Random Deck Generation Ideas

[austinio7116]

I have been making some small improvements to the deck generation features. I wanted to create a thread to record some of the other ideas that have been suggested. So far we have:

Add a "DeckDont" DeckHint so that cards can be excluded from coloured decks (Hanmac)

Allow custom formats/sets to be selected for the card pool.

Feel free to add ideas to this thread.

[Hanmac]

An Example where i would like such a DeckDont:

Stern Judge should only added to a White deck, if the Deck does not have any Black stuff (and thats why no Swamps)

"Ability$Counters" for DeckHas and DeckHints & DeckNeeds should probably extended to specify what kind of Counters it has, (like +1/+1 or -1/-1 ones)

[austinio7116]

I'd really like to see a neural network trained to create format legal decks - there are plenty of decks to train it on.

[austinio7116]

I may try deeplearning4j and see how far I get when I have some time

[Xitax]

That sounds very interesting. I'd be interested in how it turns out if you do.

[austinio7116]

Got a bit sidetracked with a more heuristic approach that I had been thinking about for a while. I have built a new deck generator that takes a large set of deck lists (currently all AI playable standard deck from Aether Revolt from Agetian's deck extract) and builds a matrix of associations between cards in the set. It's a very simple loop though all the cards counting for each other card how many times it appears in decks with it. Then the deck generator uses this matrix to get a small pool of cards based on a single source card that have high associations to that card and builds a deck (using a similar model to the limited deck generator) using the pool. This generates pseudo-random decks that are generally fairly synergistic and quite similar to human generated decks.

[austinio7116]

A few example decks:


Ruins of Oran-Rief based deck

Main:
4 Thought-Knot Seer
4 Matter Reshaper
3 Reality Smasher
4 Bearer of Silence
2 Walking Ballista
1 Metallic Mimic
1 Wasteland Strangler
1 Foundry Inspector
4 Transgress the Mind
4 Grasp of Darkness
4 Fatal Push
2 Aethersphere Harvester
1 Inspiring Statuary
4 Ruins of Oran-Rief
4 Aether Hub
4 Sea Gate Wreckage
4 Spire of Industry
4 Corrupted Crossroads
1 Swamp
2 Swamp
1 Swamp
1 Swamp

Sideboard:
4 Scrapheap Scrounger
4 Wastes
4 Concealed Courtyard


Goblin Dark-Dwellers based deck

Main:
4 Liliana, the Last Hope
4 Chandra, Torch of Defiance
3 Goblin Dark-Dwellers
4 Kalitas, Traitor of Ghet
4 Yahenni, Undying Partisan
2 Noxious Gearhulk
4 Grasp of Darkness
4 Ruinous Path
4 Unlicensed Disintegration
1 Shock
1 Fatal Push
4 Smoldering Marsh
4 Evolving Wilds
4 Foreboding Ruins
4 Aether Hub
1 Swamp
3 Swamp
1 Swamp
2 Mountain
2 Mountain

Sideboard:
3 Tireless Tracker
4 Scrapheap Scrounger
4 Bomat Courier
4 Grim Flayer

Galvanic Bombardment based deck

Main:
4 Chandra, Flamecaller
4 Thing in the Ice
2 Baral, Chief of Compliance
4 Trophy Mage
4 Galvanic Bombardment
4 Tormenting Voice
4 Harnessed Lightning
4 Cathartic Reunion
2 Take Inventory
1 Fiery Temper
1 Negate
1 Dynavolt Tower
4 Spirebluff Canal
4 Aether Hub
4 Wandering Fumarole
4 Highland Lake
4 Drownyard Temple
1 Island
1 Island
2 Mountain
1 Mountain

Sideboard:
3 Wastes
4 Woodland Stream
4 Tranquil Expanse
4 Timber Gorge

[Hanmac]

I don't know if the AI should use my decks as inspiration for deck building, they are not that good xD

[Agetian]

Sounds very interesting! Can't wait to see this used in practice! It can help create literally infinite AI opponent decks that are better than the current "fully randomly generated" decks!

- Agetian

[austinio7116]

Just a couple of bugs to sort out, plus a bit of refactoring to avoid having to generate the distance matrix on every startup and I'll check it in so you can try it.

[austinio7116]

I think this is ready to check in now. There are some things I can improve but it's a decent first version. I added some randomization so the decks for each card are different each time, ran a Modern format version of the training data to add to the Standard data and fixed several bugs. You can select the "seed" card by name in the deck selector and all the training data is now serialized and stored in the resources folders so no need to retrain on each app initialization.

[austinio7116]

Checked in SVN 33988

[Agetian]

This is really awesome! I'll play with this today after work. Thanks a lot for the effort!

- Agetian

[austinio7116]

I hope you like it - make sure to get at least svn revision 33995 for the best experience.