Would it be possible to layer this on the parameters of a design skeleton? One of Reuben's projects has been to create a basic "paint by numbers" defined skeleton for producing sets. If designers could take the basic parameters like "make cards with these mechanics, at these cmcs and these rarities..." A small amount of direction might go along way.
Yes. You could prime the network with restrictions such that it produces a card that follows those restrictions. A hand-crafted program can sit on top of the network and dictate what needs to happen.
I just signed up to say how incredibly impressive this is.
Please continue working on this. This is the most fun thing in Magic ive seen in a long time.
How much do I need to understand of neural networks and the code to try and run this myself?
Thanks! And not much, actually. The program that I used runs out of the box, but it does have some dependencies that you will need to install. Specifically, you'll need the Torch scientific computing framework. I had no problems installing it, but then again, it's the sort of thing I do every day. Now, to improve upon what I've done, we'll need to make some readjustments to the program, which does require some knowledge (or a whole hell of a lot of trial and error, take your pick).
Also, I'll warn you that I had to reorganize the data to allow for effective training. The cards in json format have most of the important information occurring at the end of the record. In my first experiment, I realized that the network interpreted the data to mean that the text in the body of the card determined the type, mana cost, and name, instead of the other way around.
Now many cards does the network need to be trained to some reasonable level?
It's best if I have all of them, the more the merrier, but I get decent results at the halfway point, which is about 7000 cards.
If anything though I feel we could do with more cards than actually exist in Magic in order to get better results, that is, we could mutate some of the inputs to expand the card pool. For example, take a creature, tweak its power, toughness, abilities, and cost ever so slightly, and then present that one as well. This is going to be necessary if we want to train a network to produce planeswalkers, since it has so few working examples. For example, take Jace, switch out some of his abilities, and resubmit it to the network.
Would making a separate network for different colors (and even card types) work better, so you don't get blue creature with red spell ability etc? I have had my neural nets course years ago, so I'm really out of the loop on this.
Partitioning the network into subnetworks could work. But what's interesting is that state of the art approaches don't even require this anymore: with the right training regime, it efficiently self-organizes. Subregions of the network become attuned to different components (a creature region, an ability region, etc.). It's very difficult to tease out what does what (that takes a lot of reverse engineering), but it's obvious that that is what is happening.
Is it possible to give it some basic design rules? Stuff like 'creatures you control gain rather than lose ____' and 'if designing an enchantment, abilities are persistant' and the like? Basically teaching it some of the context associated with the card rules?
A lot of that should come naturally with a deeper network. Right now the network is very parsimonious with its design rules because I deliberately kept everything as simple as I could.
This is *super* cool! I wish I'd thought of something like this. I don't fully comprehend what's going on (how the RNN works programatically), but I'd love to learn. Anyway, the results are awesome. I think it'd be super awesome if you published all of the results to a blog or something, I want to be able to see what those ridiculous/absurd/meaningless first cards were like, and I want to see how it improves and what kind of crazy stuff it comes up with.
Green Angel tokens reminded me of a great fantasy series called Memory, Sorrow, and Thorn (there is an important landmark in the series called Green Angel Tower).
I don't know what's so funny about "Tromple,Mointainspalk," but I have literally been laughing uncontrollably for the last ten minutes.
I started giggling as soon as Amarogge Warfos let you put a 3/1 green solider token onto the battlefield, only to cruelly force you to put it directly into your graveyard, but yeah this combination of fake keywords absolutely slayed me.
Of the more recent cards, both Gravimite and Shring the Artist are dangerously close to being printable (albeit fairly powerful in each case, and Gravimite should lose the white hybrid).
Watson has been making recipes lately. I don't see this as being radically different. If it can learn the difference between good design and bad, it might make some really great designs.
EMERGENCY EDIT: At the time of this post there's six HUNDRED thirteen people viewing this thread!
Thats because I had the silly idea of posting the thread on reddit.
