ZoomTargeting

ZoomTargeting is a hybrid of StatisticalTargeting and a new form of PatternMatching and should combine the best of both worlds. It uses SegmentedData (currently 6 dimensions) and gets this data using a variant of Waves.

Its StatisticalTargeting properties should make sure that

• it is very good at long term learning and continues to improve even in very long matches
• it can exploit weaknesses such as spikes in the enemies MovementProfile
• it can do a good educated guess when PatternMatching fails (when there's not enough data in the microsegment)

Its PatternMatching properties ensure that

• it adapts very fast to new enemies or situations.
• it can exploit weaknesses such as dodging and other types of repetitive movement
• it can get a reasonably high hitrate early on in a match if there is no stored data
• it can be more accurate than StatisticalTargeting (if there's enough data in the microsegment)

The new thing is that this targeting method basically zooms in (more PatternMatching) or out (more StatisticalTargeting) depending on the amount of data it has and the zoom level's segment size. Currently I use 8 zoom levels with increasing segment sizes. PatternMatching takes place in what I call the MicroSegments. In these tiny segments even only 1 previously reported angle is enough to deliver a firing angle with a high probability of hitting the enemy. StatisticalTargeting takes place when you zoom out, in the MacroSegments?. The segments here are much larger and contain more data to do statistical calculations with. Learning speed increases when the zoom level increases.

The PatternMatching algorithm that I use differs greatly from traditional PatternMatchers?. You might even argue if it is PatternMatching at all. I call it PatternMatching because it matches to a very specific situation. As a result it performs very well against bots that other PM guns perform well against :-). A situation is basically a very small segment with a lot of dimensions. Currently I use 6 dimensions, but if I can solve a memory issue I will use even more. I might even try a dimension that uses the movement history to make it a more official pattern matcher :-). In practise, this algorithm is a bit less accurate against truly predictable movers like SpinBot or PatternBot, but more effective against random movers. And since none of the top bots are truly predictable.........

There is a nice side-effect to using MicroSegments: I keep track of the number of Segments ('situations') that are used. This number increases more rapidly as the enemy has better movement. I am guessing this Segment count (say after 1000 rounds) could be a nice alternative Movement Index.

-- Vic Stewart

How far are you in developing this awesome gun?

• The next step is to implement the Statistical calculations of the macrosegments.
• decide whether to fire a Wave every scan or along with a bullet.
• better memory management

Yeah, yeah, yeah..... You talk the talk, but does it walk the walk?

• So far the results are very encouraging. I was unsure if the new PatternMatching algo would work, but it does! For instance, it beats all but one of the samplebots in a 10 round match, without stored data, whilst being stationary. (only TrackFire is a problem but that's due to a known bug)
• I tested against one of the challengers in the PatternMatcherChallenge and with just the new PatternMatching algo I got its index down from 32% to 25%. I wonder what will happen when I add the StatisticalTargeting part....

Sounds great, can I help in anyway?

• Yes you can :-) I have two issues:

---

/MemoryIssue /MultipleChoiceIssue

---

I am developing the gun in various stages:

/stage1 /stage2?