DraftZero FDN sealed deck builder v1 (docs/032): self-play ensembles, card order, results, model card

FDN_cardnames.txt

@@ -0,0 +1,286 @@
+Abrade
+Abyssal Harvester
+Adventuring Gear
+Aegis Turtle
+Aetherize
+Affectionate Indrik
+Ajani's Pridemate
+Ajani, Caller of the Pride
+Akroma's Memorial
+Alesha, Who Laughs at Fate
+Ambush Wolf
+An Offer You Can't Refuse
+Angel of Finality
+Anthem of Champions
+Apothecary Stomper
+Arahbo, the First Fang
+Arbiter of Woe
+Arcane Epiphany
+Archmage of Runes
+Armasaur Guide
+Ashroot Animist
+Authority of the Consuls
+Axgard Cavalry
+Bake into a Pie
+Balmor, Battlemage Captain
+Banishing Light
+Banner of Kinship
+Battlesong Berserker
+Beast-Kin Ranger
+Bigfin Bouncer
+Billowing Shriekmass
+Bite Down
+Blanchwood Armor
+Blasphemous Edict
+Bloodfell Caves
+Bloodthirsty Conqueror
+Bloom Tender
+Blossoming Sands
+Boltwave
+Brass's Bounty
+Brazen Scourge
+Brineborn Cutthroat
+Broken Wings
+Bulk Up
+Burglar Rat
+Burnished Hart
+Burst Lightning
+Bushwhack
+Cackling Prowler
+Campus Guide
+Cat Collector
+Cathar Commando
+Celestial Armor
+Cephalid Inkmage
+Chandra, Flameshaper
+Claws Out
+Clinquant Skymage
+Condemn
+Consuming Aberration
+Courageous Goblin
+Crackling Cyclops
+Crypt Feaster
+Crystal Barricade
+Curator of Destinies
+Dauntless Veteran
+Day of Judgment
+Dazzling Angel
+Diregraf Ghoul
+Dismal Backwater
+Divine Resilience
+Doubling Season
+Dragon Trainer
+Drake Hatcher
+Drakuseth, Maw of Flames
+Dreadwing Scavenger
+Dwynen's Elite
+Dwynen, Gilt-Leaf Daen
+Eager Trufflesnout
+Eaten Alive
+Electroduplicate
+Elementalist Adept
+Elenda, Saint of Dusk
+Elfsworn Giant
+Elvish Archdruid
+Elvish Regrower
+Embercleave
+Empyrean Eagle
+Erudite Wizard
+Essence Scatter
+Etali, Primal Storm
+Evolving Wilds
+Exemplar of Light
+Exsanguinate
+Extravagant Replication
+Faebloom Trick
+Fake Your Own Death
+Fanatical Firebrand
+Felidar Savior
+Felling Blow
+Fiend Artisan
+Fiendish Panda
+Fiery Annihilation
+Firebrand Archer
+Firespitter Whelp
+Fishing Pole
+Flamewake Phoenix
+Fleeting Distraction
+Fleeting Flight
+Forest
+Frenzied Goblin
+Garruk's Uprising
+Genesis Wave
+Ghalta, Primal Hunger
+Giada, Font of Hope
+Giant Growth
+Gleaming Barrier
+Gnarlid Colony
+Goblin Boarders
+Goblin Bushwhacker
+Goblin Negotiation
+Goblin Surprise
+Goldvein Pick
+Good-Fortune Unicorn
+Gorehorn Raider
+Grappling Kraken
+Grim Tutor
+Grow from the Ashes
+Guarded Heir
+Gutless Plunderer
+Hare Apparent
+Healer's Hawk
+Heartfire Immolator
+Helpful Hunter
+Herald of Eternal Dawn
+Heraldic Banner
+Hero's Downfall
+Heroic Reinforcements
+Hidetsugu's Second Rite
+High Fae Trickster
+High-Society Hunter
+Homunculus Horde