Anyway I find this fascinating as I've done programming of neural networks before but usually simple ones with back propagation.
I'll have to take a look at the source code once I'm less busy in RL.
I don't know what's so funny about "Tromple,Mointainspalk," but I have literally been laughing uncontrollably for the last ten minutes.
Yeah, same here. I think it's kind of like watching an orangutan trying to use a saw but holding it upside down. This AI works with basic pattern for Magic cards but beyond that has no idea what it's doing.
Out of curiosity, what kind of "hardware" would be required to make a more robust network? What kind of parameters would you suggest if you had no hardware limits?
Private Mod Note
():
Rollback Post to RevisionRollBack
My Moderator Helpdesk
Currently Playing:
Legacy: Something U/W Controlish EDH Cube
Hypercube! A New EDH Deck Every Week(ish)!
It would be interesting to hook this up to a website and let users rate the created cards to guide the algorithm trough the process, to see what it can come up with the assistance of humans.
I'm not sure if it would lead to any new intersting insights, since the human input would "taint" the process, but i think it would be cool to see what kind of cards would come out of it.
Spitting the output onto a website with a rating system would be awesome, but looping the results back to the system isn't necessary.
It's not like R&D looks at comments/ratings on Gatherer, either
EMERGENCY EDIT: At the time of this post there's six HUNDRED thirteen people viewing this thread!
Thats because I had the silly idea of posting the thread on reddit.
Anyway I find this fascinating as I've done programming of neural networks before but usually simple ones with back propagation.
I'll have to take a look at the source code once I'm less busy in RL.
It's a deep LSTM network trained, to my knowledge, using rounds of batched inputs followed by backpropagation.
And oh god, I didn't think about posting to reddit at all. I'm very bad at karma whoring. Thanks for stepping up to the plate! lol
Out of curiosity, what kind of "hardware" would be required to make a more robust network? What kind of parameters would you suggest if you had no hardware limits?
Well, I ran those experiments on a rather insubstantial dual-core machine with no usable GPU. In the lab, they're setting up a 64-core Intel Phi machine and I've been told that we'll be getting some Nividia Kepler GPUs. The math all boils down to relatively simple vector and matrix operations, and that can be done very efficiently with a good combination of CPUs and GPUs. The algorithms we're dealing with are extremely scalable, so the more computing power you can throw at it, the better.
As for the right number of layers and neurons? Hard to say: parameter optimization is still an open problem on that front. But more layers and thousands more neurons wouldn't hurt, and we can scale up or down from there. The neural architecture I used has over 7 million different weights that can be fine-tuned. The weights, of course, encode the notion of what Magic cards are, similar to how your brain encoded your 10th birthday party or the smell of fresh flowers. With wider and deeper networks, there's room to incorporate more complex logic, but this also means that the network takes much longer to train, and more input is usually needed.
This is all very interesting because neural nets aren't really my area of expertise. Traditionally, I use formal methods and logic to reason about complex systems. Unfortunately, that's a time consuming process that requires a lot of expertise. When I realized that these deep neural nets have the ability to learn very complex rules all on their own, I got very excited. It may be possible for us to use these kinds of algorithms for very difficult tasks. That is the direction that my dissertation research is heading in.
Talcos, I would love to see more examples of early-mid generation cards - got any more you can post?
I can make more available for laughs if you're interested. I saved copies of the network at different points in the training process, so I can provide output at any point from infancy to maturity. Of course, I'm a bit busy at the moment with my research work and the machines are all tied up doing things, so it may need to wait until later.
I, for one, can't wait to see how this absolute brilliance develops.
If Slidshocking Krow had trample and mountainhome, it might be a reasonable design with an interesting drawback, albeit one that's a relic of an earlier era of design.
Talcos, I would love to see more examples of early-mid generation cards - got any more you can post?
