Interesting take on the deck. I am not sure if Approach of the Second Sun is a needed win con for the deck with Torment of Hailfire and Sandwurm Convergence, but having access to Fumigate is probably nice. It seems like the Abzan version of the deck has a much worse control match-up, which was my main complaint with Davis' bant approach deck. The straight BG versions have much more hand disruption to work with.
Though I would almost splash white just to play Profane Procession,because that card seems awesome. As far as wildest dreams, I think it's a little slow for a maindeck slot, but in the slower matches you have time to tutor it up with acquisition.
The idea is Torment of Hailfire is marginal without having access to tons of mana and Sandwurm Convergence could be worse than Approach if opponent has ways to deal with enchantments. And vice versa if opponent is playing counters or can mill you for a few cards.
Against control you have Thaumatic Compass which means they can't win you with a single gearhulk and Ipnu Rivulet. The way Bant Approach was winning those matchups was by milling them G1 and then siding out all Approaches and siding in 6 creatures. Davis himself said that Approach is just an alternative wincon
After few playtesting - probably I did wrong removing Azor's Gateway. Also in those long games which Wildest Dreams is supposed for I often cleared graveyards a few times so it was less effective.
I've started playing UR control few weeks ago and went 7-0-2 in matches in last 3 local tournaments (played vs mirror & esper control, vs jund gods, vs mardu machines, etc)
One interesting thing I've found is that my sideboard plans all had one thing in common:
-4 Torrential Gearhulk
-3 Censor (I play only 3)
-something else, depending on matchup
+4 Thing in the Ice
+2 Dragonmaster Outcast
+4 Dispel
+1 Negate
+something else
So, what your thoughts on siding out all gearhulks and becoming more lighter tempo deck?
Hey, I've been trying similar idea few weeks ago and my last decklist looked this way: http://tappedout.net/mtg-decks/ur-paradox-combo
Some points:
1. Whir of Invention allows you to tutor for Paradox Engine when control is tapping out for Torrential Gearhulk EOT.
2. Bunch of instants + Fall of the Titans as wincon allows to win with Torrential Gearhulk still on stack.
3. Commit // Memory is backup 'draw 7' card for going off which also doubles as counterspell to push combo through or removal to stall game a bit.
4. During one game I was able to combo out with charged Pyramid of the Pantheon, Paradox Engine and 6 lands in game only (starting with Commit // Memory), so I have some doubts if we really need other mana stones.
5. I view this deck not as combo deck but as control with combo wincon instead of 'bash with gearhulk' plan which doesn't always gets there.
How are you planning to reliably win vs emrakul? Opponent can simply wait till your Demonic Pact, mindslaver you and select 'You lose the game' option.
Hi! I'm developing website www.top8draft.com.
Top8Draft is website with draft simulator which uses deep neural networks for predicting player picks. AI can analyze cards picked by player and estimate strength of each card in pack in relation to picked cards. After completion of draft you'll be able to see these estimations, compare your choices with computer's and see where you probably made an error. Or you can upload draft log from MTGO and analyze it to get same estimations - this is a fast way to get some feedback on your draft.
Currently, these formats are supported: 3xORI, OGW-OGW-BFZ, 3xSOI, EMN-EMN-SOI, 3xEMA
I strongly believe we'll see a Serra Planeswalker card in the future. Most likely on a suppelmental product, as she's long dead, but it's definetly on the table after the sucess of Commander 2014. We still need an enchantment centered Planeswalker and one that spams Angel tokens. Who better than Serra?
On a side note, Urza will probably get a card eventually. Probably our first 5 color Planeswalker
I could only hope Serra PW gets printed in a new Commander product, eventually, although 2015 didn't really have much in the way of characters we have seen before, besides Ezuri. More call backs to past characters were probably the best part of C14, Feldon is the most favorable card in all of Commander, I think.
Urza seems more 4 color than 5. Was he ever shown doing a lot with white mana, or any of the white characteristics?
I'd say 'chaotic' and 'reckless' definitely weren't Urza's traits. Same for 'interdependent' or 'peaceful'. So he should be in esper colors.