+Hungry Ghoul
+Icewind Elemental
+Imprisoned in the Moon
+Incinerating Blast
+Infernal Vessel
+Infestation Sage
+Inspiration from Beyond
+Inspiring Call
+Inspiring Paladin
+Involuntary Employment
+Island
+Joust Through
+Juggernaut
+Jungle Hollow
+Kaito, Cunning Infiltrator
+Kellan, Planar Trailblazer
+Kiora, the Rising Tide
+Koma, World-Eater
+Krenko, Mob Boss
+Kykar, Zephyr Awakener
+Lathril, Blade of the Elves
+Leyline Axe
+Lightshell Duo
+Liliana, Dreadhorde General
+Llanowar Elves
+Loot, Exuberant Explorer
+Luminous Rebuke
+Lunar Insight
+Macabre Waltz
+Make Your Move
+Marauding Blight-Priest
+Meteor Golem
+Micromancer
+Midnight Snack
+Mild-Mannered Librarian
+Mischievous Mystic
+Mischievous Pup
+Mocking Sprite
+Mossborn Hydra
+Mountain
+Muldrotha, the Gravetide
+Needletooth Pack
+Nessian Hornbeetle
+Nine-Lives Familiar
+Niv-Mizzet, Visionary
+Omniscience
+Overrun
+Painful Quandary
+Paradise Druid
+Perforating Artist
+Phyrexian Arena
+Pilfer
+Plains
+Preposterous Proportions
+Prideful Parent
+Progenitus
+Quakestrider Ceratops
+Quick-Draw Katana
+Quilled Greatwurm
+Raise the Past
+Ravenous Amulet
+Reassembling Skeleton
+Reclamation Sage
+Refute
+Resolute Reinforcements
+Revenge of the Rats
+Rise of the Dark Realms
+Rite of the Dragoncaller
+Rogue's Passage
+Ruby, Daring Tracker
+Rugged Highlands
+Run Away Together
+Rune-Scarred Demon
+Rune-Sealed Wall
+Sanguine Syphoner
+Savannah Lions
+Scavenging Ooze
+Scoured Barrens
+Scrawling Crawler
+Searslicer Goblin
+Secluded Courtyard
+Seeker's Folly
+Seismic Rupture
+Self-Reflection
+Serra Angel
+Shivan Dragon
+Sire of Seven Deaths
+Skyknight Squire
+Skyship Buccaneer
+Slagstorm
+Slumbering Cerberus
+Snakeskin Veil
+Solemn Simulacrum
+Soul-Shackled Zombie
+Soulstone Sanctuary
+Sower of Chaos
+Spectral Sailor
+Sphinx of Forgotten Lore
+Sphinx's Tutelage
+Spinner of Souls
+Spitfire Lagac
+Squad Rallier
+Stab
+Strix Lookout
+Stroke of Midnight
+Stromkirk Bloodthief
+Strongbox Raider
+Sun-Blessed Healer
+Sure Strike
+Swamp
+Swiftblade Vindicator
+Swiftfoot Boots
+Swiftwater Cliffs
+Sylvan Scavenging
+Tatyova, Benthic Druid
+Temporal Manipulation
+Think Twice
+Thornwood Falls
+Thousand-Year Storm
+Thrill of Possibility
+Time Stop
+Tinybones, Bauble Burglar
+Tolarian Terror
+Tragic Banshee
+Tranquil Cove
+Treetop Snarespinner
+Twinblade Blessing
+Twinflame Tyrant
+Uncharted Voyage
+Valkyrie's Call
+Vampire Gourmand
+Vampire Nighthawk
+Vampire Soulcaller
+Vanguard Seraph
+Vengeful Bloodwitch
+Vivien Reid
+Wardens of the Cycle
+Wary Thespian
+Wildwood Scourge
+Wind-Scarred Crag
+Witness Protection
+Youthful Valkyrie
+Zimone, Paradox Sculptor
+Zombify
+Zul Ashur, Lich Lord