I can make more available for laughs if you're interested. I saved copies of the network at different points in the training process, so I can provide output at any point from infancy to maturity. Of course, I'm a bit busy at the moment with my research work and the machines are all tied up doing things, so it may need to wait until later.
So long as you deliver at some point, all will be forgiven
Talcos, I would love to see more examples of early-mid generation cards - got any more you can post?
I can make more available for laughs if you're interested. I saved copies of the network at different points in the training process, so I can provide output at any point from infancy to maturity. Of course, I'm a bit busy at the moment with my research work and the machines are all tied up doing things, so it may need to wait until later.
So long as you deliver at some point, all will be forgiven
Haha. Well, I should hope to do so. But in the unlikely event that I disappear, I have made sure to provide you with all of the information necessary to replicate my experiments for your own amusement.
I feel like this could make a fantastic new game mode, with two players trying to make a deck out of generated cards. Short periods of action, followed by long periods of trying to figure out what a card actually did, all interspersed with maniacal laughter.
I feel like this could make a fantastic new game mode, with two players trying to make a deck out of generated cards. Short periods of action, followed by long periods of trying to figure out what a card actually did, all interspersed with maniacal laughter.
Okay, since you asked, I have some new cards for you.
Here are some more samples from one hour into the training process:
---------
Ekemreress Doetlo 3G
Creature - Nightxolk
Flying X, T: Target creature or player, if it gets +1/+1 until end of turn.
Beash (You may put a goren creatures this combat.)
As Ekemreress Doetlo enters the battlefield, you may cast it fer exile Ekemreress Doetlo.
Draw a card
Daozfy 1B (Hound by frames a card in put all combat damage you control gains flying as long as you control with houdke until end of turn.)
1/3
Styryyroved Sturtiop 4B
Legendary Creature - Elf R: Gain lands you control:" Scary a Wall creature cae extep to each non-Herfolk.
0/1
Arnoftuee UUB
Instant
You may pay 2.
Teres Sarde 1GGB
Instant
Prevent the next end step.
Exile a charge counter on Teres Sarde.
Scold, you quiunt 2
Artifact - Equipment
Horpernt a spell's power is on the battlefield with two +1/+1 counters on it.
Suthas (Whenever Scold, you quiunt becomes a drawed a creature with flying onto the battlefield.)
---------
And then here are some from 5 hours into the training process:
---------
Handso 2U
Counter target spell. If you do, put a +1/+1 counter on Handso.
Inaader Cyclion 4B
Creature Creature - Spirit G, sacrifice a creature: You gain life equal to its converted mana cost 4 or less from your graveyard to your hand.
2/2
Hendtance Ibunder 1UU
Enchantment - Aura
Enchant creature
Enchanted creature gets +3/+3. R: Hendtance Ibunder gets +3/+1 and has protection from red only as a sorcery.
---------
And finally here are some from 23 hours into the training process:
---------
Khrat Sellglade 2WW
Creature - Angel
Vigilance
At the beginning of your upkeep, if there are no time counters on it, you may look at the top card of your library until you reveal a creature card from among them, then put that card on the bottom of your library in any order.
3/3
Legan of Echince 7GG
Creature - Treefolk Avatar
Flying
At the beginning of your upkeep, put a +1/+1 counter on target creature.
6/5
Sings of Junay 3RG
Instant - Arcane
Each player sacrifices a land. (Damage dealt by the creature as though it doesn't gain have double strike.)
Otal of Shymong 1BBB
Creature - Drake
Dodaro
Reach
A creature you control gains flying, flying.
6/6
#No idea what Dodaro is, must be a keyword ability.
Pirefila Shield 1B
Creature - Elgren
Haste
When Pirefila Shield leaves the battlefield, sacrifice it and each opponent discards a card.
1/1
Slethward Bestroh 2
Artifact W: Return target creature card from your graveyard to your hand.
Anaboth Cubblue 3
Artifact 2,T: Target land becomes a 4/4 white Spirit creature with flying until end of turn.
