I just finished running my first RNN draft with 7 of my friends last night. I made notes of what cards bogged the game down with either confusion of mechanics, broken/overpowered mechanics, or some things that were simply unfun.
It was kind of hilarious to see a player use the planeswalker that adjusts your life total to turn Crypt Party into a win-con (enchantment, at beginning of upkeep, if your life total is the same as number of cards in your graveyard, you win). At the very least, making it a nonsensical threat on the board was amusing to everybody.
Overall the draft was fun but has a lot of tweaking to do before it'll be ready for round 2. I intend to put in RNN lands and artifacts to help fill it out a bit more so that I can potentially get up to a 10 person draft (right now it clocks in at 7). Once I make a few adjustments over the next week or two I'll provide a card dump link and instructions to print them out.
I have about ~25 cards left to make before the first run of the draft is playable. I've printed out a little over half of the multi-colored creatures as a test-run to see how they turned out. The print-service at my local UPS Store can print 6 on a sheet of cardstock for $0.50 a sheet. If I take advantage of their 12x18 sheet for $0.80 a pop I could get a lot more cards that way.
I'm still here, but as sixdrop said, real-life is meddling pain in the ass. I am nearing a beta version of my cubedraft within the next week and will be posting it, but probably in a new thread.
Would 3-drop Chao be more fitting? lower the +3 to +2? Or do we still think +1 is the way to go? My concern, specifically for this draft format, is that milling 5 each turn could really hurt you unless you quickly take advantage of those minus abilities.
Haha, I had completely forgotten Sidisi was already a card name, whoops. Occasionally the network will turn out stuff like "JAVON" or "ELSPELL" which are clearly references to pre-existing cards, I just missed this one.
As far as it comes to balancing out the planeswalkers, I'm open to suggestions. Sidisi's +6 ability was legitimately +6, so I'm just kinda like alright network let's playtest it and see how awful of an idea that is." So drop Sidisi's starting loyalty to what, 3? 4? Any advice for nudging the numbers into something being a bit more balanced would be appreciated.
So I just finished making a set of Planeswalkers (one of each two-color combo) after whispering and training a few different networks. I'll be honest, it's garbage compared to the success I've had with creatures/non-creatures. The system seems to generate weird, nonsensical names for planeswalkers, and it generates them with only one or two abilities, and loyalty/mana cost is way off.
How I made it work: For whatever reason my networks make really good Legendary card names, so I generate a 100k character megadump and just search "legendary" through it for names. My latest network has some kind of obsession with Kamigawa and it generated a lot of far-east themed names (Or I translated them from Gknrnju to 'Genju').
I would whisper/seed the planeswalker dump by manacost of two different colors, and then just sift through a list of the generated abilities, mixing and matching ones I thought were thematic and would go together. Unfortunately this means that I've had to make an attempt at balancing out the starting loyalty and the loyalty +/- cost of abilities.
Does that make this less "authentic" having to manually craft so much information for the cards? Probably, but I just don't think we have enough variety in planeswalkers for the RNN network to figure it out itself. What I'm planning on doing for my draft is having all the planeswalkers face-down, you nab one as a bonus card and then start your drafting your packs. For people that have no idea what colors to play as, it's a way of encouraging them to try something and helps encourage equal attention to each color.
So the following cards "are" RNN network generated with the stipulation of extreme cherry-picking and mix-matching of abilities, and slight modification of number values for balancing (I'm sure they're still not balanced). I'm not 100% sold on some of them, but I think they're a good start.
Alright so this is a work in progress of my full cube draft, and in retrospect I really need to resize all of these to be smaller images. But for now, anybody who wants to take a peek at the creatures I've done, here's the set so far: https://www.dropbox.com/s/z1mvufjx3fesl2x/MagicboxRNN Creatures Beta.rar?dl=0
I'll be able to upload a set of the non-creature spells by the end of the night.
I'm probably going to reduce the stroke around the text from 3 pts to 2 pts so that it's a little more crisp. I've printed them out on card stock using a pretty high quality toner printer. They turn out a little darker than I'd want them to. I'll fiddle around with them before I go final-version and bring it to some printshop to actually give them the full-on treatment.