README.md

@@ -0,0 +1,116 @@
+---
+license: cc-by-4.0
+tags:
+- magic-the-gathering
+- deck-building
+- self-play
+- sealed
+- reinforcement-learning
+---
+
+# DraftZero: an FDN sealed deck builder learned from self-play
+
+A deck builder for Magic: The Gathering **Foundations (FDN) sealed**, trained **only from games it played**, never from
+human decklists. Given a sealed pool (about 81 non-basic cards from six packs), it builds a 40-card deck.
+
+- **Report:** [docs/032](https://github.com/danieljbrooks/draft-zero/blob/main/docs/032-sealed-deckbuilding.md)
+  (experiment #5 of [DraftZero](https://github.com/danieljbrooks/draft-zero)).
+- **Code:** [`danieljbrooks/draft-zero` @ tag `sealed-builder-v1`](https://github.com/danieljbrooks/draft-zero/tree/sealed-builder-v1),
+  `gorge/sealed/`.
+- **Engine:** [gorge](https://github.com/adams-shaun/gorge), a Go rules engine, with DraftZero's opt-in patches
+  (`gorge/patches/`).
+
+## What it learned
+
+Trained only on its own games, the builder learns to build **simple, white-based creature decks that the bot playing
+them can pilot well**: white in about 85% of its decks, white-green in 42% (human players: 17%), fewer tricks and
+card-advantage spells than people play. More search in its training games made its decks steadily more human-like.
+It is a deck builder *for its pilot*, gorge's search with DraftZero's network, not a model of how strong humans build.
+
+## Results
+
+Win rate of the builder's decks against the decks humans built from the same pools (17lands data), with both decks
+played by gorge's search at 100 simulations. One standard error is about 1 point at 2,000 games.
+
+| Builder | Sealed test pools (1,000) | Draft test pools (1,000; never trained on) | Cards shared with the human deck (of 40, sealed) |
+|---|---|---|---|
+| **`sealed_builder_phaseC_gen1.pt`** (the final model) | **63.8%** of 2,000 | | 19.9 |
+| `sealed_builder_phaseB_gen4.pt` | **63.8%** of 2,000 | **57.5%** of 2,000 | 19.2 |
+| 17lands heuristic (best two colours by games-in-hand win rate) | 55.2% | 50.9% | 20.5 |
+| A builder trained to imitate the human decks | 53.0% | 52.7% | 23.6 |
+
+All numbers are in `results.json`. The win rate measures how well a deck suits this pilot: the human decks' strengths
+(tricks, card advantage, planeswalkers) need play the pilot doesn't have.
+
+## Files
+
+| File | What |
+|---|---|
+| `sealed_builder_phaseC_gen1.pt` | **The final model:** an ensemble of 5 deck value networks, trained on self-play games at 100 simulations (phase C, generation 1) |
+| `sealed_builder_phaseB_gen4.pt` | The phase B generation 4 ensemble (games at 25 simulations), the other builder scored on the test pools |
+| `FDN_cardnames.txt` | The 286 FDN card names, in the order of the networks' count vectors (17lands' column order) |
+| `example_pool.json` | One sealed pool (17lands val pool 0) to try it on |
+| `results.json` | The evaluation numbers above, with standard errors |
+| `SHA256SUMS` | Checksums of every file |
+
+## The model
+
+Each ensemble file holds five **deck value networks** (about 26,000 parameters each) and a little metadata (the land