Roon War Medoma G
Instant
Name a card. You gain 1 life.
Rakile Volan B
Creature - Human Knight
Whenever Rakile Volan blocks or becomes blocked, you may pay 1B. If you do, you gain 2 life.
1/1
Rilatort Treat
Land T: Add WUWG to your mana pool.
#The RNN knows lands make mana, so this seems like a perfectly acceptable card to it.
Tingras Illza R
Creature - Human Wizard U, T: Target creature you control gets +3/+2 for each eldrue counter on Tingras Illza. Untap Tingras Illza.
1/1
#Lord have mercy this would be a broken card. Fortunately, it has no way of actually generating eldrue counters.
Blancho Ogure 4W
Creature - Spidin Druid
Creatures doesn't untap during your upkeep, if it's tied by creatures it's blocking as long as you do, for each creature with power 2 or less from his or her graveyard, where X is the divight converted mana cost X or less life, where X is the number of Jells on the battlefield tapped.
2/2
#Sometimes, even the fully trained RNN just spits out noise. But it only happens very rarely.
Yes. You could prime the network with restrictions such that it produces a card that follows those restrictions. A hand-crafted program can sit on top of the network and dictate what needs to happen.
Thanks! And not much, actually. The program that I used runs out of the box, but it does have some dependencies that you will need to install. Specifically, you'll need the Torch scientific computing framework. I had no problems installing it, but then again, it's the sort of thing I do every day. Now, to improve upon what I've done, we'll need to make some readjustments to the program, which does require some knowledge (or a whole hell of a lot of trial and error, take your pick).
Also, I'll warn you that I had to reorganize the data to allow for effective training. The cards in json format have most of the important information occurring at the end of the record. In my first experiment, I realized that the network interpreted the data to mean that the text in the body of the card determined the type, mana cost, and name, instead of the other way around.
Thanks! And yes, I agree, colorshifting comes as a consequence of underfitting.
It's best if I have all of them, the more the merrier, but I get decent results at the halfway point, which is about 7000 cards.
If anything though I feel we could do with more cards than actually exist in Magic in order to get better results, that is, we could mutate some of the inputs to expand the card pool. For example, take a creature, tweak its power, toughness, abilities, and cost ever so slightly, and then present that one as well. This is going to be necessary if we want to train a network to produce planeswalkers, since it has so few working examples. For example, take Jace, switch out some of his abilities, and resubmit it to the network.
Partitioning the network into subnetworks could work. But what's interesting is that state of the art approaches don't even require this anymore: with the right training regime, it efficiently self-organizes. Subregions of the network become attuned to different components (a creature region, an ability region, etc.). It's very difficult to tease out what does what (that takes a lot of reverse engineering), but it's obvious that that is what is happening.
A lot of that should come naturally with a deeper network. Right now the network is very parsimonious with its design rules because I deliberately kept everything as simple as I could.
My LinkedIn profile... thing (I have one of those now!).
My research team's webpage.
The mtg-rnn repo and the mtg-encode repo.
Green Angel tokens reminded me of a great fantasy series called Memory, Sorrow, and Thorn (there is an important landmark in the series called Green Angel Tower).
GWB Angel PodWBR Mardu MidrangeGWB Wilted Abzan
What do you think happened to Randy Bheuller?
EMERGENCY EDIT: At the time of this post there's six HUNDRED thirteen people viewing this thread!
Modern: Delver
Legacy: OmniTell
Commander: Too many
Everywhere it works: Storm.
http://markrosewater.tumblr.com/post/121187633243/are-you-allowed-to-look-at-cards-generated-by-a#notes
Avant Block: Avant -- Stormfront
I started giggling as soon as Amarogge Warfos let you put a 3/1 green solider token onto the battlefield, only to cruelly force you to put it directly into your graveyard, but yeah this combination of fake keywords absolutely slayed me.
