General CBS

[shamizy]

Neat that they introduced a voucher option that'll be viable for most people, unfortunately my system having extra steps makes it a little tougher to implement 1-to-1 on Cobra. For traditional vouchers/squadron picks it's going to work very well just giving you the copies of whatever card(s) you need, but mine has the caveat of limitations to cashing in once per card and having to draft the actual copy of the card.

I'll probably just have to continue including the voucher options in the land box for a drafter to manually add to their Cobra decks online.

All that said, this is a pretty handy tool with some neat possibilities. I ran it for 300 in the background while working on stuff and it came out with some cool outputs:



Expected color distributions here, nothing out of the ordinary with the base breakdown.



26 clusters in my 450 sounds about right, I'm liking the criss-crossing overlap showing the ability to bleed between archetypes. I'm assuming that running for more simulations will leader to more dense clusters like @Erik Twice got with 500. Maybe I'll just go with 1000 or something at the beginning of work next week and see what happens by the end of it.



I like the visualization of these clusters with staples and other stats, really cool feature to nerd out over for hours at a time. The bots definitely have their limitations overall, but this looks way better than previous bot logic applied in single drafts. It identified just about every major archetype I've built over the years or covered in my cube primer.

 

[MusicalBen]

38 clusters found after 400 drafts

Many of these are repeated, which, after looking into the decks, sometimes makes sense and sometimes doesn't. The two "W Weenie" clusters, for example, roughly refer to two distinct groups of cards, a tokens build and an artifact aggro build. Yet, the two "WG midrange" decks have essentially the same staple cards...

Some of the themes that bridge certain colors are not represented, even though they are not an issue to draft in the cube, such as the White-Black Constellation/Saga Control deck. I think the nuances of this deck and these colors might too fine-grained for the CubeCobra bots, with options ranging/blending between Liliana's Contract/Changeling combo, Demonic Pact Blink, low-power-matters reanimator (Shirei, Shizo's Caretaker, Vesperlark, Triskelion) and so on.

Green and Red turned to be the least drafted combination, and the strange assortment of cards chosen by the bots reflect the variety of decks which might be drafted in these colors but which are, nevertheless, not always overlapping: tappers/untappers, land storm, madness, power-matters, etc.

Something I didn't expect, Dimension X Pizzasaur was the second highest picked card, being taken P1P1 98% of the time.

 

[OlegKlishin]

I played around a bit with my old cube and my WIP cube. Tried 100, 300 and 1000 drafts. At 1000 there are ton of extraneous data, like 4 same WR weenie archetypes. Also, bots obviously can't figure more niche archetypes. They figured out enchantments as WG midrange, but completely ignored BG and BW enchantments.

While I get that you can take lessons from how things connect on the map, I struggle to imagine how exactly to analyze the data to learn something actually useful. I'd be interested if someone more knowledgeable can point out what's valuable except obvious toy factor.

 

[Mown]

It's a cute tool, but it relies on a relative degree of normalcy. It can draft a storm deck in my cube and it looks quite competent, because it knows that Lion's Eye Diamond and Yawgmoth's Will often appear together in the same decks. On the other hand, it doesn't have the knowledge base to think about drafting Stoneforge Mystic and Trinket Mage to fetch Giantcraft Helm that buffs Disruptor of Currents and Ethereal Forager. To the contrary, the "exemplary" WB weenie deck plays a Stoneforge with no targets; the tool is more keen to pair the card with Ranger-Captain of Eos than Pre-War Formalwear (although they are both strongly associated). Ultimately, the bots are really good at building uninspired and boring decks that miss the point of the cube, so I struggle to see what utility I would get out of it. If I had a ton of drafts of my cube and ran it on only that data then you could maybe draw some fun conclusions.

It also picks the three SOS cards I just added 100% of the time with seemingly no context, so you get fun things like white weenies splashing Tragedy Feaster off 0 black sources.

 

[Erik Twice]

I did try with 2500 draftrs, but it doesn't lead to better data since it runs into limitations with duplication and clustering. The data is pretty much identical to 500, it's just repeated more times. It's also hard to adjust it to a reasonable level.

If sorting by colour, rather than just clusters, I think one can easily see the problem with green in my cube. Green is to the left, very heavily clustered around the lands archetype. It has UG ramp below and BG midrange on top, but little WG and, most importantly, no GR.

This makes sense since WG, UG and GR are historically weak guilds with little support and lack of powerful interaction or archetypes.