That being said, if you don't mind them being a little dark, they do print off as completely playable cards. I just use clear plastic sleeves and slip them in front of a junk common/land.
I had wondered about it too until I saw the Awaken mechanics from Battle for Zendikar. Awaken has you put +1/+1 counters on a land, THEN it becomes a creature (that is still a land). I suppose this works the same way, just one is an activated ability.
@Talcos, aw you are the man. This is just what I was hoping for. And thanks for the congratulations, I'm looking forward to getting married. To help celebrate I intend to release the draft for everybody else to enjoy as well. Over the next few days I'll host up an up-to-date version of my current pile of cards. Although incomplete, it gives you a good idea of what my RNN churns out.
Well for those of you who wanting to do a cube draft, I've still been working a lot at setting mine up; I am currently up to 227 cards, not including lands/artifacts.
I've been busy with a bunch of side projects but I'm trying to set this all up in time for my bachelor party. I'm trying to get an updated version of the corpus encoded in hardcast6drop's format, but for the life of me I can't seem to get it to output the file (my unfamiliarity with linux commands makes this tedious as all hell). If anybody has the encoded corpus that includes up to the new Innistrad set please let me know. Aside from that I do not have a complete collection of the cards posted anywhere but I've included a snapshot of my excel document I'm using to keep track of card count.
I was just being a bit silly and facetious my post, but in all seriousness I'm just excited with all the attention these networks are getting and being developed/researched further. Keep us posted on the magic art, I'm really interested to see how that goes. I've been really enjoying some of the re-imaginings with influence of other artists' styles.
Talcos in your madness have you actually created a robot that is capable of creating anime characters?
Combined with the proper feed/training of manga scans and episode synopsis you could have entirely RNN written plot/characters for a manga now. That is pretty ridiculous.
Derailing topic for just random thought: Personally, I was kind of intrigued about using this to teach AI for various videogames. Instead of having AI cheat to compensate for its lack of skill, we can give it much more organic characteristics by recording the movements and gameplay of a human behind the controls. For example, let's talk about Racing game AI and "rubber-banding". You're cutting corners and taking shortcuts, the opponents are falling behind... the computer doesn't know how to compensate for your clearly superior skill, so it increases the driving speed/handling to speed up your competition when you're not looking.
Instead, let's say we record a player on a level, and you take checkpoints and record the movement and behavior of the player for 100 races. Then you take the times for various checkpoints.
Slower checkpoints = put into training network for "Easy AI"
faster checkpoints = put into training network for "Hard AI"
Is there really too much of a risk of overtraining? You would just end up with a ghost-race at that point; the end result should be instructions on how to be a good, organic racer instead of a clunky robot trying to Race By Math.
But back to the topic of magic-cards, couldn't we use something like this to make better AI for magic the gathering pc games? You record however many thousands of games and the actions taken, it's got to be able to figure out some sort of connection of "how to play a deck" and how to recognize/identify and respond to a threat. I know I may be giving the RNN too much credit but there is a very weird logic hidden in there behind all of this, and I keep getting surprised by these results you come up with.
Well it's been awhile and I've still been churning out art for cards. Last time I did a batch I know some people were disappointed that I hadn't done any white creatures yet. Here's a batch of white creatures generated from Magicbox (my RNN).
Also worth note: Wall of Denium is the first legendary wall! Woo!
It's actually kind of funny but I was generating quite a few manlands before they were announced in the battle for zendikar set. one was a "Goblin factory" that came into play tapped, could tap for 1 or pay 2 to turn it into a 3/3 creature with trample.
Outside of this, I've had a few creatures actually be generated with a land type in their name. If you check my Green Creatures dump from the previous page you'll see that Guardian Sorya actually has Forest in the creature type. This means you could totally use Skyshroud Claim to put them into play.
Holy carp, are those cards 100% unedited from your RNN??
No I've probably posted most of these earlier in the thread, but I've had to reword a lot of them to make them legal or sensable.