+counts the training games covered; whether castability is on). A network scores a whole deck; it never outputs one.
+
+- **Input:** a deck as a count vector over FDN's 286 cards (basic lands included), plus a fixed table of 76 printed
+  facts per card (mana cost, types, power and toughness, keywords, rarity, mechanics tags, colours a land makes),
+  stored inside the file. Each card also has learned identity parameters (an embedding and corrections), which start
+  at zero.
+- **Output:** a deck strength s(deck). Trained as Bradley-Terry on game results:
+  P(deck A beats deck B) = sigmoid(s(A) − s(B) + β · [A played first]).
+- **Parts:** an additive card-quality term weighted by how castable the deck's mana makes each card; a structure MLP
+  over deck totals built from learnable card roles (land-ness, effective mana value, creature-ness, colours made); a
+  DeepSets context term. Details in the report, §4.
+
+**Building a deck** is a search against the ensemble's mean score:
+1. Score a quick deck for each colour combination.
+2. Refine the best four by one-for-one swaps until no swap helps.
+3. Keep the best.
+
+## Use
+
+```bash
+git clone --branch sealed-builder-v1 https://github.com/danieljbrooks/draft-zero.git
+cd draft-zero && pip install numpy torch
+huggingface-cli download danbrooks/draftzero-fdn-sealed-builder --local-dir hf-sealed
+```
+
+```python
+import json, sys
+import numpy as np
+sys.path.insert(0, "gorge")
+from sealed import model as M, build as B, decks as D
+
+ens = M.Ensemble.load("hf-sealed/sealed_builder_phaseC_gen1.pt")
+pool = D.C.vec(json.load(open("hf-sealed/example_pool.json"))["pool"])    # {card name: copies}, non-basic cards
+quality = np.mean([m.quality().detach().numpy() for m in ens.members], 0)
+deck = B.build(pool, B.Objective(ens, "mean"), quality, lands=ens.meta.get("lands"))
+print(D.C.dict(deck))                                                       # {card name: copies}, 40 cards
+```
+
+Building takes about a second per pool on a laptop CPU. FDN only: the networks know FDN's 286 cards.
+
+## Training
+
+- **Data:** the networks never saw a human deck. Self-play games in gorge between decks the builder made from 6,240
+  17lands sealed pools (pool contents only): about 280,000 games in the final model's lineage, out of 1.1 million in
+  the project.
+- **Curriculum:**
+  - 3 generations of games played by gorge's heuristic bot, the first on random decks;
+  - 7 of games played by the search at 25 simulations;
+  - 2 at 100 simulations.
+- **Each generation:** about 800 pools × 5 exploratory decks, 10,000-40,000 games.
+- **Compute:** a laptop, a shared 2× RTX PRO 6000 box (r1, courtesy of Dama) and community RunPod pods; about $9 of
+  cloud compute.
+
+## Licence and data
+
+The weights are released under CC BY 4.0. Pools and human decks come from
+[17lands' public datasets](https://www.17lands.com/public_datasets) (CC BY 4.0); 17lands does not endorse this project.
+Card facts come from XMage (MIT) and 17lands' card table. No Forge card scripts (GPL-3.0) are included.

