cannot be found on HSReplay.
Druid deck builder could take the hint when I offered it Malygos as seed, and it promptly created this Malygos Druid:
However, I could not make it build me a Token Druid. Most attempts resulted in weird Mecha’thun Druid decks. Neither a regular Token Druid or Treant Token Druid came up.
Warlock
Completing a Warlock deck from scratch yielded this Mecha’thun Warlock deck. It is a perfectly fine sample of the archetype and fully ready for ladder.
The deck builder is also familiar with Even Warlock: giving it Genn Greymane as a starting point resulted in a complete Even Warlock deck:
Fooling around by placing Baku the Mooneater as seed confused the Warlock deck generator and results in a strange mismash that included Prince Taldaram in combination with other three-drops in the deck.
It was probably thinking of Cubelock, I assumed, and indeed, a slight nudge with Skull of the Man'ari resulted in this Cube Warlock deck:
Mage
Completing a Mage deck from scratch created this Odd Control Mage deck. It is a novel build that opts to forgo Blazecallers in favor of Leeroy Jenkins, but still keeps Fire Fly in the deck, even though it does not require Elemental synergies for anything. A slightly strange build, but a valid one nonetheless.
But can it build me an aggressive deck? It sure can! Using Kirin Tor Mage as seed, the deck builder came up with this state-of-the-art Secret Odd Mage, complete with Subject 9. That’s impressive.
Priest
- 1Holy Smite2
- 1Power Word: Shield1
- 2Mind Blast2
- 2Radiant Elemental2
- 2Shadow Visions2
- 2Shadow Word: Death1
- 2Spirit Lash2
- 3Gilded Gargoyle1
- 3Vivid Nightmare1
- 4Eternal Servitude2
- 4Mass Dispel1
- 5Mass Hysteria2
- 7Lesser Diamond Spellstone2
- 7Prophet Velen1
- 7Psychic Scream2
- 7Shadow Essence2
- 8Shadowreaper Anduin1
- 9Zerek’s Cloning Gallery1
Completing a Priest deck from scratch provided this Clone Priest. It’s an interesting list, in fact. It is almost the pure OTK version, but it includes one copy of Gilded Gargoyle, which means that Greater Diamond Spellstone is not guaranteed to summon both Prophet Velen and Malygos, but on the other hand the deck has some potential to highroll with coins.
The Priest deck builder did not recognize Mecha’thun Priest or APM Priest, but it was able to fill out Mind Blast Priest and Wall Priest.
Budget Experiences
All of the above was with an almost full collection that includes all the relevant meta cards. What about budget players, the ones the feature is supposed to help? In order to try it out, I logged on to my NA account, where I have but a small collection and only test some budget decks occasionally.
My experience with the deck builder was not very good there. As an upside, it tends to build a decent curve for the deck, and having a good mana curve is one of the fundamentals of Hearthstone: if you cannot play cards, you cannot win. Most of the time, it also fills in blanks with generally good cards for the class.
However, it also did some weird things, such as:
- Built a sort-of Heal Zoo from scratch, except that the deck had neither Happy Ghouls nor Lightwardens. It also had no two-mana cards whatsoever, probably because I did not have Prince Keleseth in my collection. This is for an account where I have built and played Budget Zoo decks, so the main budget alternatives used in such decks were available.
- Built a sort-of Spiteful Druid from scratch. Well, it had Grand Archivist in it, but there were no spells. Literally, no spells whatsoever.
- Completed a Budget Hunter by adding in Lesser Emerald Spellstone to a deck that had no Secrets.
Conclusions
The auto-complete function is now better than it was before. In general, it will strive to build decks with a good mana curve that consist of powerful cards. However, it is far from perfect.
The deck builder is mostly based on well-established meta archetypes. If you have most of the cards for one, it will build you the base and try to fit in individually powerful cards for the remaining slots. The base it uses does not seem to be optimal, but it is close enough. Much of the time, all of this works out quite well for meta decks.
However, the deck builder has no understanding of context. This was demonstrated in multiple ways: It left gaps in the mana curve (when Prince Keleseth was not available) and it added cards that it considered powerful, but which could not function without synergy cards that were not present in the decks (the cases of Grand Archivist, Lesser Emerald Spellstone, Prince Taldaram, and Kangor's Endless Army to name a few).
