View source for Talk:Flattener

In theory, it should be possible to hit a flattener more often than a random movement because you know the flattener will avoid where it last was.

In practice, against top tier RoboRumble guns, which would be harder to hit, a typical flattener that logs where it last was and avoids it, or a type of random movement that moves to a random guessfactor each wave?

Thanks!

I have never had much success with flatteners but exploiting it is unethical imo. The logic of flattener is that your opponents targeting is learning and you can simulate what your opponent does fairly well keeping your profile fairly flat. On the other hand, assuming that your opponent assumes that your targeting is learning is not that reliable as it might just be that your gun isn't strong. Random movement guarantees that your opponent's gun will be as good as random targeting while flatteners can actually keep an advanced gun's hit rate below random.

I don't understand, what's unethical about it?

Sorry, my wording wasn't clear: in order to have a better score against all flatteners, you have to exploit it like [LittleBlackBook] does(which I don't really approve btw) because otherwise your targeting will most likely be average against flatteners.

Ah, that makes sense. But against the guns in ScalarN or DrussGT, would a standard flattener or random movement work better?

I'm pretty sure a good flattener would outperform random movement after seeing ScalarBot with a PWIN of %100 percent only with a tick flattener but WhiteFang's flattener would most definitely do worse than standard random movement =)

Sure. In Diamond, it's all in voidious/move/MoveEnemy.java. Just comment out some of these:

    addView(simple);
    addView(normal);
    addView(recent);
    addView(recent2);
    addView(recent3);
    addView(recent4);
    addView(recent5);
    addView(recent6);
    addView(lightFlattener);
    addView(flattener);
    addView(flattener2);

Oops, sorry, you'll also want to remove the hit percentage thresholds, like:

        .setPaddedHitThreshold(FLATTENER_HIT_THRESHOLD)

Flattener is only enabled at a certain hit percentage, but you want it always on.

Just making sure, this will also disable normal hit logging, so it will only dodge places its been before, not where its been hit? Will I also have to add something to its move history so it has some at the beginning?

I don't think that will be a problem. I just checked and it's hard-coded to dodge head-on targeting (GF=0) when it doesn't have any stats. It doesn't pre-load the trees or anything, it's just a special case in the code.

And yep, all the data sets are configured there, hits and visits are both "first-class citizens". ;)

Comment out "some"?

I'd leave just flattener and flattener2. The lightFlattener is pretty new and enabled at a lower hit percentage threshold, so I dunno, up to you.

Has anyone figured out a statistical test which gives a good prediction of the state of an enemy surfer's flattener? It seems that it might be useful in targeting to figure out if they are using a flattener or not.

Combat measures the peak in kernel density estimation. Low peaks are a sign of curve flattening. But it is used only for energy management. Low power bullets against flatteners, max power bullets against predictable movements.

And I'm using it in melee/team only. Doesn't work against surfers since they create peaks to screw up your data on purpose.

While you're at it, detecting if they're surfing at all might be good to have. One thought I've had on that is detecting if their attack angle is relative to you or the wave sources.

Current virtual gun systems seem to do a pretty good job deciding if they are surfing or not: until they start flattening.

Combat heading is not perpendicular to either bots or waves. It's heading aims to maximize MEA instead.

How do you determine the best attack angle to maximize MEA?

According to MEA formula, heading should be perpendicular to a bullet path targeted using head-on targeting. It IS perpendicular movement, but keep moving straight until the wave hits, instead of turning as in orbital movement.

When there are 2 or more waves flying in your direction it becomes more complex, but Combat assumes there is only 1 wave.

If close to walls, then heading should create the longest path which would hit the wall at the same time the wave hits. Combat calculates which point in the wall coincides with waves hitting it with PPMEA. It could also be calculated with precise MEA.

The side-effect is dynamic distancing becoming hard to perform.

"One thought I've had on that is detecting if their attack angle is relative to you or the wave sources."

That doesn't work with all robots, some surf relative to the enemy robot, rather then the wave source (such as Seraphim).

