Artificial intelligence: Difference between revisions
Moved Brawl rumours to their own section below the specific examples. Added list of reasons why the "learning AI" rumour is false.
(→Difficulty levels: amiibo) |
(Moved Brawl rumours to their own section below the specific examples. Added list of reasons why the "learning AI" rumour is false.) |
||
| Line 195: | Line 195: | ||
#Mighty-8 | #Mighty-8 | ||
#Nasty-9 | #Nasty-9 | ||
The flaws in ''Brawl'''s artificial intelligence are considerably lesser and less recurrent than in ''Melee'', but they can still be potent. Computer players still tend to be easier to KO and [[edge-guard]] than the average human player, even when set at level 9, due to not utilising survival techniques such as [[DI]] and usually not fighting back when being edge-guarded. CPUs do, however, [[air dodge]] when sent flying, which provides minimal cushioning for knockback although usually not enough to save them. Computer players sometimes have difficulty avoiding certain stage hazards (most notably on [[Halberd]]), and are prone to self destructing in scrolling stages, such as the [[Past and Melee Stages|Melee Stages]] [[Rainbow Cruise]] and [[Big Blue]]. They have generally poor edge-guarding abilities (though unlike the previous two games, self destructing while attempting to edge-guard is very rare), and will usually stand on the ledge and spam projectiles when the player is recovering; they will usually attack offstage intentionally only with characters that have multiple jumps, such as {{SSBB|Meta Knight}}. And while they are better at recovering than in the previous game, computer players still under-utilise recoveries, such as by not making use of alternate recovery options with some characters, or always recovering in the same direction with a recovery move. A computer player (regardless of level) will also always aim for solid platforms that can be [[edge sweet spot|sweetspotted]] when recovering, even when there are "soft" platforms off the stage. | The flaws in ''Brawl'''s artificial intelligence are considerably lesser and less recurrent than in ''Melee'', but they can still be potent. Computer players still tend to be easier to KO and [[edge-guard]] than the average human player, even when set at level 9, due to not utilising survival techniques such as [[DI]] and usually not fighting back when being edge-guarded. CPUs do, however, [[air dodge]] when sent flying, which provides minimal cushioning for knockback although usually not enough to save them. Computer players sometimes have difficulty avoiding certain stage hazards (most notably on [[Halberd]]), and are prone to self destructing in scrolling stages, such as the [[Past and Melee Stages|Melee Stages]] [[Rainbow Cruise]] and [[Big Blue]]. They have generally poor edge-guarding abilities (though unlike the previous two games, self destructing while attempting to edge-guard is very rare), and will usually stand on the ledge and spam projectiles when the player is recovering; they will usually attack offstage intentionally only with characters that have multiple jumps, such as {{SSBB|Meta Knight}}. And while they are better at recovering than in the previous game, computer players still under-utilise recoveries, such as by not making use of alternate recovery options with some characters, or always recovering in the same direction with a recovery move. A computer player (regardless of level) will also always aim for solid platforms that can be [[edge sweet spot|sweetspotted]] when recovering, even when there are "soft" platforms off the stage. | ||
| Line 252: | Line 245: | ||
*CPUs have a difficult time getting out of the underground part of [[Temple]] when opposing players are on the upper part. This is especially noticeable with CPU {{SSBB|Wolf}}<ref>http://youtu.be/a3Uus1Ql2xg</ref>. | *CPUs have a difficult time getting out of the underground part of [[Temple]] when opposing players are on the upper part. This is especially noticeable with CPU {{SSBB|Wolf}}<ref>http://youtu.be/a3Uus1Ql2xg</ref>. | ||
*CPUs will never swim on [[Pirate Ship]] or [[Delfino Plaza]]. | *CPUs will never swim on [[Pirate Ship]] or [[Delfino Plaza]]. | ||
====Rumors==== | |||
Many rumors have circulated about the AI of ''Brawl''. One rumor claims that level 9 CPUs read button commands the player inputs, thereby allowing them to perfect shield and air dodge more often. This sparked general dislike among the community, many of whom labelled the idea as "cheating". There has been no evidence however that the AI reads the player's inputs, and "evidence" presented for it has just shown that the AI has 1 frame reactions. | |||
