[Home]Recrimpo

Robo Home | Changes | Preferences | AllPages

"There is an art or rather a knack to flying, it involves throwing yourself at the ground and missing." -- Douglas Adams

Recrimpo - My new robot! -- Tango

2.4 - added wall smoothing and MusashiTrick. My tests show it's about equal to the previous version, but i think it will actually do much better, because I wasn't testing against head-on-targeters, or near-wall-segmenters, which are the 2 things that should be improved.

2.31 - when copying and pasting the new movement from a test bot to the dev version i also copied a simpler radar system that i had been testing with for simplicity (as radar only really effects my gun, not my movement), and it ended up being the radar used in the packaged version, so i've put the better one back in. I doubt it will make much difference, but might get a couple of places.

2.3 - new movement again...

2.2 - new movement

version 2.1 - bugfix

version 2.0 - complete rewrite, now with GuessFactorTargeting with Waves

version 1.1 - codesize reduced to 197, now a nanobot!! :-)

http://www.robocoderepository.com/BotDetail.jsp?id=1952

Where did you get the name?

The name is short for "Recreational Impossibilities" which is the heading in the fictional "Hitch Hikers Guide to the Galaxy" under which the above quote comes.

How does it move?

It trys to stay perpendicular, and changes direction randomly.

How does it fire?

GuessFactorTargeting with Waves

How does it dodge bullets?

It doesn't.

How does the melee strategy differ from one-on-one strategy?

It is only a one-on-one bot.

How does it select a target to attack/avoid in melee ?

See above.

What does it save between rounds and matches?

Guess factors between rounds, nothing between matches.

Can I use your code?

Not yet, but keep asking.

What's next for your robot?

A better movement.

What other robot(s) is it based on?

None. This is the propably the bot all my other bots will be based on.


=== Comments:===

How can you fit Wave's in 197 bytes? -- PEZ

I can't. V 1.1 was 197 bytes, this one is a lot bigger. I haven't checked it's codesize, but it's probably a megabot, or maybe a big minibot. The only thing it shares with version 1.1 is the name, really. -- Tango

Ok, it's 1265 bytes, so it's a MiniBot. -- Tango

Ah, but it's still a good size for a Wave bot. Especially since you probably didn't think all that much about code size when you build it. I think 1265 is about the same size as VertiLeach without movement. -- PEZ

I haven't done anything to shrink codesize, no. My coding is usually very bad so i could probably get rid of a few 100 bytes if i tried hard enough. Unfortunately it's ranking seems to have gone down, which is weird. It briefly went up, and then went down again. It's currently lost 20 points. Hopefully with a few more battles it will go back up again. -- Tango

You must have at least one huge bug in that bot. Have you tested the gun in the TargetingChallenge? It's a good measure I think. If you can't get your GF gun to beat Gouldingi's gun then you have a bug in the gun. -- PEZ

I think you're right... i belive it's somewhere in the using the data bit, rather than the getting the data bit, because the data looks about right... -- Tango

Ok, found it. I changed the number of segments just before releasing it, and forgot to change it in the targetting code. Stupid mistake. Should work a little better now. I'll upload it when the repository starts working properly again. -- Tango

But you must still have some bugs in that gun. A GF gun should definately be able to beat Gouldingi's gun, not to mention BeerPoet's. Bug hunting is hard, but you probably must look at the units of code and what you expect them to return from known input and see if it holds tight. And, sometimes it pays to keep the source open, i know that from experience. =) -- PEZ

It's a very primative GF gun. I need to improve the segmentation, for a start. Just give me time, and it'll get there. -- Tango

Wow!! either something is amiss, or you really have Ender's number - 97.6% over the first battle... --Brainfade

Yeah, that happens on a few bots, they crash when they fight me, so i thrash them. Same happened on Wolverine. -- Tango

I think the trig is wrong somewhere, i've corrected a few minor bugs, but it still isn't working quite right. Hopefully it will be all fixed by the end of the weekend. -- Tango

