Difference between revisions of "Glacier/Version History"

From Robowiki
Jump to navigation Jump to search
m (Document Glacier 0.2.12)
m (clarify)
 
(4 intermediate revisions by the same user not shown)
Line 8: Line 8:
 
----
 
----
 
<!-- {{subst:rumblestats:meleerumble|ags.Glacier x.x|RumbleStatsDefault}} -->
 
<!-- {{subst:rumblestats:meleerumble|ags.Glacier x.x|RumbleStatsDefault}} -->
 +
 +
* '''0.3.2''' - 25/06/2021
 +
** 1v1: ?
 +
** Fix a bug in BulletHitBullet handling
 +
** Fix a bug in determining intended next-tick location
 +
 +
* '''0.3.1''' - 24/06/2021
 +
** 1v1: {{RumbleStatsDefault|link=http://literumble.appspot.com/BotDetails?game=roborumble&name=ags.Glacier%200.3.1 |rumble=RoboRumble|scorelabel=APS|score=81.16|rank=60th|win=1105|loss=75|plrank=65th|glicko2=N/A|pwin=93.64|vote=0.18|anpp=84.7|score2label=Survival|score2=87.23}}
 +
** Adjust distancing policy for 1v1 to prefer distance=400 (in contrast to melee's "the further the better")
 +
 +
* '''0.3.0''' - 24/06/2021
 +
** 1v1: {{RumbleStatsDefault|link=http://literumble.appspot.com/BotDetails?game=roborumble&name=ags.Glacier%200.3.0 |rumble=RoboRumble|scorelabel=APS|score=72.81|rank=127th|win=986|loss=194|plrank=157th|glicko2=N/A|pwin=83.56|vote=0|anpp=0|score2label=Survival|score2=76.87}}
 +
** Revert the "little experiment" in 0.2.12, but keep the other changes.
 +
** Make the path-wave-intersection danger adaptive in 1v1, based on [[RougeDC]]'s movement's very old and very rudimentary KNN-GF model. Still avoids both HOT and Linear by default, until there's enough adaptive data.
 +
** Borrow [[Diamond]]'s 1v1 fire power selection, as a simple starting point
  
 
* '''0.2.12''' - 19/06/2021
 
* '''0.2.12''' - 19/06/2021
 +
** 1v1: {{RumbleStatsDefault|link=http://literumble.appspot.com/BotDetails?game=roborumble&name=ags.Glacier%200.2.12 |rumble=RoboRumble|scorelabel=APS|score=62.73|rank=322nd|win=835|loss=345|plrank=310th|glicko2=N/A|pwin=70.76|vote=0|anpp=65.3|score2label=Survival|score2=63.19}}
 
** [https://literumble.appspot.com/BotCompare?game=roborumble&bota=ags.Glacier+0.2.12&botb=ags.Glacier+0.2.11 1v1 comparison vs 0.2.11]
 
** [https://literumble.appspot.com/BotCompare?game=roborumble&bota=ags.Glacier+0.2.12&botb=ags.Glacier+0.2.11 1v1 comparison vs 0.2.11]
−
** A little experiment where I change the HOT+Linear avoidance in 1v1 to move as far as possible from danger, instead of the the minimum distance from danger. (As might be expected, it helps dramatically against some bots, and hurts dramatically against some others)
+
** A little experiment where I change the HOT+Linear avoidance in 1v1 to move as far as possible from path-wave-intersection danger, instead of the the minimum distance from danger. (As might be expected, it helps dramatically against some bots, and hurts dramatically against some others)
 
** Also include a little subtle targeting randomization for anti-shielding
 
** Also include a little subtle targeting randomization for anti-shielding
 
** And improve initial turret direction at start of round.
 
** And improve initial turret direction at start of round.
  
 
* '''0.2.11''' - 18/06/2021
 
* '''0.2.11''' - 18/06/2021
−
** {{RumbleStatsDefault|link=http://literumble.appspot.com/BotDetails?game=meleerumble&name=ags.Glacier%200.2.11 |rumble=MeleeRumble|scorelabel=APS|score=69.63|rank=7th|win=390|loss=8|plrank=8th|glicko2=N/A|pwin=97.99|vote=4.77|anpp=94.07|score2label=Survival|score2=47.41}}
+
** 1v1: {{RumbleStatsDefault|link=http://literumble.appspot.com/BotDetails?game=roborumble&name=ags.Glacier%200.2.11 |rumble=RoboRumble|scorelabel=APS|score=66.52|rank=226th|win=835|loss=345|plrank=310th|glicko2=N/A|pwin=70.76|vote=0.28|anpp=69.03|score2label=Survival|score2=68.43}}
 +
** Melee: {{RumbleStatsDefault|link=http://literumble.appspot.com/BotDetails?game=meleerumble&name=ags.Glacier%200.2.11 |rumble=MeleeRumble|scorelabel=APS|score=69.63|rank=7th|win=390|loss=8|plrank=8th|glicko2=N/A|pwin=97.99|vote=4.77|anpp=94.07|score2label=Survival|score2=47.41}}
 
** Apply BulletHitEvent & HitByBulletEvent to cancel out expected energy change for purposes of [[Waves|wave]] detection.
 