Actually, I've tested a deck with similar idea - t2 Heir of Falkenrath or Pale Rider of Trostad, discard Stitchwihg Skaab or Ghoulsteed, t3 activate zombie in graveyard discarding Prized Amalgam. Bam! It also had Asylum Visitor and other cheap madness spells.
Even a couple pieces of targeted removal doesn't stop that. If you try to go under them, they just let you hit them because they know they'll lifelink back to stability. Only Radiant Flames or Kozilek's Return really has a chance.
rx303: Actually, I don't want to teach the game how to play Magic. I want to teach it the general 'good strategy' for a general set. IOW, I want to teach it a general strategy for 'drafting well'.
And then, I want to use, for each new set, the NN in its state from the previous set. Both to watch the evolution of 'good drafting strategy' in general and also to find out how the 'conventional wisdom' as applied to a particular draft format fails.
That would've worked if MTG card could be easily encoded. But major feature humans use to evaluate card is rules text. I doubt NN can extract relation between power level of card and its text. A lot of sets introduce new keywords which never appear further. Often small changes in text cause huge changes in strength of card. It also depends on context - all other cards in draft format. For example, look at Makindi Aeronaut. Even LSV thought that card is bad based on his previous experience. But it proved otherwise because that card played well with other cards in set.
I think that it is easier to learn cards relationship from game and card rules represented as hardcoded game engine rather than text.
rx303: Thanks for the explanation! A couple questions.
Can you use the NN to extrapolate beyond the current cardset? I.e., this seems like a very brute force approach to doing this. As I understand correctly, you are not encoding anything about the card itself other than its name, and then deducing what you should pick by what the Network 'knows' about the cards and how they relate to other cards, right?
This will work with proper training data (which will be the crux in any new format) for any given environment. However, can you build a NN that actually learns 'card rules' so that it can actually generalize beyond what you're giving it.
You wouldn't have to build a new NN, just give the current one that info and it will be able to figure out the relationships between that data and what it current;y knows and then apply that to any format.
Exactly. The only data I use are names of the cards and then network figures relations between them for particular draft format.
To build NN which makes decision based on card rules, we need to teach it play magic first. And to do so we need to build a proper model of game state and then we could apply reinforcement learning algorithms to it. Currently, I'm trying to figure out how should I encode game state. There are some tricky aspects - like unlimited number of permanents on the battlefield with different effects on them or selecting and storing targets for spells.
The idea is Torment of Hailfire is marginal without having access to tons of mana and Sandwurm Convergence could be worse than Approach if opponent has ways to deal with enchantments. And vice versa if opponent is playing counters or can mill you for a few cards.
Against control you have Thaumatic Compass which means they can't win you with a single gearhulk and Ipnu Rivulet. The way Bant Approach was winning those matchups was by milling them G1 and then siding out all Approaches and siding in 6 creatures. Davis himself said that Approach is just an alternative wincon
After few playtesting - probably I did wrong removing Azor's Gateway. Also in those long games which Wildest Dreams is supposed for I often cleared graveyards a few times so it was less effective.
2 Duress
4 Fatal Push
3 Baffling End
3 Thaumatic Compass
1 Doomfall
1 Profane Procession
4 Gift of Paradise
3 Ixalan's Binding
3 Mastermind's Acquisition
1 Vraska's Contempt
2 Yahenni's Expertise
2 Fumigate
3 Hour of Promise
2 Wildest Dreams
1 Arch of Orazca
4 Blooming Marsh
2 Desert of the Glorified
2 Desert of the Indomitable
2 Forest
1 Hashep Oasis
1 Ifnir Deadlands
1 Ipnu Rivulet
2 Plains
2 Scattered Groves
1 Scavenger Grounds
1 Shefet Dunes
3 Sunpetal Grove
3 Swamp
1 Baffling End
2 Duress
2 Fumigate
1 Gideon's Intervention
3 Naturalize
1 Overwhelming Splendor
1 Protection of the Hekma
1 Ramunap Excavator
1 Sandwurm Convergence
1 Wildest Dreams
Thaumatic Compass is a huge sleeper. After you've played it in a deck with Hour of Promise it would be an immediate candidate to any deck with the Hour.
