Rank: Admin
Groups: Administration
Joined: 10/20/2007
Posts: 1,525
Points: 3,139
|
Been playing a lot lately, finally beat Sorin Markov (Exp 2) last night. The AI is absolutely amazing in this game and was doing some scouring for info about it and found this (Warning, pretty geeky read):
The game's artificial intelligence (AI) is based on creating a multistep "lookahead" decision tree from the current state of the board, evaluating possible known moves by its opponents, scoring the current and resulting situations both positively and negatively, and then selecting the option with the highest score. The artificial intelligence utilizes threading to run the main intelligence algorithm alongside three "sub-contractors" that evaluate the possible future states and report back to the main algorithm. Each of these instances of the AI uses a special engine that both implements the rules of Magic, while also providing "undo" actions so that they may explore up and down the decision tree. The game's intelligence algorithm runs on a separate CPU core than the main game to avoid creating framerate issues with its display while the player is considering their actions.[4] When the player makes an action, the AI threads are interrupted and brought to the same state as the game, then continue processing, in some cases, reusing existing branches on the decision tree that match with the player's selection.[4] The AI itself does not employ any strategy in terms of strong card combinations, but instead, when such combinations improve the computer's situation, the chances of playing these combinations will "ripple" back through the decision tree. In this manner, the computer AI will play one card of such a combination should it be the best choice at that time, and follow up with the second card of the combination should that option still remain the best after any other player actions.[4]
In order to simplify the task of forward evaluation, the lookahead algorithm was optimized in several ways, considering "the idea that the AI needs to be excellent 99% of the time, even if it's at the cost of being sub-optimal that other 1%".[4] The computer will not consider actions that are considered "pointless", such as healing an opponent, though as noted, there are combination of cards where such actions are useful in the normal card game; as this was found to be a useful feature, the game also implements a similar check on player actions, though the player may disable this feature. The computer also skips any further calculations while it has control of the top action on the stack. The algorithm only looks about three moves (or six total actions) ahead to avoid considering too many deep options. In order to respond quickly, certain actions, such as attack and blocking resolution, are permuted and evaluated (a considerably fast process), and then are sorted in order; the computer then proceeds the lookahead from each option, in order, with the likelihood that if the computer should run out of time, one of the high scoring options will produce the best possible result.[4]
As a result of these means of programming the AI, it was noted that the component opponents may exhibit behavior not found by typical Magic players, such as not holding back attackers to act as blockers for the opponent's turn, though as noted by the developers, such aggressive play is generally a winning strategy for the computer.[4] The AI also does not have any special knowledge of the constructed decks itself or what cards an opponent might have given the visible cards in play; again, the developers noted that approach can sometimes work better instead of contemplating "big and clever" plans.[4] The developers noted that they observed unusual but effective strategies from working with the AI that they were not aware of before, primarily based on the timing of playing certain cards and effects.[4]Wikipedia: Dual of the PlaneswalkersMagic: The Gathering Duel of the Planeswalkers
|