So it's obviously not perfect, but I still think it's a useful signal. A reliable surf-detection system would probably need quite a few of them.

Has anyone made a targeting system capable of hitting a good sandboxFlattener (Perhaps a top bot with surfing disabled and flattener enabled by default) better than top bots hit strong random movement? This would be a good start for hitting combined surfing and flattening, which is driving hitrates of top guns into the ground ( Below 10, sometimes even 9%).

What makes you think you could predict a curve flattening movement better than a "random" movement? A good curve flattening movement is a lot like random with a forced even distribution, like Java's Math.random(). It should always have a higher entropy than a naive random, I think.

It seems like its theoretically possible to get a higher hitrate (without simulating their movement) because you know that they will move away from where they have moved before, while in random movement there shouldn't be a correlation between where they have moved and where they do move.

It's possible. A random with an even distribution is less predictable than a skewed random, but it's only a hunch that curve flattening resembles the former.

But while knowing where they won't be is something, and could increase your hit rate, it's really quite far from knowing where they will be. You could know where they won't be 100% of the time and never hit them. :-) And of course there is a correlation in random movement between where they have moved and where they will move, since they are determined by the same algorithm. Isn't it more useful to know where they will be vs where they won't be?

To know where a flattener will be, you'd have to figure out quite a lot about their stats to be able to shoot at the least visited reachable point with any accuracy. And that truly is the holy grail of Anti-Surfer guns, since you could also destroy a normal surfing system if you knew that much (ie, not just a flattener). When surfing first came along (~2005), a lot of folks thought this would happen almost immediately, but so far they are still wrong.

Knowing where they won't be increases your average hitrate over time, by limiting the places they could be. Imagine a targeting system which fires outside of their MEA vs one that always fires in the MEA.

Oh for sure. My most sophisticated attempts at fancy Anti-Surfer guns were along these lines - assume they will go to the safest reachable point on all of the waves in the air at fire time, using precise prediction to figure out reachable points, then restrict my choice of firing angles to the range of remaining reachable points.

What I really want is a mask to apply to the whole range of firing angles, deduced from the probability of all possible enemy movements as they surf the waves in the air. I tried a very (very!) crude version of this once in Dookious.

Oh, and I don't mean to sound too discouraging! I actually think of improvements in Anti-Surfer targeting as one of the remaining frontiers in Robocode, and probably one of the nearer ones. I've spent quite a lot of time trying Anti-Surfer stuff myself, including stuff that precise predicts enemy movements or penalizes places they've been hit. But the best I've come up with in Diamond is still just a simple fast-decaying / well-tuned gun.

I think a major problem is that hitting flatteners and surfers is two different things (Despite them often being combined). When running top bots (Diamond vs DrussGT), I notice that once their flatteners are enabled, they often can't decide to use main or anti surfer targeting systems. To me, it seems like an optimized gun for hitting surfers (Fast data decay, fire where they have gone in the past) is not the same as an optimized gun for hitting flattened movement (??? perhaps data decay rate that is opposite of theirs? perhaps a gun that fires at the lowest point of their estimated movement curve in the instead of the highest?), and advanced bots now just have a main + anti surfer gun, with no effective method of dealing with flatteners, let alone flattening combined with surfing. I can imagine the problem would only get worse if bots started having a % rating for their flattener, instead of a simply on or off, allowing to vary how much flattening they do.

YersiniaPestis has a non-binary flattener, like you describe. He is pretty tough, and I do like the idea. I have several thresholds in Diamond's surf stats, as well, not just 2 modes.

But I don't think some thrashing between guns is a big deal - I certainly don't read that much into it. You have fairly little data, two very strong guns, and an enemy that's constantly changing its movement profile, so a lot of fluctuation doesn't seem surprising. I don't see anything inherently bad about thrashing, either. I still think it's the right move to choose the highest rated gun at any given point, even if it changes every shot.

I don't know, I just don't see a strong distinction between targeting vs flatteners or surfers. It's all the same algorithms and using information that's available to both of you. But I'd love to see some innovation vs either or both. :-)