Of the more recent cards, both Gravimite and Shring the Artist are dangerously close to being printable (albeit fairly powerful in each case, and Gravimite should lose the white hybrid).
Excited to see the next batch!
Thats because I had the silly idea of posting the thread on reddit.
Anyway I find this fascinating as I've done programming of neural networks before but usually simple ones with back propagation.
I'll have to take a look at the source code once I'm less busy in RL.
Are you designing commons? Check out my primer on NWO.
Interested in making a custom set? Check out my Set skeleton and archetype primer.
I also write articles about getting started with custom card creation.
Go and PLAYTEST your designs, you will learn more in a single playtests than a dozen discussions.
My custom sets:
Dreamscape
Coins of Mercalis [COMPLETE]
Exodus of Zendikar - ON HOLD
Yeah, same here. I think it's kind of like watching an orangutan trying to use a saw but holding it upside down. This AI works with basic pattern for Magic cards but beyond that has no idea what it's doing.
Currently Playing:
Legacy: Something U/W Controlish
EDH Cube
Hypercube! A New EDH Deck Every Week(ish)!
Spitting the output onto a website with a rating system would be awesome, but looping the results back to the system isn't necessary.
It's not like R&D looks at comments/ratings on Gatherer, either
It's a deep LSTM network trained, to my knowledge, using rounds of batched inputs followed by backpropagation.
And oh god, I didn't think about posting to reddit at all. I'm very bad at karma whoring. Thanks for stepping up to the plate! lol
Well, I ran those experiments on a rather insubstantial dual-core machine with no usable GPU. In the lab, they're setting up a 64-core Intel Phi machine and I've been told that we'll be getting some Nividia Kepler GPUs. The math all boils down to relatively simple vector and matrix operations, and that can be done very efficiently with a good combination of CPUs and GPUs. The algorithms we're dealing with are extremely scalable, so the more computing power you can throw at it, the better.
As for the right number of layers and neurons? Hard to say: parameter optimization is still an open problem on that front. But more layers and thousands more neurons wouldn't hurt, and we can scale up or down from there. The neural architecture I used has over 7 million different weights that can be fine-tuned. The weights, of course, encode the notion of what Magic cards are, similar to how your brain encoded your 10th birthday party or the smell of fresh flowers. With wider and deeper networks, there's room to incorporate more complex logic, but this also means that the network takes much longer to train, and more input is usually needed.
This is all very interesting because neural nets aren't really my area of expertise. Traditionally, I use formal methods and logic to reason about complex systems. Unfortunately, that's a time consuming process that requires a lot of expertise. When I realized that these deep neural nets have the ability to learn very complex rules all on their own, I got very excited. It may be possible for us to use these kinds of algorithms for very difficult tasks. That is the direction that my dissertation research is heading in.
My LinkedIn profile... thing (I have one of those now!).
My research team's webpage.
The mtg-rnn repo and the mtg-encode repo.
I can make more available for laughs if you're interested. I saved copies of the network at different points in the training process, so I can provide output at any point from infancy to maturity. Of course, I'm a bit busy at the moment with my research work and the machines are all tied up doing things, so it may need to wait until later.
My LinkedIn profile... thing (I have one of those now!).
My research team's webpage.
The mtg-rnn repo and the mtg-encode repo.
If Slidshocking Krow had trample and mountainhome, it might be a reasonable design with an interesting drawback, albeit one that's a relic of an earlier era of design.
