MTG Salvation

Quote from moronstudios »

How's your RNN coming along with its learning? Is it able to produce working planeswalkers?

I'm at a satisfactory spot with it, pending any new developments that would let me train something bigger and better on the architecture I have access to. I can whip up new cards of a quality similar to the ones from my last couple posts on demand, which is good enough to keep my cube supplied for as infrequently as I get to play it!

Planeswalkers, well. Working yes, resembling anything in print, no. I get about one planeswalker or leveler per 300 cards I generate, and they always have only one loyalty ability or level up line. Four examples:

Geeler's Mivily 3B
Planeswalker ~ Gilt (common)
+2: Target player loses 2 life and you gain 1 life.
((4))

Voldering Vouggant 4B
Planeswalker ~ Dispersat (common)
+2: Voldering Vouggant deals 2 damage to target creature with flying and target creature or player. You lose 3 life.
((3))

Long of Jeggant 3BB
Legendary Creature ~ Faerie Rogue (rare)
Flying
Level Up 1B
Level 1~3
4/4
Pay 3 life: Return target nonlegendary creature card from your graveyard to your hand.
(3/3)

The Spike Wood 3W
Creature ~ Elf Wizard (common)
Level Up 2
Level 1~2
4/3
Lifelink
Mountainwalk
(1/1)

I'm throwing a couple into the cube as novelties, but it would be nice to get something more complete! I guess this is all one can expect from common rarity planeswalkers, though...

Those are amazing! I love that the Eldrazi got Devoid even though it doesn't require colored mana to begin with.

I find myself equally interested in "where did this art come from?" and "whose RNN spawned this?"

Man, I would poach that for the RoboCube I've got on hand for whenever, but 3 for 4/4 flying is a bit over the curve!

Quote from sirleandenor7 »

Oh look, your RNN made a delirium card.

Yeah! I also had one Investigate:

Paracoshu, Wirror's Changeling 2W
Creature ~ Serpent (common)
Defender
Flying
When Paracoshu, Wirror's Changeling becomes the target of a spell or ability, investigate.
(3/3)

There were also a ton of cards that said to transform them in different circumstances, but alas, none of them had associated flip sides.

For the folks who watch this thread for the sole purpose of seeing more generated cards, a random selection from my latest. No cherry-picking, no edits beyond capitalization/filling in "@"s/forum formatting.

Quote from hardcast_sixdrop »

That said, the library is getting out of date. I haven't updated it to include the latest sets, so the encoding algorithm probably breaks with the new colorless-only mana symbol.

Aha! I'd wondered why so many cards from AllSets.json were rejected as invalid. That could be at least part of it!

Because things were quiet here, I bulled ahead with the existing process. I had hoped to pull off a size 768, 3 layer network, but that ran out the memory on an EC2 g2.2xlarge instance. I fell back to size 512, which worked fine. Quirks of this net (at least at temperature 0.8+) included lands with mana costs and a mysterious card type "Rander" that showed up once, ha.

The biggest change is that char-rnn is no longer state of the art as a character-level language model backend. There is a full rewrite of it called torch-rnn that offers significant performance improvements and greatly reduced memory usage, which should make it possible to train much large networks. I've worked extensively with the codebase, and I'm currently in the process of adding much more powerful training and sampling interfaces to my own fork. It's a terrible mess of lua and python, but once I've got it working for my research project it should be easy to come back and do some clever things to improve magic card generation, like experimenting curriculum learning and other techniques to mix up the format during training to get as much out of the limited data as we can with a larger neural net.

Exciting! I will stay tuned here as always.

I managed to fix my problem by running through the tutorial using an Amazon EC2 instance instead of a local VM. No idea what the difference between Ubuntu environments is there, but it did the trick.

Is the tutorial at https://github.com/billzorn/mtgencode still the state of the art for card generation, or are there tweaks worth incorporating for the latest greatest output?

I'm trying to rebuild my card generation NN now that VirtualBox seems to be working for me again. Unfortunately, I'm running into an error when installing Torch.

I'm following the directions at http://torch.ch/docs/getting-started.html on an Ubuntu VM (the version packaged with VirtualBox). Cloning the git repo works fine, as does "cd ~/torch; bash install-deps;" When I attempt "./install.sh", however, I get this output:

Prefix set to /home/sabrecat/torch/install
No submodule mapping found in .gitmodules for path 'extra/iTorch'
Installing Lua version: LUAJIT21
CMake Error at exe/CMakeLists.txt:1 (ADD_SUBDIRECTORY):
The source directory
/home/sabrecat/torch/exe/luajit-rocks
does not contain a CMakeLists.txt file.

Any ideas what's going on there?

I know there are some cases where people have edited, but there are also cases where NNs with particular parameters get cautious and clone existing cards, only putting different names on them. It's tough to tell!

Wow, those are absolutely amazing! Your effort is paying off, there.

What do you plan to do with these once you've got the big commercially-printed pile of them?

Quote from maplesmall »

How many times do you think you'll do the voting process?

Actually, after playing another draft yesterday, I think I'm not going to do that after all. A big chunk of the fun is the novelty of the things--"creature gets +2/+2 and has feel" is hilarious the first time you see it, but next time around it merits no more than a smile and a nod.

