Quote:
Plenty have already done this. DotP did this. There is a big difference between playing the game, which is quite trivial to implement, playing the game well, and playing the game perfectly (which is what the Go and Poker AI's in that past couple years were built to do).
Yes, your method has to be demonstrably better than or similar to the best humans, otherwise it doesn't provide valuable input for strategy/tactics discussions. This is the point I have been trying to make all along
Quote:
This example tracked the number of lands drawn at 24, 25, and 26 land
Why would you use a Monte Carlo method to approximate this instead of just using a Hypergeometric distribution to give you the correct answer?
(unless your whole point was to illustrate that stochastic processes with a large sample size will always have a small inherent margin of error, in which case, fine)
Quote:
People always go back and forth on land counts, they cut a land because they feel like they're flooding, or add a land because they were mana starved but the reality is that those types of situations play out over a handful of games, at the most 100 games. When 100,000 games doesn't even give you a solid base to conclude that 24, 25, or 26 lands are correct, and considering that land drops are the most consistent action in almost all decks, then every conclusion you make about how a deck functions is also on equally shaky ground.
Like I say, you can calculate the exact percentages for all the entries in that table, but it still wouldn't tell me whether 24/25/26 lands are correct, because all it tells me is how many lands I need to play in order to see X number of lands with Y probability in the first Z cards
This information is not totally worthless (hence why Frank Karsten's articles are so popular), but it doesn't actually tell me how many lands I should play in order to maximise my win percentage. To put it another way, it doesn't tell me what are the values of X or Y or Z that I should be aiming for.
How many lands should I play considering my mana curve? How many lands should I play considering the number of cantrips in my deck? Should I play more/fewer lands if I think my deck gets punished more for getting screwed/flooded? Can I play fewer lands if I have basic lands that don't get destroyed by wasteland? If I have my own wastelands, do I count them as lands, or as spells? If the meta has a lot of waste/stifle decks in it, do I need to play more lands? I can give contextual answers to all of these questions. Yes, my answers are based on 'shaky' human experience, but my point is that we don't currently have a computational process that accounts for these factors, so human experience is the best method we have.
Quote:
I've been testing the system for years now. The first variant played Modern Knightfall before that deck had any tournament success, and therefore no lists to base a deck skeleton off of. I was able to identify one, and that skeleton ended up getting extremely close to the deck in it's modern day form. After that I used Burn, and was able to identify cards that weren't being played that had potential. I've then added those cards and played it at tournaments, to decent success.
You are implying that you fed a database of every modern legal card into your solver, and ab initio it spat out a list similar to what Kelvin Chew designed . This claim defies credulity.
If what you're suggesting is instead "I came up with a deck skeleton myself, and then a computer program suggested a few tweaks based on some guidelines I decided upon myself, and in the end I got a list that was a bit worse than what a pro player came up with" then this is entirely believable, but not useful or noteworthy at all (insofar as being an endorsement of your computational system).
From someone who suggested that 100,000 games isn't enough to get good data only a few sentences earlier, it's amusing to see "I changed a couple of cards in my burn deck and won a few games with it, so the system works"
Quote:
It's not perfect, but no one said it was. Echelon and I were just explaining to the person with their matrix that if you want to follow that approach, there's a much better way to do it which is with rule based evaluation card by card. Not a point based system for effects.
I'm not insisting that it has to be perfect, only that it has to improve upon what we can achieve without it.
I find it bizarre that you would insist one wonky numerical method is "better" than another when you don't have any way to validate that assertion.
Simply, it seems your advice was 'Rather than try to think about what your cards do in isolation, it may be better to think about how they match up against specific cards that you expect to face'. I can't really disagree with this statement.
But it's a huge (currently: impossible) leap to go from that idea to having a system that objectively tells you the best cards to play, so pretending that such a system currently exists is misleading (or at least unhelpful).
Quote:
Edit: And all of this has an actual impact on discussing deck building choices. To give an example, something you'll see very often in Nic Fit is people playing a 61st card. This tends to go against almost all conventional wisdom for deck building. However, through various methods (logic/reasoning, a simulation such as my above example, a little bit of practice, and so on) we have over the years come to the conclusion that the toolbox value of having one additional card in the deck so that it's there when you need it, outweighs the inconsistency added to the deck. I've proven this in several decks now, so it's not just a toolbox related effect. Having an out to something unexpected that you can potentially draw in your deck as an additional card is much more valuable than adhering strictly to 60 cards, and that's something I've even started applying to other decks and formats.
"Proven" is a very strong word
Quote:
For example, Pod is one of the weakest archetypes and BUG is the weakest color combination. Given the cost of the format, it's only fair to tell people up front if their build is or isn't viable.
From MTGGoldfish results from the last 12 months it seems like there aren't any Birthing Pods, but BUG looks pretty evenly split with Junk and GB. It seems fair to tell people upfront whether your claims are based on theorycrafting or metagame data or your own testing or whatever else.
Quote:
To pile onto Brael: in software development the scale of work generally isn't considered when talking about stuff like the core workings of an AI. Consider the AI that beat pro dota players last year. This sums up how it works: It's a neural network that loses points for taking damage and gains points for inflicting damage. Sounds simple, right? Also sounds like easy to make, right? It isn't though, it took a team a good while to build it. But that doesn't change the core principle.
In our case we weren't discussing how we built AIs to become MtG gods, but how some of the core principals of building one would work. Very different things. My issue was that you didn't understand what it was you were reading and yet you kept pushing on. It's probably my fault for not explaining well enough.
I work in a medical research lab; it isn't my own specific area of expertise, but my colleagues do the exact kind of deep-learning for image analysis mentioned a few posts back and we frequently discuss it. What exactly is it that you think I don't understand? I haven't commented at all on the techniques you might use to invent a better system. My only problem is the claim that there is a computational approach that builds and plays decks today and that this current system can build decks at a level comparable to professional/skilled players. These claims are unverified (if not outright false) so it's wrong to suggest that the best cards/decks are found by following any sort of numerical scheme.
Even Brael's proposed solution is incredibly limited because as-described it doesn't actually involve any machine learning. As long as the "rule based evaluation from card to card" is given to the system from its designer, and not generated by the system itself, there will always be a significant cap on the insight that can be obtained beyond what the designer already knows or believes (assuming those beliefs are even correct in the first place).
Quote:
Thoughtseize beats Liliana of the Veil
?