 

[Nanonox]

While I get that you can take lessons from how things connect on the map, I struggle to imagine how exactly to analyze the data to learn something actually useful.

Whether it's correct or not, I am treating the data as a way to show the more obvious archetypes that a cube contains. Another leap of dubious logic and now I'm equating that to what a newer player would find signals for.

I'll be posting more in my blog, but for example my lands/ramp cluster doesn't really communicate with the storm one, even though they should go well together. So as a more obvious signal, I am thinking of adding Sail into the West, a card I've considered before, but not as an explicit signpost to go into Simic combo.

 

[Zoss]

I'd be interested if someone more knowledgeable can point out what's valuable except obvious toy factor.

As someone who interpreted somewhat more UMAP/clustered spatialized data plots than the average bear (though never made them), here are my two cents on what these are good for.

--Identifying new clusters you may have missed (i.e. all intended decks are considered "draftable" from historic human draft data).

--Make sure that all the clusters you think are there are there and that they have roughly the right number of members per cluster (i.e. color combination balance).




Then, once you've moved on from there, ask card-by-card questions.

--Ensure that cards are where you think they will be. (Want Goblin Bombardment to support Aristocrats? Make sure it's not in your Izzet Spellslinger pile)

--Identify cards that may be specific or nonspecific to archetypes. (Bone Shards getting in your UB Control? Yikes. Do you have LOTS of nonspecific cards? Maybe there's too much good stuff.)


In a nutshell, these are sniff tests--are the cards where you expect them to be? These are necessarily data reductions, though, so you're losing a lot of nuance (as we've already figured out in this thread).



A note on cluster annotations:

I don't know how much you can actually count on the bots knowing what they are doing. After a draft I checked what the bots have been drafting and they've named a slightly weird black grindy sacrifice deck: Black Tempo. Maybe they don't really know what they are doing actually.

The bots absolutely, 100% do not know what they're doing, in the sense that I would completely disregard the name, but I think this is still valuable. It strongly suggests that these cards have a track record of working together (or at least that humans draft these cards together), though it doesn't know how these cards work together or why. Has anyone drafted a deck like this? If not, might be interesting to force it a few times!

From reading the documentation, it sounds like Gwen & co. manually set out a pool of cards for each cluster label. If you see something that doesn't make sense, it's likely because you've done something way off the beaten path and the computer grabbed the nearest label out of desperation. In which case, congrats! I want to see those decks.

 

[alvoi]

I feel very stupid, but... I tried using it, many times, on many cubes, and it always gives me 800 different clusters, each one with exactly one deck. I am not sure about what I am doing wrong...

 

[Zoss]

I feel very stupid, but... I tried using it, many times, on many cubes, and it always gives me 800 different clusters, each one with exactly one deck. I am not sure about what I am doing wrong...

This sounds like a fascinating problem! Would you mind sharing a screenshot of your settings? Idk if any of us will be able to diagnose the issue, but it sure sounds interesting.

 

[alvoi]

Okay, so, I tried for the Cube they used at the 2007 Invitational: https://cubecobra.com/cube/list/invi07-og?view=Mainboard (so that I am sure the problem is not that my Cubes are so badly done that even cubecobra refuses to touch them)

I haven't change anything in the settings, everything very normal

All 800 drafted decks belong to a cluster that consists of just them, and all these clusters are overlapping in the map. Also, the name of each cluster is (colour of the deck) + Big Ramp -- and this is true in any Cube I try:

 

[Zoss]

Big Ramp Problems

Can't wait to see their bug report on this--at least it got G Big Ramp correct!

 

[Mown]

What's the pick priority distribution like? Completely even?

 

[Zoss]

One more thought on what this cluster map *could* be good for:

This feature isn't included because it would be computationally expensive, but where cluster analysis of UMAP is really really good for is comparisons of a map under similar but distinct conditions. For example, if I added in the aforementioned Goblin Bombardment, what does it do to my clusters? What does it enable, and what does it make less likely to coalesce? Which clusters are more prevalent, and which decrease?

To make this really explicit, if you're thinking of some additions or cuts, implement them, rereun the simulation, then compare the maps they draw.

More later because I'm super tired this week, but I wanted to get this in there in case someone else has the bandwidth to do so. If not I'll probably illustrate this tomorrow.

 

[alvoi]

What's the pick priority distribution like? Completely even?

No, that one works well! I think the only problem is with the clustering.

You are right though, probably I can still extrapolate useful information from that