SHA256SUMS

@@ -0,0 +1,6 @@
+9b6fc635736aa515bd66d1a4e2e57b42cbf6cbad13aa72a1baa7c197892d8dc0  README.md
+30bd8a108bf821166341b3cf1e70f6a4ee16037489aecf51912d6acb8d9124c5  FDN_cardnames.txt
+9242d33d26f782323d440b1b630ffef0944454c4929609aa495ecb927dc93482  example_pool.json
+4ac4a55def98ad4ab270c2079004c20d6fd68072abb180671496f93f1217254c  results.json
+66aecc3323487fd151f766188f21b22c123a5cccd71839ee65b21663f4fb4b85  sealed_builder_phaseB_gen4.pt
+58aa9aad59992326bd697176306685034121e8aa52a5eebef7f1eda4a7eca736  sealed_builder_phaseC_gen1.pt

example_pool.json

@@ -0,0 +1,73 @@
+{
+ "source": "17lands FDN sealed, val pool 0 (CC BY 4.0)",
+ "pool": {
+  "Apothecary Stomper": 1,
+  "Armasaur Guide": 1,
+  "Axgard Cavalry": 1,
+  "Bake into a Pie": 1,
+  "Banishing Light": 1,
+  "Bite Down": 1,
+  "Boltwave": 1,
+  "Brineborn Cutthroat": 1,
+  "Cackling Prowler": 2,
+  "Campus Guide": 2,
+  "Cat Collector": 1,
+  "Claws Out": 1,
+  "Consuming Aberration": 1,
+  "Courageous Goblin": 1,
+  "Crackling Cyclops": 1,
+  "Crypt Feaster": 2,
+  "Dauntless Veteran": 1,
+  "Day of Judgment": 1,
+  "Dazzling Angel": 1,
+  "Diregraf Ghoul": 1,
+  "Dismal Backwater": 1,
+  "Dragon Trainer": 2,
+  "Eaten Alive": 2,
+  "Elementalist Adept": 1,
+  "Elfsworn Giant": 1,
+  "Empyrean Eagle": 1,
+  "Erudite Wizard": 1,
+  "Evolving Wilds": 2,
+  "Exsanguinate": 1,
+  "Faebloom Trick": 1,
+  "Fanatical Firebrand": 1,
+  "Firespitter Whelp": 2,
+  "Fleeting Flight": 3,
+  "Giada, Font of Hope": 1,
+  "Gnarlid Colony": 1,
+  "Goblin Negotiation": 1,
+  "Good-Fortune Unicorn": 1,
+  "Gorehorn Raider": 1,
+  "Hero's Downfall": 1,
+  "Incinerating Blast": 1,
+  "Inspiring Call": 1,
+  "Inspiring Paladin": 1,
+  "Macabre Waltz": 1,
+  "Make Your Move": 1,
+  "Meteor Golem": 1,
+  "Midnight Snack": 1,
+  "Mocking Sprite": 2,
+  "Pilfer": 1,
+  "Prideful Parent": 1,
+  "Rise of the Dark Realms": 1,
+  "Ruby, Daring Tracker": 1,
+  "Rugged Highlands": 1,
+  "Scoured Barrens": 1,
+  "Serra Angel": 1,
+  "Soulstone Sanctuary": 1,
+  "Sower of Chaos": 1,
+  "Spectral Sailor": 1,
+  "Stab": 1,
+  "Stroke of Midnight": 1,
+  "Tolarian Terror": 1,
+  "Tragic Banshee": 1,
+  "Tranquil Cove": 1,
+  "Uncharted Voyage": 2,
+  "Vampire Gourmand": 2,
+  "Vampire Soulcaller": 2,
+  "Vanguard Seraph": 1,
+  "Wildwood Scourge": 2,
+  "Zimone, Paradox Sculptor": 1
+ }
+}

results.json

@@ -0,0 +1,61 @@
+{
+ "note": "Win rate of the builder's decks against the decks humans built from the same pools; both decks played by gorge's search with docs/031's loop B gen 7 network (cpuct 0.5). Pools crossed (i, j), 4 games a pool pair. 17lands data, CC BY 4.0.",
+ "test_sealed_100sims": {
+  "sealed_builder_phaseC_gen1": {
+   "winrate": 0.638,
+   "se": 0.01,
+   "games": 2000,
+   "cards_shared_of_40": 19.9,
+   "main_colours_match": 0.26
+  },
+  "sealed_builder_phaseB_gen4": {
+   "winrate": 0.638,
+   "se": 0.01,
+   "games": 2000,
+   "cards_shared_of_40": 19.2,
+   "main_colours_match": 0.25
+  },
+  "17lands_gih_heuristic": {
+   "winrate": 0.552,
+   "se": 0.01,
+   "games": 2000,
+   "cards_shared_of_40": 20.5
+  },
+  "imitation_of_human_decks": {
+   "winrate": 0.53,
+   "se": 0.011,
+   "games": 2000,
+   "cards_shared_of_40": 23.6
+  }
+ },
+ "test_draft_100sims": {
+  "sealed_builder_phaseB_gen4": {
+   "winrate": 0.575,
+   "se": 0.01,
+   "games": 2000,
+   "cards_shared_of_40": 31.6
+  },
+  "17lands_gih_heuristic": {
+   "winrate": 0.509,
+   "se": 0.011,
+   "games": 2000,
+   "cards_shared_of_40": 33.2
+  },
+  "imitation_of_human_decks": {
+   "winrate": 0.527,
+   "se": 0.011,
+   "games": 1999,
+   "cards_shared_of_40": 34.5
+  }
+ },
+ "val_sealed": {
+  "sealed_builder_phaseC_gen1": {
+   "100sims": 0.669,
+   "300sims": 0.652
+  },
+  "sealed_builder_phaseB_gen4": {
+   "100sims": 0.661,
+   "300sims": 0.656
+  }
+ }
+}

sealed_builder_phaseB_gen4.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:66aecc3323487fd151f766188f21b22c123a5cccd71839ee65b21663f4fb4b85
+size 1102883

sealed_builder_phaseC_gen1.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:58aa9aad59992326bd697176306685034121e8aa52a5eebef7f1eda4a7eca736
+size 1106825