** Apply BulletHitEvent & HitByBulletEvent to cancel out expected energy change for purposes of [[Waves|wave]] detection.
  
Line 27: Line 44:
 
**** "much more accurate" means that in a test where Glacier never fires, 1v1 versus [[TrackFire|sample.TrackFire]], it now dodges flawlessly, often with very precise dodges. Glacier 0.2.8 in contrast would make some mistakes. Getting the basics like this to be accurate really counts
 
**** "much more accurate" means that in a test where Glacier never fires, 1v1 versus [[TrackFire|sample.TrackFire]], it now dodges flawlessly, often with very precise dodges. Glacier 0.2.8 in contrast would make some mistakes. Getting the basics like this to be accurate really counts
 
*** Note that when enemy robot scans are interpolated, the wave detection code will generate multiple waves, one for each tick where firing could have occurred. The wave then has an associated probability of existing, which danger score caused by the wave is weighted based upon.
 
*** Note that when enemy robot scans are interpolated, the wave detection code will generate multiple waves, one for each tick where firing could have occurred. The wave then has an associated probability of existing, which danger score caused by the wave is weighted based upon.
−
*** Associated with this, the "[[GoTo|go-to]]" path simulation in the [[Minimum Risk Movement||minimum risk movement]] engine is replaced with a more accurate version.
+
*** Associated with this, the "[[GoTo|go-to]]" path simulation in the [[Minimum Risk Movement|minimum risk movement]] engine is replaced with a more accurate version.
 
** Movement intent is now accounted for when aiming.
 
** Movement intent is now accounted for when aiming.
 
** Keep dodging anything that needs to be dodged in the moments after all other bots are gone.
 
** Keep dodging anything that needs to be dodged in the moments after all other bots are gone.
−
** Commentary:
+
** This version appears to do worse in 1v1 rumble than 0.2.8, despite doing better in melee rumble. I presume this is due how, while it dodges HOT+linear more reliably, it also is more predictable versus certain other targeting. This is fine, since 1v1 is only a diagnostic measure of where things can be improved for now.
−
*** This version does measurably worse in overall 1v1 rumble than 0.2.8, despite doing better in melee rumble. I presume this is due how, while it dodges HOT+linear more reliably, it also is more predictable versus certain other targeting. This is fine, since 1v1 is only a diagnostic measure of where things can be improved for now.
 
−
*** <s>Interestingly, this version is now a real huge [http://old.robowiki.net/cgi-bin/robowiki?ProblemBots ProblemBot] for [[Tron]], with -60.82 [https://literumble.appspot.com/ScoreExplanation KNNPBI] as of this writing.</s> <- This is because Tron is currently broken in Robocode 1.9.4.2 & OpenJDK 11 for some reason.
 
−
*** It's also by a hefty margin the lowest overall ranking bot in 1v1 rumble to have >0.7% in the "Vote" column, which counts for something in my eyes.
 
  
 
