StatisticalPatternMatching

Take some pattern-matched angles, then select the most visited angle (like GuessFactorTargeting).
In a word, PatternMatching + MultipleChoice

(Feel free to edit the definition.)

Bots using this Targeting:

• Tron
• Matchup

I'm curious why you say Tron uses this. It really doesn't use patterns to match old situations in its log, it uses DynamicClustering (aka TronsGun), which is more of a log-based system to replace Segmentation / VisitCountStats. ABC has described it as a pattern recognition gun that can work in melee, but it's not a true PatternMatcher any more than Dookious is. Different versions choose firing angles in different ways, but I think most have a form of MultipleChoice for selecting the angle once its found the N closest matches. -- Voidious

• On second thought, perhaps you know all that and the definition at the top just needs to be refined. =) -- Voidious

• What I understood of TronsGun was that it used DynamicClustering to find a 'similar' scan, and then 'played it forward' using normal pattern matching. This simplifies on 'normal' GF-DC guns because it doesn't have to use waves, and adds an element of PatternMatching into it. It is a stat gun because it doesn't match on exact past movement, and is a PatternMatcher because it 'plays it forward'. Thus StatisticalPatternMatching seems right to me. -- Skilgannon

• Ugh - I really disagree with this notion that playing a movie forward to project a firing angle is what PatternMatching is. :-\ (Sorry to be so blunt.) PatternMatching is, to me, when you take the most recent enemy movements (the "pattern") and you match it against a log of previous movements (the "matching"), then find the appropriate firing angle from that point in the movement log (traditionally, the "playing a movie forward" part). A couple other points of note: the current Shadow uses GuessFactors instead of path retrace (says on DynamicClustering page), and neither Chalk nor Lukious (both DC guns) use path retracing either. I don't think that's the defining aspect of a gun. On some level, you could argue that all targeting is looking for predictability (i.e., patterns) in enemy movements, but as a term, PatternMatching has already been coined as a specific implementation that's a bit more literal, IMO. -- Voidious
  • I'm convinced! -- Simonton
  • I wasn't aware that Shadow no longer plays forward. Thanks for the info =). For me, pure PatternMatching means that the data is stored in order of the time recorded, with only enough data to reconstruct their (future) movement. StatisticalPatternMatching keeps more info than is required to rebuild the enemy movement, and tries to find the best match on more data than is required to rebuild the enemy movement. There may also be several angles to shoot at, and these may be filtered using MultipleChoice. Pure statistical targeting does not take the time the scan was made into account, or what scans were made after it. The data is not sorted by time, as is the PatternMatching varieties, although the data may be decayed over time in order to make the fresh scans more heavily weighted. -- Skilgannon
    • Actually Shadow 'does' play it forward, but its not a pattern matcher by my definition (alot of confusion about this), read below. --Chase-san
      • Well, Shadow's not OpenSource, so we only have what ABC says - and he says on the DynamicClustering page that Shadow currently uses GuessFactors. But it's kinda moot anyway, I think it's clear you can have a TronsGun with or without path retracing. -- Voidious
        • Closed-source, right.. hem, forgot. --Chase-san
  • Well, let's not bring another term ("statistical targeting") into the mix =) ... I think all learning guns are statistical in some way. I like this: "StatisticalPatternMatching keeps more info than is required to rebuild the enemy movement." (In fact, my Shaakious was like that.) As long as you're still matching a pattern in the first place, that is, which I still say TronsGun doesn't - at least no more than any form of learning gun does. The main difference between DynamicClustering and VisitCountStats (traditionally GuessFactorTargeting) is that the former keeps raw attribute data and analyzes that before firing, while the latter processes that data on the fly into a data structure (multi-dimensional arrays); both ways will see patterns to some degree with attributes like "time since velocity change" and "distance last eight ticks", but I don't personally think it makes them PatternMatchers. -- Voidious
  • But, the current version of Matchup keeps no more info than is required to rebuild the enemy movement. -- Stelokim