I doubt I'll get to do this sort of thing on anything approaching a regular basis (I only got to pull this off due to a local games convention), but if I were, I'd try this: have the RNN output enough cards for a cube. Print 'em and go. At the end of the draft, each player picks perhaps five of their favorite cards; those go into the next cube, but the rest gets generated brand new.

Quote from MaximumC »

What I considered doing to fix this problem was actually use rarities. That is, commons are x4 in the cube, uncommons are x3, rares are x2, and mythics are x1. This leads to more expensive print run, but it allows you to manipulate the cards by moving absurd things up and necessary things (REMOVAL!) down in rarity.

I considered that too, but decided not to because in this goofy format, the more variety, the better. I think if I get to do the vote on/off process a few times, we'll beef up some needed niches by iteration.

They don't look all that special; they're prints using MSE defaults (I didn't take the time to get art for everything, yikes) on a printer that's starting to falter with old age, haha.

My fave from my deck was Cornepority, though I didn't actually get to play it. I also like Tormenter's Charm and Ethereal Bolt.

Player2 had a lot of respect for Goblin Marpher and Bloodruin Dead, in the voting round. Not sure what he most liked in his own deck other than the obvious power cards.

Player3 expressed a lot of love for Ancient Donger and Rold Dragon (she played Rold Dragon despite not doing much red, thanks to the green Suspend cost).

Player4 liked the token generators, particularly Steel's Flapse.

Gift of Caanndy got the most laughs, which landed it in the keep pile despite being kinda dubious otherwise. And of course everyone chuckled over Ancient Donger and Necropousy (M:tg came out when I was a preteen; it takes me back, I guess!). The text on Ageseal's Artifact broke Player3 pretty good, too.

I held my first robo-cube draft today!

To get my set, I pulled all @RoboRosewater cards, added a chunk of removal and black from early in this thread, and grabbed 180 at random to print off. I went the "damn the torpedoes" route advocated by MaximumC, and it proved to be a great time. We went for consensus ruling of the four players whenever weirdness came up, with an axiom of "where possible, interpret card text such that it does something."

My deck started off URW, which I swiftly pared down to UR after struggling to get the right mana in my first game. I went for high-efficiency creatures like Slidshocking Krow (ruling was that Tromple and Mointainspalk did nothing, alas; folks felt 4/2 for U was quite enough), bolstered with punchy R effects like Necropousy (+3/+3 totem armor) and Lightning Wind (+5/+5 and flying for the turn).

Player2 went BW, aiming for massive card draw with plays like Garruk, Leap in the Carapas (creature with 1: draw a card) to bring out the game-ending Giant Loyod (land, BB, T: deal 5 damage to a player).

Player3 was G with a splash of U, getting lots of use out of aggressive creatures like Emberty (3/2, haste, shadow--no other creatures with shadow were in the cube!), Callerwast (land that on ETB spawns a hasted 3/3 Saproling) and Colelenhan Right (3/1 for G with a tap ability to draw a card for bringing a creature onto the battlefield).

Player4 was BGU, fielding a main force of hitters like Onoch Wall (5/5, may cast for free) with artifacts like Untarswell of Aira (1, tap for G or G) and Steel's Flapse (3, same tap effect as Untarswell but also generates a tapped 2/2 Goblin each turn) to help ramp up.

After getting crushed by Player2 in first game, I tried to bring in some white hate in the form of Living Paladin (enchantment, XUU, white spells cost 3 more to cast), but he just shrugged and played mono-black. I couldn't keep his Aacotelb Mright down (5/5 for 3B, pay 2B to bring back to hand from graveyard).

My matchup against Player4 went much better; he only managed to get small token creatures out while I hammered away with Amel'rpen Threestel (5/5 for 3U, returns two creature cards to hand from graveyard every upkeep).

I had Player3 on the ropes for a little while with a cool sacrifice cycle going: Goblin Marpher (1/1 for R, sacrifice a creature max 1/turn to get 2 +1/+1 counters) fed by the Bisor Slaver (2/1 defender for 1U, returns to hand from graveyard whenever a creature ETB). But then she dropped Deers of Aodo (1G destroy all permanents), and recovered much better than I did from the board reset.

I don't know too much about what happened in the matchups I didn't play. I did catch that Player3 swapped in her sideboarded Path of Bone (enchantment, GGGG, players can't cast spells without paying their mana costs) to keep Onoch Wall out of her hair. ^.^

After the round-robin was done, we decided to hold a voting round where we went thumbs-up/thumbs-down on whether to keep each card in the cube. The voted-out cards will be ditched and replaced with new RNN cards for a later iteration of the cube!


EDIT: Here's the MSE to spare you having to hunt around to see what any of the above are!

Apparently I skipped way more Twitter @RoboRosewater cards than I'd thought. After I went back and put in every card I'd skipped, it was 182 creatures, 114 other.

I can probably make a 4-person cube out of what I've got, though as MaximumC noted, there is a serious lack of removal spells, hah. Definitely still interested in getting suggestions from folks for cards outside the Twitter set.