I have changed the casting cost of a few just to nudge it in the right direction (a LOT of cards were 2 drops initially).
It's been a pretty time consuming process to set each template up in photoshop (each file is like 200mb, so they're all very high-res) but I'm going to be moving onto White and Colorless/Multi-colored pretty soon. I just wander around online for art and I kept finding a lot of red-themed ones that fit the bill for the red cards I had.
Also my training is pretty basic, I've adjusted it only very lightly over the past few months, but I've probably sifted through 10 500,000 character dumps (like 60 pages long each): "th train.lua -data_dir data -gpuid -1 -rnn_size 600 -num_layers 3 -dropout .5 -init_from data/input.tx"
The key thing though is all of my cards are taken under epoch 11. Anything around epochs 20 and 30 are just too vanilla; you get way more interesting results at lower epochs, even if you have to decipher a lot of gibberish and change the wording to make it legal. Epochs 9-11 have been my sweet spot.
It was kind of hilarious to see a player use the planeswalker that adjusts your life total to turn Crypt Party into a win-con (enchantment, at beginning of upkeep, if your life total is the same as number of cards in your graveyard, you win). At the very least, making it a nonsensical threat on the board was amusing to everybody.
Overall the draft was fun but has a lot of tweaking to do before it'll be ready for round 2. I intend to put in RNN lands and artifacts to help fill it out a bit more so that I can potentially get up to a 10 person draft (right now it clocks in at 7). Once I make a few adjustments over the next week or two I'll provide a card dump link and instructions to print them out.
I have about ~25 cards left to make before the first run of the draft is playable. I've printed out a little over half of the multi-colored creatures as a test-run to see how they turned out. The print-service at my local UPS Store can print 6 on a sheet of cardstock for $0.50 a sheet. If I take advantage of their 12x18 sheet for $0.80 a pop I could get a lot more cards that way.
Here's a preview of how they turned out
As far as it comes to balancing out the planeswalkers, I'm open to suggestions. Sidisi's +6 ability was legitimately +6, so I'm just kinda like alright network let's playtest it and see how awful of an idea that is." So drop Sidisi's starting loyalty to what, 3? 4? Any advice for nudging the numbers into something being a bit more balanced would be appreciated.
How I made it work: For whatever reason my networks make really good Legendary card names, so I generate a 100k character megadump and just search "legendary" through it for names. My latest network has some kind of obsession with Kamigawa and it generated a lot of far-east themed names (Or I translated them from Gknrnju to 'Genju').
I would whisper/seed the planeswalker dump by manacost of two different colors, and then just sift through a list of the generated abilities, mixing and matching ones I thought were thematic and would go together. Unfortunately this means that I've had to make an attempt at balancing out the starting loyalty and the loyalty +/- cost of abilities.
Does that make this less "authentic" having to manually craft so much information for the cards? Probably, but I just don't think we have enough variety in planeswalkers for the RNN network to figure it out itself. What I'm planning on doing for my draft is having all the planeswalkers face-down, you nab one as a bonus card and then start your drafting your packs. For people that have no idea what colors to play as, it's a way of encouraging them to try something and helps encourage equal attention to each color.
So the following cards "are" RNN network generated with the stipulation of extreme cherry-picking and mix-matching of abilities, and slight modification of number values for balancing (I'm sure they're still not balanced). I'm not 100% sold on some of them, but I think they're a good start.
https://www.dropbox.com/s/z1mvufjx3fesl2x/MagicboxRNN Creatures Beta.rar?dl=0
I'll be able to upload a set of the non-creature spells by the end of the night.
Edit: Non-creature spells are on this download here below:
https://www.dropbox.com/s/j82sxdusov9cq6o/MagicboxRNN Non-Creatures Beta.rar?dl=0
I'm probably going to reduce the stroke around the text from 3 pts to 2 pts so that it's a little more crisp. I've printed them out on card stock using a pretty high quality toner printer. They turn out a little darker than I'd want them to. I'll fiddle around with them before I go final-version and bring it to some printshop to actually give them the full-on treatment.