I̟̥͍̠ͅn̩͉̣͍̬͚ͅ ̬̬͖t̯̹̞̺͖͓̯̤h̘͍̬e͙̯͈̖̼̮ ̭̬f̺̲̲̪i͙͉̟̩̰r̪̝͚͈̝̥͍̝̲s̼̻͇̘̳͔ͅt̲̺̳̗̜̪̙ ̳̺̥̻͚̗ͅm̜̜̟̰͈͓͎͇o̝̖̮̝͇m̯̻̞̼̫̗͓̤e̩̯̬̮̩n͎̱̪̲̹͖t͇̖s̰̮ͅ,̤̲͙̻̭̻̯̹̰ ̖t̫̙̺̯͖͚̯ͅh͙̯̦̳̗̰̟e͖̪͉̼̯ ̪͕g̞̣͔a̗̦t̬̬͓͙̫̖̭̻e̩̻̯ ̜̖̦̖̤̭͙̬t̞̹̥̪͎͉ͅo͕͚͍͇̲͇͓̺ ̭̬͙͈̣̻t͈͍͙͓̫̖͙̩h̪̬̖̙e̗͈ ̗̬̟̞̺̤͉̯ͅa̦̯͚̙̜̮f͉͙̲̣̞̼t̪̤̞̣͚e̲͉̳̥r͇̪̙͚͓l̥̞̞͎̹̯̹ͅi͓̬f̮̥̬̞͈ͅe͎ ̟̩̤̳̠̯̩̯o̮̘̲p̟͚̣̞͉͓e͍̩̣n͔̼͕͚̜e̬̱d̼̘͎̖̹͍̮̠,͖̺̭̱̮ ̣̲͖̬̪̭̥a̪͚n̟̲̝̤̤̞̗d̘̱̗͇̮͕̳͕͔ ͖̞͉͎t̹̙͎h̰̱͉̗e̪̞̱̝̹̩ͅ ̠̱̩̭̦p̯̙e͓o̳͚̰̯̺̱̰͔̘p̬͎̱̣̼̩͇l̗̟̖͚̠e̱͉͔̱̦̬̟̙ ̖͚̪͔̼̦w̺̖̤̱e͖̗̻̦͓̖̘̜r̭̥e͔̹̫̱͕̦̰͕ ̗͔̠p̠̗͍͍̱̳̠r̰͔͎̰o͉̥͓̰͚̥s̟͚̹̱͔̣t͉̙̳̖͖̪̮r̥̘̥͙̹a͉̟̫̟̳̠̟̭t͈̜̰͈͎e̞̣̭̲̬ ͚̗̯̟͙i͍͖̰̘̦͖͉ṇ̮̻̯̦̲̩͍ ̦̮͚̫̤t͉͖̫͕ͅͅh͙̮̻̘̣̮̼e͕̺ ͙l͕̠͎̰̥i̲͓͉̲g̫̳̟͈͇̖h̠̦̖t͓̯͎̗ ̳̪̘̟̙̩̦o̫̲f̙͔̰̙̠ ̹̪̗͇̯t͖̼̼͉͖̬h̹͇̩e͚̖̺̤͉̹͕̪ ͚͓̭̝̺G͎̗̯̩o̫̯̮̟̮̳̘d̜̲͙̠-̩̳̯̲̗̜P̹̘̥͉̝h͍͈̗̖̝ͅa͍̗̮̼̗r̜̖͇̙̺a̭̺͔̞̳͈o̪̣͓̯̬͙̯̰̗h̖̦͈̥̯͔.͇̣̙̝
So long as you deliver at some point, all will be forgiven
Can someone post renders of some of these cards with art and everything? I'd love to see Light of the Bild and its Angel token
Fuseback is genius!
I'm excited to see what the network will come up with after a couple more days of learning.
Haha. Well, I should hope to do so. But in the unlikely event that I disappear, I have made sure to provide you with all of the information necessary to replicate my experiments for your own amusement.
My LinkedIn profile... thing (I have one of those now!).
My research team's webpage.
The mtg-rnn repo and the mtg-encode repo.
Okay, since you asked, I have some new cards for you.
Here are some more samples from one hour into the training process:
---------
Ekemreress Doetlo
3G
Creature - Nightxolk
Flying
X, T: Target creature or player, if it gets +1/+1 until end of turn.
Beash (You may put a goren creatures this combat.)
