Actually, a RoboRosewater for his writing style would have a decent chunk of source material to learn from. 700+ articles, 130+ podcast transcripts, and 40,000+ short answers on tumblr. I'm not sure how tough it would be to scrape the data, especially the more media-rich articles.
Do you (or anyone else) have any other features you'd like to see improved/implemented that original Gatherer doesn't have? I've been looking a bit at what it'd take to implement a "better Gatherer" and it seems pretty doable.
It'd be fun to include Word2Vec as a sort option. For example, you could ask "What's a Giant Growth variant from Zendikar block?" and it would show you Prey's Vengeance, Vines of Vastwood, etc.
EDIT: That would be useful for Commander deckbuilding too!
I took a look at your results, and they have some pretty funny stuff! What kind of corpus are you using though? I see @'s, but not much else that look like cards.
Is it possible to build the discriminator to accept bigger images than the generator?
One of the things that's nice about upsizing pixel art with waifu2x is that a place where the pixel art is wonky ends up getting resized into something that looks wonky as well. So you can use it effectively as a way to spot problems you're a little too close to the art to see. In other words, it tends to make bad parts more noticeable to the human eye. So hey, why not see if it makes them more noticable to a NN?
In principle, doing this should also decrease overfitting a little as well, since we're tweaking the output after processing in a way that the NN doesn't see, and make sure the discriminator is looking more at the content than the style of a piece.
By the way, I just realized that the implementation of the algorithm that I was using actually has the networks handling 96x96-sized images rather than 64x64 ones. Whoops. The Torch version of the code used 64x64-sized images.
Anyway, the size of image that the discriminator will accept is fixed. I can feed the script images of any size, but they get resized to 96x96 before being sent to the discriminator. I can change that, but it'd take several days to retrain the network to deal with the new image size.
And yeah, I can set up some experiments to test out your ideas later. I'll report on my findings.
I've been pondering whether applying the style network before the discriminator would help or hurt. On one hand, the generator+style combo would probably fool the discriminator earlier, which seems counterproductive. On the other hand, the descriminator might learn that it can't judge by style, and both would spend more effort on the content.
This isn't really helpful, but the Norseman reminds me of something out of Jim Hensen's creature shop. In particular, Farscape's Pilot. That's pretty exciting to me, because it mirrors human creativity, while still being quite novel for Magic art!
On another note, just how many values are in the vectors that describe these images?
EDIT: I'm also curious what happens if you look at how the evolutionary algorithm modifies the vectors, and instead go in the opposite direction. Does it become a mess, or create a viable alternative?
I came up with a mechanic the other day that has a similar "networking" feel as Recall, and I'm curious what everyone thinks of it. I was trying to find an Archer tribal mechanic:
Elvish Skyhunter 1G
Creature - Elf Archer (U)
Quiver 2 (When ~ enters the battlefield, you get 2 arrow counters) GT, 1 arrow counter: ~ deals 2 damage to target creature with flying.
Not an archers would deal damage. Many would have "trick arrows" that tap, force an attack, or remove keywords from a creature.
EDIT: I used the same parameters as before (a two layer network with 4 stacks), but I increased the dropout to 35%. That's up from like 20% or so. The highest you could conceivably raise the dropout would be to 50%; any more than that and neurons are more likely to be off than on, and that's not good. There's a sweet spot somewhere inbetween 0% and 50% where we get good benefits, and I have the feeling that I hit closer to that sweet spot.
This reminds me of some advice I've heard for game designers: If part your design is underpowered, your next iteration should have numbers that are clearly overpowered. That way you can quickly narrow down the proper range. If you try numbers that seem right, you might need several iterations before you find an iteration that's too powerful. (The reverse also applies to overpowered designs.)
There may be a flaw in applying that here though: Can you tell from a single data point if your dropout rate is too high or too low?
Actually, we're sorta already doing this with our mana symbols, right?
Right now, every mana symbol requires two characters so we can represent hybrid costs and such, so "WW" is W, "GW" is (G/W)
Hmm, converted mana costs could be easier to learn if colorless mana was encoded as "^^" instead of "^". That is, "&^^^^WW" would be 2W, not 4W. That might let the network generalize CMC by looking at the length of the cost. The networks seem to be doing fine as it is, but it could be worth experimenting with.