That being said, if you don't mind them being a little dark, they do print off as completely playable cards. I just use clear plastic sleeves and slip them in front of a junk common/land.
I had wondered about it too until I saw the Awaken mechanics from Battle for Zendikar. Awaken has you put +1/+1 counters on a land, THEN it becomes a creature (that is still a land). I suppose this works the same way, just one is an activated ability.
@Talcos, aw you are the man. This is just what I was hoping for. And thanks for the congratulations, I'm looking forward to getting married. To help celebrate I intend to release the draft for everybody else to enjoy as well. Over the next few days I'll host up an up-to-date version of my current pile of cards. Although incomplete, it gives you a good idea of what my RNN churns out.
Well for those of you who wanting to do a cube draft, I've still been working a lot at setting mine up; I am currently up to 227 cards, not including lands/artifacts.
I've been busy with a bunch of side projects but I'm trying to set this all up in time for my bachelor party. I'm trying to get an updated version of the corpus encoded in hardcast6drop's format, but for the life of me I can't seem to get it to output the file (my unfamiliarity with linux commands makes this tedious as all hell). If anybody has the encoded corpus that includes up to the new Innistrad set please let me know. Aside from that I do not have a complete collection of the cards posted anywhere but I've included a snapshot of my excel document I'm using to keep track of card count.
Combined with the proper feed/training of manga scans and episode synopsis you could have entirely RNN written plot/characters for a manga now. That is pretty ridiculous.
Derailing topic for just random thought: Personally, I was kind of intrigued about using this to teach AI for various videogames. Instead of having AI cheat to compensate for its lack of skill, we can give it much more organic characteristics by recording the movements and gameplay of a human behind the controls. For example, let's talk about Racing game AI and "rubber-banding". You're cutting corners and taking shortcuts, the opponents are falling behind... the computer doesn't know how to compensate for your clearly superior skill, so it increases the driving speed/handling to speed up your competition when you're not looking.
Instead, let's say we record a player on a level, and you take checkpoints and record the movement and behavior of the player for 100 races. Then you take the times for various checkpoints.
Slower checkpoints = put into training network for "Easy AI"
faster checkpoints = put into training network for "Hard AI"
Is there really too much of a risk of overtraining? You would just end up with a ghost-race at that point; the end result should be instructions on how to be a good, organic racer instead of a clunky robot trying to Race By Math.
But back to the topic of magic-cards, couldn't we use something like this to make better AI for magic the gathering pc games? You record however many thousands of games and the actions taken, it's got to be able to figure out some sort of connection of "how to play a deck" and how to recognize/identify and respond to a threat. I know I may be giving the RNN too much credit but there is a very weird logic hidden in there behind all of this, and I keep getting surprised by these results you come up with.
Also worth note: Wall of Denium is the first legendary wall! Woo!
Outside of this, I've had a few creatures actually be generated with a land type in their name. If you check my Green Creatures dump from the previous page you'll see that Guardian Sorya actually has Forest in the creature type. This means you could totally use Skyshroud Claim to put them into play.
No I've probably posted most of these earlier in the thread, but I've had to reword a lot of them to make them legal or sensable.
I have changed the casting cost of a few just to nudge it in the right direction (a LOT of cards were 2 drops initially).
It's been a pretty time consuming process to set each template up in photoshop (each file is like 200mb, so they're all very high-res) but I'm going to be moving onto White and Colorless/Multi-colored pretty soon. I just wander around online for art and I kept finding a lot of red-themed ones that fit the bill for the red cards I had.
Also my training is pretty basic, I've adjusted it only very lightly over the past few months, but I've probably sifted through 10 500,000 character dumps (like 60 pages long each): "th train.lua -data_dir data -gpuid -1 -rnn_size 600 -num_layers 3 -dropout .5 -init_from data/input.tx"
The key thing though is all of my cards are taken under epoch 11. Anything around epochs 20 and 30 are just too vanilla; you get way more interesting results at lower epochs, even if you have to decipher a lot of gibberish and change the wording to make it legal. Epochs 9-11 have been my sweet spot.