Another rumor claims that there's a system in ''Brawl'' that enables AI players to "learn" from human players.<ref name="CECPUexpBlog">{{cite web| url = http://allisbrawl.com/blogpost.aspx?id=8848| title = The Wonders of Brawl's CPU Experience System| accessmonthday = Jan 2| accessyear = 2011| author = [http://allisbrawl.com/blog.aspx?id=10347 ChurroEmiliano]| date = 2008-12-31| format = blogpost| work = | publisher = [http://allisbrawl.com/ AllIsBrawl.com]| quote = When developing Super Smash Bros. Brawl, Sakurai integrated an experience system in which CPUs (mostly level 9/nasty) would actually learn from other players and try to mimic such strategies. This experience system is subject to all characters. / The way you and others play on your save data, the way they will intend to play, keeping several of their own ways of playing as well. You can tell a huge difference, by comparing...a new fresh save data of Brawl, to a very much used save data.}}</ref><ref name="CECPUexpVideo">{{cite web | |||
| url = http://www.youtube.com/watch?v=g4HvhIPOz4I| title = The Wonders of Brawl's CPU Experience System| accessmonthday = Jan 2| accessyear = 2011| author = [http://www.youtube.com/user/ChurroEmiliano ChurroEmiliano]| date = 2008-12-29| format = video| work = | publisher = YouTube| quote = Churro (Snake) vs. Lvl 9 Link}}</ref><ref name="CEFalconPunch">{{cite web | |||
| url = http://www.youtube.com/watch?v=YMcEUsMd9Cw| title = Falcon Punch Much? 0_0| accessmonthday = Jan 2| accessyear = 2011| author = [http://www.youtube.com/user/ChurroEmiliano ChurroEmiliano]| date = 2008-12-31| format = video | work = | publisher = YouTube| quote = After having several Falcon Punch free-for-alls, the CPU wanted to fit in. | |||
}}</ref> [[All is Brawl]] blogger Churro Emiliano documented much of this widely for the first time in a blog post made at the end of 2008. Aside from some advanced techniques (such as [[dash dancing]] and [[DACUS]]ing) and play-styles, the AI have also been reported to "learn" to [[taunt]] a KO by crouch-spamming,<ref name="jerkette">{{cite web| url = http://www.youtube.com/watch?v=diJX703iC0A| title = ...jerkette| accessmonthday = Jan 2| accessyear = 2011| author = [http://www.youtube.com/user/Hoidsa Hoidsa]| date = 2009-05-13| format = video| work = | publisher = YouTube| quote = }}</ref> and to overuse Falcon Punches after humans play several Falcon Punch free-for-alls.<ref name="CEFalconPunch" /> This rumour is demonstrably false for several reasons: | |||
*The AI code has been disassembled and examined during hacking (such as when KingClubber made changes to the AI for Project M) and no evidence of any learning mechanisms have been found. | |||
**Disassembling a Brawl save file also reveals that no AI traits are retained in the save. | |||
*Despite the existence of videos showing the AI having supposedly learned various techniques or strategies, the AI will never under any circumstances learn to perform many actions (all actions documented in the "specific examples" section above exist in the AI regardless of the save file being played). | |||
*There are save files in existence which have logged several hundred hours of versus matches against CPUs, but they do not have any clear, objectively measurable differences in their AI to other save files. | |||
**This applies to both differences between fresh save files that will not have had the opportunity for its AI to learn and to save files with AI which, if the rumour was true, have learned from different players and therefore developed a different playstyle. | |||
*If the rumour was true then replays would desynchronise when the AI learned from the player and changed its behaviour after the replay was saved. Replay data does not contain the inputs of the AI; it contains the player inputs, information such as the stage and characters used, and the random seed of the original match. When the replay is played back, the AI determines its actions from scratch in exactly the same way as it did in the original match (the same actions are produced because all variables that would affect the AI's decisions are identical). Since the replay data has been disassembled to reveal that it does not contain any information about what the AI had learned at the time the match was saved, a learning AI would make different decisions in the replay than it did during the original match, causing the replay to desync, as it would be using its current behaviour (with the adjustments it had made after the original match) and not the behaviour it used in the original match. | |||
*No documented objective tests (for example, only using one move in several matches on a fresh save file and seeing if the AI uses that move significantly more than on a save file which has been used normally) have been performed which have concluded that the learning mechanism exists. | |||
==External links== | ==External links== | ||