Talk:Diamond

I can't believed it, it almost the same thing I'm doing in my TheRiver! Anyway, nicely done! Here are my result that I just ran if you want (EDIT: more, and the debugging graphics is the exactly same as one I'm creating) » Nat | Talk » 06:16, 15 May 2009 (UTC)

Rank	Name	Score	Surv.	Bonus	Bullet Dmg.	Bonus	Ram	Bonus	1st	2nd	3rd
1st	voidious.Diamond 1.0	14361 (13%)	10850	540	2807	164	0	0	6	9	5
2nd	abc.Shadow 3.83c	13453 (12%)	10100	630	2633	89	1	0	7	3	5
3rd	rz.Aleph 0.34	12804 (12%)	9600	270	2861	73	0	0	3	4	5
4th	abc.Tron 2.02	12261 (11%)	8450	630	3052	127	2	0	7	3	3
5th	kawigi.mini.Coriantumr 1.1	11909 (11%)	450	360	2998	101	0	0	4	3	3
6th	florent.XSeries.X2 0.7	9472 (9%)	695	90	2349	84	0	0	1	4	2
7th	ags.surreptitious.MiniSurreptitious 0.0.1	9411 (9%)	6550	0	2785	68	8	0	0	4	5
8th	tzu.TheArtOfWar 1.2	9240 (9%)	6750	90	2355	42	2	0	1	3	4
9th	abc.tron3.Tron 3.11	8050 (7%)	5400	360	2245	42	4	0	4	2	1
10th	davidalves.Phoenix 1.02	7443 (7%)	5550	180	1679	32	2	0	2	0	2

Does Diamond use a flattener in 1v1? I haven't seen any mention of one.

Yep, it does. In spirit, the decision criteria and configuration are similar to Dookious. Maybe the biggest implementation difference, besides VCS vs DC, is that Diamond has a more sophisticated system to decide when to enable it. Dookious has "tiers" - like never enable it in first few rounds, then hit % threshold is 9% until round 10, 10% until round 20, etc. Diamond just uses a single hit % + a margin of error based on how much data he's collected, similar to an election polling formula.

If you want to check out the code, see "initSurfViews()" in voidious/move/MoveEnemy.java. The last 3 are for the flattener. A "view" is a kd-tree + all associated configuration data, like parameters to decide on k, max data points, decay rate, and hit % threshold before it's enabled.

how does a flattener work with wave surfing

see Flattener. Instead of moving to the places with the least amount of danger, move to the places you have been less often. Ofcourse depending on your configuration and segmentation. In case of a wavesurfer, recording every fired bullet instead of only the times the enemy really hit you, would result in a 'flat' movement => flattener. The biggest problem with a flattener is the decision when to switch to the flattener and when not. (pro tip: search the wiki for info)

Well, sorry to ruin Christmas break for you but...

I was watching a battle with Diamond, and I noticed that there were bullets in the area marked safe. The shadow was quite large, (large enough for me to be able to say it was one shadow) so, I don't think it's possible that the bullet that supposedly made the shadow hadn't passed yet (unless there were two shadows near each other and a third bullet that would have connected them had just been fired). In short it appears there is a bug with your bullet shadows.

Thanks for the notice! If it was a big shadow, it's probably a bug with merging shadows or angle normalization. I think that stuff's pretty tight so it might be tough for me to duplicate it enough to debug it. But I'll give it a shot when I have some time. Thanks. :-)

Oh, and finding a bug for me to fix in Diamond doesn't ruin anything, it's great. If I can't get to #1 with bug fixes, I have to actually think of innovative new features. ;)

Just to be clear - did your see if the bullet in the safe area hit one of my bullets before it reached me? Because that would not be a bug, it would just mean Diamond knew beforehand that one of his fired bullets would intersect any bullet in that range.

Yeah, I'm pretty sure it didn't but I MIGHT have missed it. I think I have a bug with bullet shadows it Gilgalad, and to test I just ran two 1000 round battles against hawkOnFire with and without bullet shadows. HawkOnFire's bullet damage without bullet shadows was 103 and with bullet shadows it was 192, so it appears I will be suffering with you...

Still searching for the similar bug in Gilgalad. I just found that it occurs in my intersection calculations (rather than the merging) which will make it more annoying to find. Perhaps I should have tested for bugs before optimizing those...

I discovered my bullet shadows didn't contain any 'impossible bullets' when I moved my bullets a tick forward - how about trying an empirical test for having your bullets a tick ahead/behind where you think they are?

So I tried that, and and fixed the possible "bullet start and end positions are in the wave, but it passes outside of it" bug. Still short of perfection (though there is a measurable improvement against HawkOnFire over 1000 rounds). Would someone mind explaining how robocode deals with bullets that hit both the enemy robot and their bullet on the same turn?

I'm not sure how Robocode deals with simultaneous impacts. I think you'd have to dig in the source code to find out exactly what happens.

Just in case anyone else is curious, the bulletHitRobot event is dealt with first.

I guess after 17k battles and still no 1. in melee general it is time to spring the congratulation cards :)

Congratulation for taking the melee Crown. Thats a hell of a bot you put there together.

Take Care

Congradulations Void! -Jlm0924

I noticed only now. Diamond took the PL throne away from DrussGT.

Thanks :-) I'm very pleased with my PL improvements, but the truth is I still get 49.3% vs DrussGT over 1000 battles and Tomcat has been kicking my butt a lot lately. So it's still very much a crown in dispute, but I'll keep working on it. =)

Tomcat actually had the highest PL for about a day or two. It's a very close running.

Pretty sure I've got it now. =) (Well, for now.) I'm running 1000 seasons vs Shadow, DrussGT, and Tomcat to be sure. After ~250, the lowest score was over 51% (Tomcat). I'll post results soon.

Of course, it's after a change to improve APS that I get the PL boost I was looking for. :-P

Hey, remove Tomcat from yours testbed - it's just little sweet kitty!:)
Enough to hurt my defenseless kitty!:) He even can not do multiply wave surfing right:)

Pairs won: 100%. It's beautiful.

Very nice. Good job, now just take the crown.

Very nice there! Note to self: Better throw some of my new top-secret bulletpower research into Scarlet, to attempt make it more challenging for folks to get 100% in PL ;)

Some Diamond 1.6.17 scores over 500 battles:

• vs Shadow 3.83c: 53.08
• vs DrussGT 2.4.4: 52.36
• vs Tomcat 3.46: 51.25

For the first time since 2004, I have assembled a testbed, but now I notice something I had not though about on beforehand. Some bots like Phoenix and ad.Quest, save data and that could have influence on the outcome. Do you use data-saving bots in your testbed and if so, how do you arrange that every version starts with a clean sheet against those. I want to alter my 40-bot testbed anyway, because 1 run (35 rounds) takes about 30-35 minutes. I still use a single core 2.66 GHz P4, although I have a i7 8-core laptop from work available. --GrubbmGait 14:39, 11 September 2011 (UTC)

I do use data saving bots and am just ignoring the issue right now, actually, which is perhaps stupid. But it should be pretty easy to modify RoboResearch to delete the robot.database and .data after every season, so they get rebuilt and saved data is cleared. That would be a very nice option to have. Are you using RoboResearch?

One little note, is another option would be perhaps to modify RoboResearch to enable the option in new versions of Robocode, to obfuscate/randomize enemy names?

I thought that would return a consistent name for the same bot each time, just preventing you from pre-load bot-specific behavior? I started browsing the code just now to try and say for sure, but I wasn't able to uncover the answer as quickly as I'd hoped...

A wait... it looks like the "anonymous names" are just like "#1" and "#2" according to their position in the participants in the battle... That's unfortunate, because because unlike randomized ones, that would really confuse data saving robots by giving them polluted results.

See "getAnnonymousName()" in net.sf.robocode.host.RobotStatics

On second thoughts, it is not really an issue. Most bots that save data do it for each version separate. Only bots like cx.BlestPain that save data without version information I should remove from my testbed. And ofcourse regularly removing the contents of 'working_dirs'. Any tips about contents of the testbed? I now selected 10 of my 20 worst White Wales, 10 bots out of the top-30, and 20 bots ranging from 40-140 with a PBI close to zero. Still busy with 30 seasons for 30 minutes each, sigh. --GrubbmGait 19:24, 11 September 2011 (UTC)

These days I generally go for big test beds generated by BedMaker. Like right now, my main test bed (when working on APS) is 100 random bots that Diamond scores <= 80% against. Your test bed sounds well put together to me - improving against problem bots is good, and having a diverse set of other bots helps to ensure you're improving in general instead of just specializing against a different set of bots.

Does Diamond still has more than 100 bots it scores less than 80% against? That's a bit disappointing isn't it ;-)

Ok, me and my big mouth. Diamond has around 200 bots that qualify for the above statement, GresSuffurd still has around 300. But running such testbeds for 30 seasons or so will cost you alot of time, certainly if you like to improve gradually (like me) instead of big-bang.