Wildest Dreams is another nice way to transform mana advantage into card advantage (first one in Arch of Orazca). I think it is a better maindeck slot than Vraska.
Ixalan's Binding is a clean answer to The Scarab God, Hazoret the Fervent or other problematic cards like Anointed Procession
Gift of Paradise is better than Beneath the Sands because it gives double coloured mana for Fumigate and Mastermind's Acquisition and 3 life is a significant bonus.
I wonder how many matches of those 2 5-0 did BG won by time?
It was small this time though. 24 people only, PT qualification for top-2.
Other finalist was on Mardu Machines.
One interesting thing I've found is that my sideboard plans all had one thing in common:
-4 Torrential Gearhulk
-3 Censor (I play only 3)
-something else, depending on matchup
+4 Thing in the Ice
+2 Dragonmaster Outcast
+4 Dispel
+1 Negate
+something else
So, what your thoughts on siding out all gearhulks and becoming more lighter tempo deck?
Some points:
1. Whir of Invention allows you to tutor for Paradox Engine when control is tapping out for Torrential Gearhulk EOT.
2. Bunch of instants + Fall of the Titans as wincon allows to win with Torrential Gearhulk still on stack.
3. Commit // Memory is backup 'draw 7' card for going off which also doubles as counterspell to push combo through or removal to stall game a bit.
4. During one game I was able to combo out with charged Pyramid of the Pantheon, Paradox Engine and 6 lands in game only (starting with Commit // Memory), so I have some doubts if we really need other mana stones.
5. I view this deck not as combo deck but as control with combo wincon instead of 'bash with gearhulk' plan which doesn't always gets there.
2. Ishkanah, Grafwidow
3. Incendiary Flow - not only answers advocate, but also hangarback and gisella
4. Spell Queller
5. Liliana, the Last Hope - eats WW humans and UW spirits for breakfast
6. Thalia, Heretic Cathar
7. Stromkirk Condemned
8. Grim Flayer
9. Bloodhall Priest - vampires are real deal
10. Hamlet Captain
Top8Draft is website with draft simulator which uses deep neural networks for predicting player picks. AI can analyze cards picked by player and estimate strength of each card in pack in relation to picked cards. After completion of draft you'll be able to see these estimations, compare your choices with computer's and see where you probably made an error. Or you can upload draft log from MTGO and analyze it to get same estimations - this is a fast way to get some feedback on your draft.
Currently, these formats are supported: 3xORI, OGW-OGW-BFZ, 3xSOI, EMN-EMN-SOI, 3xEMA
I'd say 'chaotic' and 'reckless' definitely weren't Urza's traits. Same for 'interdependent' or 'peaceful'. So he should be in esper colors.
I feel like UB Demonic Pact deck is a sleeper. It has a lot of options vs Westvale Abbey decks like Languish, Disperse, Crush of Tentacles and Silumgar's Command
That would've worked if MTG card could be easily encoded. But major feature humans use to evaluate card is rules text. I doubt NN can extract relation between power level of card and its text. A lot of sets introduce new keywords which never appear further. Often small changes in text cause huge changes in strength of card. It also depends on context - all other cards in draft format. For example, look at Makindi Aeronaut. Even LSV thought that card is bad based on his previous experience. But it proved otherwise because that card played well with other cards in set.
I think that it is easier to learn cards relationship from game and card rules represented as hardcoded game engine rather than text.
Exactly. The only data I use are names of the cards and then network figures relations between them for particular draft format.
To build NN which makes decision based on card rules, we need to teach it play magic first. And to do so we need to build a proper model of game state and then we could apply reinforcement learning algorithms to it. Currently, I'm trying to figure out how should I encode game state. There are some tricky aspects - like unlimited number of permanents on the battlefield with different effects on them or selecting and storing targets for spells.