Rob Cashwalker wrote:I must have forgotten to upload the change to the deck editor I made on my home copy. I just added the set code support in the draft edit mode, it should work now in the latest rev.
Thanks for fixing it mate, I can confirm that it works correctly now. Great job, your awesome work on the draft mode is much appreciated!
Here are just a few observations I made regarding the AI drafting that maybe will be helpful to you in your future endeavors:
1) I noticed that the AI doesn't focus much on choosing specific colors when drafting and commonly drafts something of everything (possibly the best card it sees no matter the color? I don't know the algorithm so I'm not sure what happens exactly), and it doesn't try to strengthen its early play (at least it's my observation from, like, 7 or so drafts that I made over the past few days, playing with all the 7 opponents that were generated each time). In the end the AI commonly ends up with 4- or 5-colored decks (only very few decks are 2/3-colored, like one or two out of seven per draft) which it can't employ correctly because its early play is stuck. Four- or five-colored AI decks seem to, in most cases, end up being stuck early on in the game (it's not uncommon to see the AI do nothing except play various lands for the first 4 or even 5 or more turns) and once it finally gets to playing its creatures and spells the positional advantage I have by then is typically too overwhelming. The more focused, two- or three-color decks seem to play better for the AI as it seems to be better at the early play and transitioning into the mid-game with such decks cause it usually has some stuff to cast earlier on to hold off my earlier attacks and perhaps even create some early pressure (depending on what it has and what I have in hand). And also, it seems like generally speaking, 2-color or maybe 3-color (with good mana source micromanagement) decks with a good mana curve and with enough cards at CMC slots 1-4 and some late-game cards and finishers at CMC 5+ seem to play generally better than 5-color decks that do not have a particular focus and that generally have to rely on a high random factor (not only the player would have to get the cards that he would be able to employ both early in the game and later on, but also he would need to get the appropriate mana sources to actually cast those spells). By going 5-color, by the way, the AI also creates a massive disadvantage for itself against the landwalk cards - for instance, today I was able to beat the AI deck with my single swampwalk 3/3
Anaconda simply because the AI played a
Swamp which it didn't actually use in any way (it was about turn 7 or so and it hasn't cast anything that required a black mana by then - and it only had one
Swamp and something like 2
Mountain / 2
Forest / 1
Island plus a nonbasic land that provided UW or UB (can't remember which); for the same reason it wasn't able to cast much by turn 7 because probably whatever options it actually had in hand were inaccessible early on due to wrong mana colors). I'm not sure if it's just a subjective impression perhaps, maybe the AI was just terribly unlucky in most games I played against it (to be perfectly honest, I haven't found time yet to actually analyze the decks the AI creates when drafting - the draft mode is just way too addictive for me for now to actually do something like that
). At any rate, I keep drafting every day, so I'll keep observing what's going on. I know that it's very difficult to come up with a solid, strong drafting AI, so I wish you the best of luck in improving the algorithms should you choose to do it!
2) The AI drafts cards that it can't correctly use and then puts them into the deck. This leads to a bigger number of glitches than in regular random or quest games, because the AI will put, for instance, pain lands into its deck and then it'll use them without getting hurt (someone over at the buglist thread pointed out that it's a known fact that the AI can't and shouldn't use pain lands for now cause it just can't). To tell the truth, I'm not sure what can be done here, cause if the AI drafts them but then doesn't put them into the deck, it might create a disadvantage for the AI as it will draft potentially useless cards for itself... Perhaps it should only resort to drafting cards that it can use in the first place then?... As more cards get supported (I believe there's some kind of a flag that tells which cards the AI can and can't use), the AI drafting will also improve on its own cause it'll be able to add more and more cards to its arsenal over time? This, of course, leads to the problem - what should happen when only the no-AI cards are remaining in the draft pool?... Oh well, unfortunately, this is a puzzle I was unable to resolve, so I'm afraid I don't have a very good suggestion here, sorry...
EDIT: Perhaps the second drawback can be improved by a little more prioritization of what the AI drafts and puts into the deck. I was thinking about the algorithm along these lines (simplified and not perfect, but might be something to think about if you haven't considered it yet): When drafting, the AI should look if there are cards to draft that are not marked as NoAI. If there are, it should draft the best card from those (using whatever algorithm it's using to decide which card is the best). If all the cards from which to draft are marked as NoAI, it should as a last resort option draft the best card from those, using once again its algorithm to determine what is best. When deckbuilding, the AI should try to complete its deck using the cards that were drafted and that do not have the NoAI flag set. As such, it should prirotize any nonbasic or basic land that does not have the NoAI flag over a nonbasic land that has the NoAI flag, etc. If all the working cards are exhausted but the deck is still not complete, then the AI should try, as a last resort option, to complete the deck with the remaining NoAI cards that it has. Feel free to ignore any of it, just my $0.02.
At any rate, even the current level of AI drafting is massively fun, enjoyable, and impressive, congratulations on this important breakthrough and good luck in your future work on this awesome mode!
Please don't take anything here as an unwarranted criticism or anything, I'm merely trying to be helpful and provide some feedback based on my experience and impressions.
Posted: 16 May 2011, 13:52