* '''0.2.9''' - 17/06/2021
 
* '''0.2.9''' - 17/06/2021

Latest revision as of 12:12, 25 June 2021

Glacier Sub-pages:
Glacier - Version History - Hypothermia - GlacialHawk
  • 0.3.2 - 25/06/2021
    • 1v1: ?
    • Fix a bug in BulletHitBullet handling
    • Fix a bug in determining intended next-tick location
  • 0.3.1 - 24/06/2021
    • 1v1: RoboRumble ‒ APS: 81.16% (60th), PL: 1105-75 (65th), Survival: 87.23%
    • Adjust distancing policy for 1v1 to prefer distance=400 (in contrast to melee's "the further the better")
  • 0.3.0 - 24/06/2021
    • 1v1: RoboRumble ‒ APS: 72.81% (127th), PL: 986-194 (157th), Survival: 76.87%
    • Revert the "little experiment" in 0.2.12, but keep the other changes.
    • Make the path-wave-intersection danger adaptive in 1v1, based on RougeDC's movement's very old and very rudimentary KNN-GF model. Still avoids both HOT and Linear by default, until there's enough adaptive data.
    • Borrow Diamond's 1v1 fire power selection, as a simple starting point
  • 0.2.12 - 19/06/2021
    • 1v1: RoboRumble ‒ APS: 62.73% (322nd), PL: 835-345 (310th), Survival: 63.19%
    • 1v1 comparison vs 0.2.11
    • A little experiment where I change the HOT+Linear avoidance in 1v1 to move as far as possible from path-wave-intersection danger, instead of the the minimum distance from danger. (As might be expected, it helps dramatically against some bots, and hurts dramatically against some others)
    • Also include a little subtle targeting randomization for anti-shielding
    • And improve initial turret direction at start of round.
  • 0.2.11 - 18/06/2021
    • 1v1: RoboRumble ‒ APS: 66.52% (226th), PL: 835-345 (310th), Survival: 68.43%
    • Melee: MeleeRumble ‒ APS: 69.63% (7th), PL: 390-8 (8th), Survival: 47.41%
    • Apply BulletHitEvent & HitByBulletEvent to cancel out expected energy change for purposes of wave detection.
  • 0.2.10 - 18/06/2021
    • MeleeRumble ‒ APS: 69.71% (6th), PL: 390-8 (8th), Survival: 48.35%
    • Revert to 0.2.8
    • Replace virtual bullets with more accurate precise intersection waves
      • This was based upon importing code I wrote around May 2012 or so, yet never actually became part of a released bot.
      • Still only tries to dodge HOT and linear targeting like before, but now done with much more accurate waves.
        • "much more accurate" means that in a test where Glacier never fires, 1v1 versus sample.TrackFire, it now dodges flawlessly, often with very precise dodges. Glacier 0.2.8 in contrast would make some mistakes. Getting the basics like this to be accurate really counts
      • Note that when enemy robot scans are interpolated, the wave detection code will generate multiple waves, one for each tick where firing could have occurred. The wave then has an associated probability of existing, which danger score caused by the wave is weighted based upon.
      • Associated with this, the "go-to" path simulation in the minimum risk movement engine is replaced with a more accurate version.
    • Movement intent is now accounted for when aiming.
    • Keep dodging anything that needs to be dodged in the moments after all other bots are gone.
    • This version appears to do worse in 1v1 rumble than 0.2.8, despite doing better in melee rumble. I presume this is due how, while it dodges HOT+linear more reliably, it also is more predictable versus certain other targeting. This is fine, since 1v1 is only a diagnostic measure of where things can be improved for now.
  • 0.2.9 - 17/06/2021
    • MeleeRumble ‒ APS: 69.06% (8th), PL: 389-9 (10th), Survival: 46.68%
    • A quick hack of a movement tweak. Seems it made things slightly worse in melee (though 1v1 improved slightly?). Will be reverting.
  • 0.2.8 - 16/06/2021
    • MeleeRumble ‒ APS: 69.67% (6th), PL: 390-8 (8th), Survival: 47.09%
    • Fix a bug in targeting. The 'acceleration' feature for KNN search was badly miscalculated.
  • 0.2.7 - 29/10/2009
    • 16/06/2021 results: MeleeRumble ‒ APS: 68.93% (8th), PL: 389-9 (10th), Survival: 42.3%
    • 30/10/2009 results: MeleeRumble ‒ APS: 69.32% (4th), PL: 278-4 (5th), Survival: 90.51%
    • Rewrite the radar, to create a truly optimal melee radar. First of it's kind that I know of.
  • 0.2.6 - 29/10/2009
    • MeleeRumble ‒ APS: 69.14% (4th), PL: 279-4 (5th), Survival: 90.23%
    • Don't use gun turn to assist radar sweeps except 1) at start of battle, and 2) when changing radar directions
      • Using gun to assist radar sweep was previously being bad in the circumstances where the radar just spun in one direction, because it caused the gun to aim in poor directions where it would take longer to aim again from.
  • 0.2.5 - 28/10/2009
    • MeleeRumble ‒ APS: 69.04% (4th), PL: 277-5 (5th), Survival: 90.55%
    • Add the optimal-radar-starting-direction back again, with a stupid bug fixed
    • Don't start moving or recording data until all enemies are known
    • Use the gun/body to to increase the size of radar sweeps ;)
  • 0.2.4 - 27/10/2009
    • MeleeRumble ‒ APS: 69.21% (4th), PL: 279-3 (3rd), Survival: 90.04%
    • Revert to 0.2.2, and then make PIF wall boundary checking more lenient like 0.2.3
  • 0.2.3 - 25/10/2009
    • MeleeRumble ‒ APS: 69.17% (4th), PL: 278-3 (4th), Survival: 90.07%
    • Re-release of 0.2 to check melee score drift
  • 0.2.2 - 24/10/2009
    • MeleeRumble ‒ APS: 68.88% (4th), PL: 277-3 (4th), Survival: 89.64%
    • Revert radar changes of 0.2.1
  • 0.2.1 - 16/10/2009
    • MeleeRumble ‒ APS: 68.28% (4th), PL: 278-3 (4th), Survival: 89.06%
    • Use 'turn towards center' for initial radar direction
    • Use 'shadow' style Play-If-forward for now
      • Pure BFT approach appears to be notably faster in my tests, but my implementation is locking up from time to time. Not sure it's worth the effort.
  • 0.2 - 26/9/2009
    • MeleeRumble ‒ APS: 69.21% (4th), PL: 277-3 (4th), Survival: 90.31%
    • Quick hack to dodge linear targeting in addition too.
  • 0.1 - 25/9/2009
    • Melee APS diffs:
    • MeleeRumble ‒ APS: 68.95% (4th), PL: 279-4 (4th), Survival: 89.64%
    • Uses GlacialHawk 1.10 targeting.
    • Uses old movement HOT-dodging movement developed just prior to begining the GlacialHawk experiments.
Retrieved from "http://robowiki.net/w/index.php?title=Glacier/Version_History&oldid=56748"