OK, i've been staring at the same code for too long, and its all starting to blur before my eyes. I think it's something wrong in the code that works out of it's going clockwise or anticlockwise, and then using that to work out if it should be a +ve or -ve GF. I can't see what's wrong, so i'm going to open source it, and if anyone has some spare time, could they please look at it. I expect it is something simple, so anyone who hasn't been staring at it for ages will probably spot it in no time. The only thing that will slow them down is the terrible coding style. It's full of debug code, and has very few comments. I might tidy it up later today, if you're lucky. So, if you are bored, please look at /Code. Thanks in advance. -- Tango

For me, the + or - angle thing was by FAR the hardest part of getting a GF gun to work. I spent days and days and days on it, and I have megs upon megs of useless trigonometric drawings and overlays of the battlefield, and tried all sorts of bogus calculations, and when it finally worked, I never touched it again and just let it work. Unfortunately, I think there's now a bug somewhere in there; if you turn on Fractal's gl graphing, you'll note that there's always peaks at the edges of its graphs. This shouldn't be possible, because I set the range from -60 to 60 for debugging purposes. This means I'll have to revisit my wave to angle conversion stuff and figure out what in the heck is going on; maybe this is another of Fractal's performance problems. In any case, I'm not entirely sure how it works, but I can post some code out of Fractal to show you what I have. Once Fractal knows the angle but isn't sure if it's + or -, it does this (this directly follows its angle finding code). 'angle' is the found angle, and 'then' is a class containing robot data on the enemy at the time the wave was fired. The anglePoints method as used figures out the angle from where Fractal was to where the enemy was at the time the bullet was fired. add() simply adds the data to the given bin angle.

        if (then.velocity == 0) {
            //enemy was stopped; add symetrically
            add(angle, wave.getSize(), 1, 0, wave, data);
            add(-angle, wave.getSize(), 1, 0, wave, data);
        } else {
            double newheading = then.heading + 90 * (BotMath.sign(then.velocity) - 1);
            double origin = BotMath.anglePoints(wave.xt,wave.yt,wave.xs,wave.ys);
                // -(origin - newheading)
            double sign = BotMath.sign(BotMath.normalRelativeAngle(newheading - origin));

            angle *= sign;
            add(angle, wave.getSize(), 2, 0, wave, data);
        }

Not sure if that makes any sense, but you multiply the angle by the sign of the the enemy's fixed heading (i.e. if his velocity is negative, turn it 180) minus the angle from the robot to the enemy at the time the bullet was fired. The thing is that angle is not an absolute value; it could turn out positive or negative, and if it's negative this calculation being negative cancels it out. This is why my calculations are so confusing, and so I'm not sure that I'd be any help in debugging Recrimpo. I can post the angle-finding code if you'd like to see it. In any case, good luck with the debugging!

-- Vuen

I'm missing something here. I just subtract the old angle from the new one to find out the delta and then sign() on this delta tells me if it's clockwise or not. But you can also just use LateralVelocity instead. My tests shows that these (delta bearing and lateral velocity) are very equivalent. RoboGrapherBot also had the problem with registering peaks at the edges of the graph a while. In my case it turned out I was using the enemy bullet velocity from the last tick when calculating the max escape angle. Once I fixed that I don't even have to use min() or max() to limit the calculated range. I suggest you check your calculated max escape angles against the escape angle reported by the waves. -- PEZ


Having trouble in the MovementLaboratory? (Since Recrimpo dived some 20 steps when you changed it.) While you work with your gun you migh try using a mirroring movement:
class MirrorMovement {
    AdvancedRobot robot;

    public MirrorMovement(AdvancedRobot robot) {
	this.robot = robot;
    }

    public void doMove() {
        robot.goTo(new Point2D.Double(robot.field.getCenterX() - (robot.enemyLocation.getX() - robot.field.getCenterX()),
	    robot.field.getCenterY() - (robot.enemyLocation.getY() - robot.field.getCenterY())));
    }
}
-- PEZ