I'm not sure if we'd want to use "^^" for non-mana numbers though. It would be more consistent, but would expand the length of the average card even more.
I posted a loooong time ago, thinking changing how costs and abilities containing X in them could be parsed differently.
The idea I had was to seperate cards that contain X in the text into three different categories:
Cards that have X in its text, but not X in mana nor any ability cost. For example Death's Shadow.
The notation is very similar, but they function very differently, for the first category the meaning of X is always the cost, where in the second it is (always?) a static count of some sort.
My solution was to change the text of cards of the second category, replacing ever instance of X with Y; hoping that the network will be able to parse the categories better!
Currently I am testing the code on my laptop, the only problem is my hardware; it takes around 45 s for each batch =(
I did a very slight modification of the code of hardcast_sixdrop, I can share it if you are interested.
So, on this- I read the paper and I realised there's an important difference with our case, namely that M:tG is not a computer game so there's no "native" representation of the game state that can be learned.
If I can gloss over the details a bit for brevity here, what I mean is that M:tG being a physical game and not a computer game, it doesn't have a text- or graphics- based interface between its internal state and the outside world, or rather, a computer.
So what that means is that for an AI agent to interact with the game, someone needs to provide a representation of the game in a computer-readable format, be it text or graphics or whatever. And we know that such representations are a classic AI PITA.
Does it help that there are fan-made implementations of Magic? Some of them can handle 99% of all cards.
If you want to see what that format might look like, take a look at the Forge fan project.
Anybody want to start working on this?
Actually, a RoboRosewater for his writing style would have a decent chunk of source material to learn from. 700+ articles, 130+ podcast transcripts, and 40,000+ short answers on tumblr. I'm not sure how tough it would be to scrape the data, especially the more media-rich articles.
It'd be fun to include Word2Vec as a sort option. For example, you could ask "What's a Giant Growth variant from Zendikar block?" and it would show you Prey's Vengeance, Vines of Vastwood, etc.
EDIT: That would be useful for Commander deckbuilding too!
I've been pondering whether applying the style network before the discriminator would help or hurt. On one hand, the generator+style combo would probably fool the discriminator earlier, which seems counterproductive. On the other hand, the descriminator might learn that it can't judge by style, and both would spend more effort on the content.
On another note, just how many values are in the vectors that describe these images?
EDIT: I'm also curious what happens if you look at how the evolutionary algorithm modifies the vectors, and instead go in the opposite direction. Does it become a mess, or create a viable alternative?
EDIT: I just discovered that Google can help with this! Here's a bunch of card images that Google thinks look like "badart_2."
Here's how, in Chrome:
Elvish Skyhunter 1G
Creature - Elf Archer (U)
Quiver 2 (When ~ enters the battlefield, you get 2 arrow counters)
GT, 1 arrow counter: ~ deals 2 damage to target creature with flying.
Not an archers would deal damage. Many would have "trick arrows" that tap, force an attack, or remove keywords from a creature.
Thoughts?
Spire Spirit 3B
Creature - Spirit (Rare)
Whenever a creature enters the battlefield, if it was kicked, destroy target nonblack creature.
2/2
Kicker matters!
This reminds me of some advice I've heard for game designers: If part your design is underpowered, your next iteration should have numbers that are clearly overpowered. That way you can quickly narrow down the proper range. If you try numbers that seem right, you might need several iterations before you find an iteration that's too powerful. (The reverse also applies to overpowered designs.)
There may be a flaw in applying that here though: Can you tell from a single data point if your dropout rate is too high or too low?
And my suggestion is to use "^^BB" instead. That way, colorless mana uses two characters, just like colored mana.
Hmm, converted mana costs could be easier to learn if colorless mana was encoded as "^^" instead of "^". That is, "&^^^^WW" would be 2W, not 4W. That might let the network generalize CMC by looking at the length of the cost. The networks seem to be doing fine as it is, but it could be worth experimenting with.
I'm not sure if we'd want to use "^^" for non-mana numbers though. It would be more consistent, but would expand the length of the average card even more.
This sounds very similar to what jml34 was doing, so you might want to look at his work first. Here's how he split X, and here's his best result (click the third spoiler).
Does it help that there are fan-made implementations of Magic? Some of them can handle 99% of all cards.