The deck builder also goes completely astray when presented with an off-meta deck. Even when the deck is almost completely built, it adds powerful cards from one of the archetypes it recognizes with no consideration on how they fit in the deck they are added to.
As long as the deck you are adding cards to is one of the main archetypes that the deck builder recognizes, it is able to add appropriate cards to fill the deck. The moment you have something else in your hands, it can get very confused and add in completely useless cards. While this is an improvement over the previous state of the function, the new deck builder does not in fact provide you with the most powerful decks even if you have all the cards available, and its ability to fill in good replacements depends on how close your deck is to a meta deck so that it does not get confused.
There is still plenty of deck building left in Hearthstone, the new Complete My Deck function does not replace creative work in the slightest.
Old Guardian
Ville "Old Guardian" Kilkku is a writer and video creator focused on analytic, educational Hearthstone, and building innovative Standard format decks. Youtube: https://www.youtube.com/c/OldGuardian Twitch: https://www.twitch.tv/old_guardian
[…] you want our input on the feature, Old Guardian has already tested it quite extensively in this article, and while it’s a massive upgrade over the previous “auto complete” feature, […]
Overall i really like it. It simply throws the highest winrate cards into a specific archetype. You can throw Baku into a Warlock Deck and press the complete button. Sounds fun how it works. But if you choose something like lets say Paladin Quest it will put the OTK Shirvalla Deck together. So i think this needs rather more data or more time. If you choose Eterium Rover and Dr.Boom, to hope for MechWarrior, it will build you Odd/Warrior. So its kinda meh….
Can’t wait for the wild format to be available as well so it can build me a Wild Reincarnate Shaman deck with Kel’Thuzad
Thanks for the leg work – but it proves, as suspected, that it is almost completely useless!
For the sake of Devil’s Advocate: just because it has no direct benefit to you doesn’t mean its useless. I suspect Blizzard’s main goal was to make the game more “casual-friendly”–i.e., now you can build reasonably competitive decks without ever visiting a 3rd party site, so less involved players might be less discouraged and a play a little more which is generally good for everyone.
But yeah, I’ll go back to my earlier comment and say I too am pretty disappointed that the tool isn’t more competent. Hopefully it improves with time.
It built me a very sweet murloc deck for shaman (questless as i did not have the card), so it’s possible to create off-meta decks depending on the collection
Rogue also recognizes Hooktusk tempo rogue and Myracle Rogue, at least on my account.
The article does not aim to provide a comprehensive list of all the recognized archetypes. That said, only fairly popular ones seem to be recognized, based on all the ones that are not.
True true. What is interesting is the Miracle Rogue list it built for you is the exact same that Kyouma used to reach top 3 legend earlier today. Somewhat spooky lol.
(source: https://twitter.com/Kyouma_HS/status/1106569262997463040 )
Just logged in to EU server, where I don’t have much of a collection. The Innkeeper made me a deck for one of the welcome back quests, looks like it used the deck building system to create it. It was a passable tempo rogue deck, but it had 1x Cavern Shinyfinder with no weapons; especially strange, because I actually have Kingsbane in that collection.
**IF** it is based on machine learning algorithms, maybe it just needs more time…
The developers said that it takes a couple of days for it to gather enough data for a new set. But this is not a new set, the meta is stable. Here’s the interview: https://www.hearthstonetopdecks.com/upcoming-hearthstone-game-updates-smarter-deck-builder-random-card-backs-arena-seasons-sneak-peek-at-the-new-single-player-experience/
Good work…!!
Another great article, Guardian, although it’s very disappointing news. I was REALLY hoping for something like “hey, I’ve opened Undertakah and Tirion, can you make me an interesting off-meta deck around that?” This sounds like almost the exact opposite: it reinforces “proven” meta synergies and just fails to recognize new or off-meta synergies.
Yes, that is exactly how it works. I too was expecting off-meta insights based on Blizzard data, although I guess that would have taken some of the creativity away from deck building. Alas, the way it works is that it steers you towards established meta archetypes.