Difference between revisions of "Watermelon"

From Robowiki
Jump to navigation Jump to search
Line 20: Line 20:
 
== Movement ==
 
== Movement ==
  
Uses [[Wave Surfing]], implemented in a straightforward manner. <small><small>To the extent that that's possible <sub>grumble grumble</sub>.</small></small> I used [[User:Simonton|Simonton]]'s non-iterative [[Wall Smoothing]] code, and my own future position prediction. It took forever to realize that the future prediction needs to call the wall smoothing function for each imaginary tick, but I finally got it.
+
Uses [[Wave Surfing]], implemented in as straightforward a manner as I can manage. I used [[User:Simonton|Simonton]]'s non-iterative [[Wall Smoothing]] code, and my own future position prediction. It took forever to realize that the future prediction needs to call the wall smoothing function for each imaginary tick, but I finally got it.
 
 
It kind of uses the [[Musashi Trick]], by seeding the bins with a single hit at [[Guess Factor]] 0. The magnitude of the initial seed is such that the first real hit overrides it. Actual hits are smoothed, with a rolling average preferring new data to old.
 
  
 
Bot width is considered, averaging the danger from covered bins. There's no dive-in protection; bot width takes care of that.
 
Bot width is considered, averaging the danger from covered bins. There's no dive-in protection; bot width takes care of that.

Revision as of 00:55, 12 June 2009

/Development

Watermelon
WatermelonStockPhoto.jpg
Author(s) Synapse
Extends AdvancedRobot
Targeting GuessFactor Targeting (traditional)
Movement Wave Surfing
Released 4 June 2009
Current Version 5
Download

Structure

It's built on a modular framework, coded cleanly. If I ever retire this bot I might release the source to the community. Occasionally I'll post snippets for free - see Watermelon/Code. Currently developing in Eclipse. Early on I used an educational IDE called BlueJ (its interface is a UML diagram, very pretty). I test by running battles against a few favorites, including CigaretBH and some others. For more specific testing I post a release to RoboRumble and see what happens.

Movement

Uses Wave Surfing, implemented in as straightforward a manner as I can manage. I used Simonton's non-iterative Wall Smoothing code, and my own future position prediction. It took forever to realize that the future prediction needs to call the wall smoothing function for each imaginary tick, but I finally got it.

Bot width is considered, averaging the danger from covered bins. There's no dive-in protection; bot width takes care of that.

Guess Factors are segmented by velocity and acceleration.

Firing

Guess Factor gun, segmented on velocity and acceleration. Up next: Automatic Segmentation.

Radar

Uses the same radar in melee and in 1v1 conflicts - turns the radar just past the furthest angle that the least recently seen bot could have reached since it was last spotted.

Debug Graphics

WatermelonDebugGraphics.png

Paints waves for enemy bullets and its own bullets, with brighter segments where the bins are fuller. Also marks precisely predicted future positions.

What I'm Working On

Automatic Segmentation

I'm converting both the gun and the movement to automatically select a segmentation, based respectively on the following:

  • whether that segmentation has enough data to be useful
  • how many axes it involves
  • how spiky it is (ratio of maximum to average)

The general strategy behind this is outlined on the old wiki at AutomatedSegmentation.

Saving Data

This one's a tough one - with the sheer quantity of bots in the rumble, and stored data stops being portable pretty quickly. I think the best approach may be just to dynamically alter the rolling average and not save data at all.