As it happens, the 20 step dive was when i added mirror movement (actually it wasn't exactly mirror, it was just based on it. If the current movement doesn't help, i'll try regular mirror)... I've changed it to an oscilator and my tests (i haven't done many, because i'm too lazy ;-)) show it should be a little better. I'm waiting for it to get a stable ranking on the RR atm. -- Tango

I don't know if you're still supposed to be using a stat gun (from what I can tell above you are) but Muffin uses Musashi style movement on you. I ran a match and watched it, and sure enough, it certainly doesn't seem that you can fire anywhere near guessFactor 1.0, if in fact you're firing anything other than direct. Might want to work out why this is happening and you'll probably see Recrimpo jump up quite a large amount in the rankings. 2.4 doesn't seem to have this issue. -- Kuuran

At the moment Recrimpo has a simple pattern matcher, because the GF gun is very buggy and doesn't seem to work. I can't work out why though, I've graphed the GFs and compared it to StatistBot?'s graphs, and although they aren't identical, they do tend to vaguely agree (I put the difference down to segmentation, and suchlike), so I think the error is in the picking the best factor, or in the moving to point in that direction bit... -- Tango

Whatever gun it's using 2.41 only hits direct or near it. Put it up against a MusashiTricked bot and watch. You might want to check to make sure you aren't mixing radians and degrees somewhere, I usually find that's what causes stuff to not spin for me. As for the GF gun, could be something similar, or maybe your calculation for guess factors is a bit off, might want to compare it to existing work and see if there's some important difference. -- Kuuran

Maybe you share the same bug as Falcon against TheArtOfWar TC. That score sure doesn't look like it should for a working GF gun. No matter which way it's segmented. -- PEZ

I'm not sure there is anything specifically wrong, I just think it's a poorly coded gun. I intend to start from scratch in the next few days. I've worked out why the last update on the TC was so small, there's a major bug in the latest segment I added, so I'll fix that and re-run the challenge tomorrow. -- Tango (Of course, the movement isn't great either, it needs some tuning)

But even the poorest coded GF gun should nail down TheArtOfWar down to an index of at least 95 if it's bug free. It's just a matter of figuring out that there's a HUGE spike at +0.2 something. Falcon nails it 100% 4 times out of 5. The 1/5th time there it just doesn't hit it at all. Truly strange. But it's a bug alright. One way to nail bugs is to start from scratch. And, to be honest, I would have started from scratch again if I were you. The code you posted earlier on Recrimpo's gun was a bit on the untidy side of things. =) -- PEZ

Recrimpo's gun was intended more as a learning experience than a competetive bot, and it's worked for that. I have a feeling it might work better with LESS segmentation, in the first axis (the GFs). atm i have 50 bins, and no smoothing, which probably hides spikes, so I will reduce that to 30 and see if it helps at all. -- Tango

No, that lost 2% on the TC. I'll put it back to 50 and try running a TC with a fixed version of the acceleration segment. -- Tango

I've got just over 80%, but there is certainly something wrong against TAOW. If I remove it from the challenge and recalculate all the averages I would be just above the 2 versions of Muffin. I guess I have to watch a few battles now, and see what's going on... -- Tango

If you figure it out, please let me know what it was. -- PEZ

After graphing the visits my gun records, there is a massive (like 95% of all visits) spike at just over head-on (GF 0.04 to be precise). This doesn't match what other graphers find... Time to think now... -- Tango

What happens in Falcon is that it most often finds the correct bin, but then all of a sudden, for reasone completely unknown to me it falls back and uses the middle bin all the time. It's as though it throws away the gathered stats completely and then just never picks it up again. Is it that Recrimpo also does something like that? -- PEZ

Check the tutorial page. I've had an idea. -- Tango


Robo Home | Changes | Preferences | AllPages
Edit text of this page | View other revisions
Last edited May 4, 2006 23:59 EST by PEZ (diff)
Search: