Robo Home | SandboxDT | Changes | Preferences | AllPages

Phew! I think I will even wait a few minor revision numbers on Marshmallow before even downloading that beast. =) -- PEZ

Many minor and one major revision later Marshmallow still don't stand a chance against SandboxDT. -- PEZ

Times have changed and while Marshmallow still doesn't beat DT, it now wins some 10/50/100 round battles now and then. (June 2 2003). -- PEZ

It seems that version 1.81 of SandboxDT has a memory leak bug or some such. It's almost impossible to run it in 5000k+ round battles. I first thought it was a problem with Marshmallow but its the same with any bot you put up against it. -- PEZ

This is true and will be fixed in the next release - it seams that the Java garbage collector can not recover objects a robot has referenced at the end of each battle - in the case of DT it uses an ArrayList of Wave objects. The next version will make this ArrayList static and will be initialised at the beginning of each round (this solves the issue). -- Paul Evans

Um... I just looked at the EternalRumble stats, and SandboxDT is sitting at #1 on melee... (*:-(* WE STILL LOVE YOU TRON!) [edit] Hey, I ALSO just noticed that Tron hasn't yet been updated to 1.5 in the melee. How come? -- Vuen
I haven't been following the recent discussions on this page until now, but I have been experimenting with movement systems a lot recently (on paper and with code), and I have finally perfected a nifty little puzzle for all of you -- behold, TheSpikeConundrum.

Question 5: How does it work? ;)

-- nano

Very well. -- jim

Questionably (I'm still a little skeptical, your strength before was what you had between 0 and 1, not keeping 0 and 1 even. You seem to be solving Fhqwhgads' problem, not your own). -- Kawigi

In response, I must hasten to point out that Spike is merely an entertaining experiment, and by no means a solid grounds in and of itself for a serious movement system. I mean, try any other bot against him! They should fairly rip him to shreds. :) Pure-statistical guns that do not segment on acceleration will fall for Spike's trap, however. -- nano

The good news is that bots with shorter attention spans (bots who discard old data quicker) will see the (bigger) spikes appear and will be able to take advantage of it. This contrary to bots who average all data who'll see a flatter profile with two spikes. The hunt for the optimal 'attention span' is on :) We might start looking into the direction statistical relevance of differences ... -- FnH

Thank you Paul,did you run them 1000 rounds,and then run 1000 rounds again(with pre saved data)?I think the next 1000 rounds BlestPain could do better than DT(maybe can beat SandboxDT 2.01,in my once testing it did),what happened when DT save or load datas?.I have not plan to add the PM gun to BlestPain this time except I have a good enough way to evaluate the Virtual Guns.I am glad that you say BlestPain's dance is nice,did you noted that he has two different Victory Dance,the one most time you could see is design from a song of Lacrimosa which named "not every pain hurts",another is design from my being excited.Hope you like both of them,and hope you like its ramming running too(It may never happen when he vs DT as DT never be disabled). -- iiley

I think I ran a 5000 round battle - either way the issue is moot - DT 2.11 is released - it has a score advantage of 52.5% to 47.5% after training (better DT results in the first 800 rounds). Sorry. -- Paul Evans

People must be getting better, you are releasing versions much faster now. ;-) -- Tango

Either that, or he's messing up more often. But this time he said he's improved both his targeting and his movement. If he knew ways to improve it before, why didn't he do them? Or have we brought his weaknesses to light for him? -- Kawigi

I'd go with Tango's analysis. I have always suspected Paul had quite a few improvements up his sleave. They haven't really been called for before. Now iiley's movement and Kawigi's guns asks questions DT hasn't been asked before. It's way cool! -- PEZ

The targeting improvement was as a result of looking at iiley's movement - it was a simple change in the segmentation - no extra guns/ideas but enough to give me the edge. Up until the last version changing the movement was always a risky problem - but with the movement adaptor code most of the fine tuning is done automatically so changes are easier. The change here causes problems with bots collecting stats every tick - DT 2.11 has a tendency to move on bullet fire giving a false profile for Sedan - for Sedan to get accurate targting stats he will have to run 11 to 16 times more rounds. I keep nothing pre-coded ready for realease - it was simply more urgent to get a realease this time for obvious reasons. -- Paul Evans

I've only done 500 rounds, so this isn't a particularly acurate result, but it is interesting:

 1st: pe.SandboxDT 2.01	43781	13600	2720	24295	3165	0	0	273	228	0
 2nd: pe.SandboxDT 2.11	38846	11400	2280	22599	2565	1	0	228	272	0

I would understand the new bot being no better than the old one, but being worse...? Is this simply changes designed to make it better against some bots will automatically make it worse against others? I will run a longer battle tomorrow, but 500 rounds should be enough for a guess. -- Tango

While I don't know if this release is better or worse against FloodHT than the last one, I did notice that this release does worse against HT than against Sedan, which was noticeably not the case with the last release. And, it seems, I've actually convinced Paul to take my view on things to a degree concerning bullet-dodging :-p (since he was always an advocate of pure, good, unpredictable random movement). Regardless of whether my information gathering is correct or not, it would appear that this edition "flattens" itself against FloodMini significantly better than the old version (I think FloodMini tends to hit more often when he fires as his opponent is accelerating/decelerating, which caused him to hit in the +1.0 region and made more of a mountain of DT's profile). This may or may not be due to this. I'll do more looking probably later. -- Kawigi

Most versions of DT has reacted on bullet fire for all I know, at least up to bullet travel times below 27 or so. What I have heard Paul say is that there are risks connected with doing any such dodging since it makes it harder to keep your profile flat. I think that what he's now saying is that with his profile adapter in place this risk is more or less eliminated for him and thus DT can do this without messing up its profile. Note that he also seems to be able to exploit some aspect of iiley's movement. Wanna bet if this aspect might have to do with reacting on bullet fire? =) For when it comes to improving your bot against some bots and getting a punishment against others, I think that this is very hard to avoid. But at least you shouldn't worry about beating the old version of your bot. You should seldom meet that one in the leagues anyway. -- PEZ

Tango, I can confirm your results - I ran a 5000 round battle last night and it seems that the new DT 2.11 movement is weaker than the 2.01 movement - especially with the segmented stats. It looks like iiley's bot does not take advantage of this, but the new segmantation is better against BlestPain - I think new or old movement was about the same against BlestPain.

Kawigi, I did some graphs with FloodMini 1.3 - the result is that the peaks etc firing a wave each tick are different to the peaks when a wave is fired when you fire a bullet. A bot collectig stats each tick will now be at a disadvantage - it will not shoot in the best place. Be sure to fire waves only when you fire when graphing DT.

PEZ, you are correct, up until the last version I had code which may change direction when it detected enemy fire at closer distances. It looks like this DT's movement is weaker against segmented guns collecting data on bullet fire only. DT's guns are well adapted to learning fast in these conditions - others may need a few rounds to get there :) -- Paul Evans

The new(?) habit of sharply changing strafe angles is curious. Is this to create false peaks and bait, or is this something the movement adapter does to flatten dips in your profile without slowing down (and thus becoming fodder for acceleration segmentation)? -- Kuuran

You will see 4 'strafe' angles - a neutral angle designed to keep the same distance form the opponent, a sharpish retreat angle to get to an sensible working distance from the opponent if they are to close, a slight attack angle to get to the working distance if far away, and a sharp attack angle to ensure I kill an opponent if they are low on power (to collect the bonus). Only the slight attack angle is new (I think it was introduced in version 2.01) - it keeps DT out of the corners. The main reason for the retreat angle is to decrease the probability of a hit (a retreating bot gives a smaller hit angle than one that stays neutral or approches). I think anything other than a 90 degree straight line will make a bot easier to hit - the angles are not designed to flatten the curve for acceleration segmentation because, if anything, they become fodder for linear guess factor aiming, or segmentation based on the 'attack angle' of the bot.

I hadn't noticed the angles as prominently before, anyway. As for the segmentation based on attack angles, sure, but no one does that compared to segmenting on acceleration ;) Thanks for the explanation, appreciate it. :) -- Kuuran

Oh yes, various versions of my bots do this and it might be something I will let GloomyDark do. I think Kawigi has experimented with this as well. -- PEZ

LoL?, of course I have. You'll often see my bots changing direction laterally and also doing a significant (usually 60 degree) turn at the same time to maintain an 'attacking' or 'retreating' angle. Since many of my bots put effort into distance control, I'm doing this quite often, especially if my opponent disagrees about the distance. -- Kawigi

Robo Home | SandboxDT | Changes | Preferences | AllPages
Edit text of this page | View other revisions
Last edited May 4, 2006 22:25 EST by GrubbmGait (diff)