As Ekemreress Doetlo enters the battlefield, you may cast it fer exile Ekemreress Doetlo.
Draw a card
Daozfy 1B (Hound by frames a card in put all combat damage you control gains flying as long as you control with houdke until end of turn.)
1/3
Styryyroved Sturtiop
4B
Legendary Creature - Elf
R: Gain lands you control:" Scary a Wall creature cae extep to each non-Herfolk.
0/1
Arnoftuee
UUB
Instant
You may pay 2.
Teres Sarde
1GGB
Instant
Prevent the next end step.
Exile a charge counter on Teres Sarde.
Scold, you quiunt
2
Artifact - Equipment
Horpernt a spell's power is on the battlefield with two +1/+1 counters on it.
Suthas (Whenever Scold, you quiunt becomes a drawed a creature with flying onto the battlefield.)
---------
And then here are some from 5 hours into the training process:
---------
Handso
2U
Counter target spell. If you do, put a +1/+1 counter on Handso.
Inaader Cyclion
4B
Creature Creature - Spirit
G, sacrifice a creature: You gain life equal to its converted mana cost 4 or less from your graveyard to your hand.
2/2
Hendtance Ibunder
1UU
Enchantment - Aura
Enchant creature
Enchanted creature gets +3/+3.
R: Hendtance Ibunder gets +3/+1 and has protection from red only as a sorcery.
---------
And finally here are some from 23 hours into the training process:
---------
Khrat Sellglade
2WW
Creature - Angel
Vigilance
At the beginning of your upkeep, if there are no time counters on it, you may look at the top card of your library until you reveal a creature card from among them, then put that card on the bottom of your library in any order.
3/3
Legan of Echince
7GG
Creature - Treefolk Avatar
Flying
At the beginning of your upkeep, put a +1/+1 counter on target creature.
6/5
Sings of Junay
3RG
Instant - Arcane
Each player sacrifices a land. (Damage dealt by the creature as though it doesn't gain have double strike.)
Otal of Shymong
1BBB
Creature - Drake
Dodaro
Reach
A creature you control gains flying, flying.
6/6
#No idea what Dodaro is, must be a keyword ability.
Pirefila Shield
1B
Creature - Elgren
Haste
When Pirefila Shield leaves the battlefield, sacrifice it and each opponent discards a card.
1/1
Slethward Bestroh
2
Artifact
W: Return target creature card from your graveyard to your hand.
Anaboth Cubblue
3
Artifact
2,T: Target land becomes a 4/4 white Spirit creature with flying until end of turn.
Roon War Medoma
G
Instant
Name a card. You gain 1 life.
Blood Rhast
1G
Creature - Goblin Shaman
3G: Blood Rhast gains "Sacrifice a land: Blood Rhast deals 3 damage to target player."
2/2
Rakile Volan
B
Creature - Human Knight
Whenever Rakile Volan blocks or becomes blocked, you may pay 1B. If you do, you gain 2 life.
1/1
Rilatort Treat
Land
T: Add WUWG to your mana pool.
#The RNN knows lands make mana, so this seems like a perfectly acceptable card to it.
Tingras Illza
R
Creature - Human Wizard
U, T: Target creature you control gets +3/+2 for each eldrue counter on Tingras Illza. Untap Tingras Illza.
1/1
#Lord have mercy this would be a broken card. Fortunately, it has no way of actually generating eldrue counters.
Blancho Ogure
4W
Creature - Spidin Druid
Creatures doesn't untap during your upkeep, if it's tied by creatures it's blocking as long as you do, for each creature with power 2 or less from his or her graveyard, where X is the divight converted mana cost X or less life, where X is the number of Jells on the battlefield tapped.
2/2
#Sometimes, even the fully trained RNN just spits out noise. But it only happens very rarely.
My LinkedIn profile... thing (I have one of those now!).
My research team's webpage.
The mtg-rnn repo and the mtg-encode repo.