# what's the secret to making a good robot in robocode

Fragment of a discussion from User talk:Tmservo

I actually don't feel like GFs make any major assumptions besides clockwise vs counter-clockwise being treated the same. And it would take some actual effort to make that assumption false. The firing angle you use is relative to where you are firing from, regardless of how the enemy moves. And that's the output you need from a targeting algorithm, not the exact location of an enemy.

22:12, 18 December 2013

GF assumes symmetric movement, which is safe. Also assumes orbital movement, which is not as safe, but still hard to exploit.

GF combined with statistical targeting, also assumes non-adaptable movement, which is not safe at all. Exploited by all surfers.

MN (talk)22:31, 18 December 2013

How does it assume orbital movement? If you're not staying perpendicular to me, my GF gun is going to crush you even harder.

22:36, 18 December 2013

In practice, yeah, because the whole point of perpendicularity is to maximize your MEA. But, GF guns are actually assuming that the enemy is orbiting when it visits those GFs. For example, SittingDuck and RamFire both stay at GF0, even though they have very different movements. I'm definitely not saying that this assumption would be easy to exploit, especially with segmentation, but it could be, in theory.

23:22, 18 December 2013

I still disagree that GuessFactors introduce that assumption. A GF is a scalar value representing the whole range of firing angles the enemy could reach. Orbital movement maximizes this range and distributes the values most evenly across it.

23:30, 18 December 2013

The way I think of it is... besides the clockwise/counterclockwise thing, the other assumption that GuessFactors make is that the movement profile either stretches with BFT (or more directly MEA) or that BFT is rather constant.

Raw bearing offsets are similar except that they don't apply such a stretching, and instead assume that either BFT is rather constant, or... bots are breaking the physics rules... no that doesn't count... they just rigidly assume BFT is rather constant.

The main innovation of GuessFactors over raw bearing offsets is that it generalized what was assumed as to be less rigid, but it still makes some assumptions. Further, things such as segmentation by BFT or MEA tend to mitigate the effect of this specific assumption.

23:58, 18 December 2013

Basically, I don't think GuessFactors assume anything (besides symmetry) that isn't already true based on the physics of Robocode. Your input to firing a shot at your already chosen bullet power is an absolute angle (modulo your current gun heading). A wave collects what should have been the input to that API call. To me that seems about as close to "making no assumptions" as you can get.

So I'm forced to ask myself: why don't you use GFs in Melee? Doesn't that prove they assume an orbital movement? But I really think it's more that you can assume non-orbital movement in Melee, so it's advantageous to do so (PIF, displacement vectors).

03:55, 19 December 2013

I think of "assume non-orbital movement in Melee" and "assume orbital movement in 1v1" as two sides of the same coin. Either way it's an assumption. You could go somewhere in the middle by having both GF-based and PIF-based (and displacement-vector-based) methods and picking the one that seems to best characterize the current enemy, but even then you still make an assumption: That the assumption of a fixed method is suboptimal often enough that it's worth the penalty of sometimes picking the wrong one ;)

05:40, 19 December 2013

In practice, they are two sides of the same coin, but I still disagree. Only the "assume non-orbital movement" involves extrapolating enemy movements (relative to anything) and translating to and from firing angles. Waves collect very close to the exact value that the Robocode API takes as input to decide if you hit an enemy. That just doesn't feel like adding any assumption that isn't already introduced by Robocode itself.

05:57, 19 December 2013

But on further thought, it's also no excuse to ignore an assumption made by the API. As in, coming up with a model that eliminates an assumption made by the Robocode API would (edit: well, could...) be an improvement to a targeting algorithm. And I'll concede that the Robocode API taking a firing angle as input introduces an assumption of lateral displacement scaling by bullet flight time.

06:12, 19 December 2013

It seems to me that assuming lateral symmetry could be exploited by an opponent. Imagine a bot which started moving clockwise, went to GF .5 . Then for the next wave, its moving counterclockwise, a gun assuming symmetry will fire ahead if it at GF .5 relative to its orbit direction. It is not looking at GF relative to orbit direction, but absolutely in terms of clockwise and counterclockwise. It moves again to absolute GF .5. It continues to always move to absolute GF .5, but predicts enemy firing times and arranges to alternate orbiting clockwise and counterclockwise when they fire. This simple version would probably be much more effective against guns with fast data decay.

02:20, 20 December 2013

Hmm. But if you're already tracking stats from the enemy's perspective, isn't it more effective to just use the same model as they are and avoid the dangerous spots? I don't need any symmetry tricks to know the enemy will shoot at GF 0.5 (from his perspective) and not go there.

I'm curious what MN and Rednaxela have to say. :-) I'm thinking it might be exploitable by a non-surfing movement, but for a surfing movement you are sacrificing more than your'e gaining.

02:34, 20 December 2013

Im not saying my system is a particularly good one, it just shows that using relative GFs should technically be exploitable. What you are saying, knowing enemy will shoot at .5, is basically using a flattener.

03:22, 20 December 2013

I don't know, I think to me the definition of "exploiting" means you can use it to improve your system. Tweaking a random movement to have a non-symmetric profile could be an improvement. With surfing, your goal is to model the enemy's targeting data model, which is done correctly by assuming symmetrical GFs. It seems to me a surfing movement is already aware of the GF symmetry and taking it into account as best it can.

03:34, 20 December 2013

From the perspective of normal (relative to orbit direction) GFs, what it comes down to is that you'd have is a movement which alternates between GF 0.5 and GF -0.5.

It's true like you say that this could have some level of effectiveness against targeting systems that are using firing waves only AND are missing certain segmentation dimensions...

I would say it could be 'exploiting' yes, but I would also guess there are very few targeting systems you'd reliably trick with this. It seems extremely fragile. Your maneuvers to "arrange to alternate orbiting ..." would give away which of GF 0.5 and GF -0.5 you're heading to in certain targeting segments/dimensions (easy near-100% hit rate against it), and even without that, asymmetry in the result of non-firing waves would cause a lean toward one or the other, causing an easy ~50% hit rate.

It would be interesting to see a demonstration of what in practice would be tricked by it though.

16:05, 20 December 2013

I'll give you that it assumes a movement profile that scales with bullet flight time. But that's a very light assumption and not the same as assuming orbital. What types of movement don't scale with BFT? Just fixed patterns?

PS: Mathematically, I think bearing offsets would assume bullet power is constant, not BFT, right? ;)

04:04, 19 December 2013

Fixed patterns and bullet dodging movements like enemy fire oscillation.

04:28, 19 December 2013

No, because bearing offsets have no relationship to MEA. I think that raw MEA is the source of the orbit assumption in most guns. The raw MEA is assuming that the enemy can reach all those possible future locations by the time the wave hits, which it can't if it isn't orbiting. Precise MEA pretty much solves this.

03:32, 19 December 2013

Why is that an assumption they will ever go there? It's only taking into account that they could. If they never go there, what's the difference? If your bot has a min GF of -0.1 and a max of 0.3, it wouldn't make my gun any more accurate to change my MEA to match.

03:41, 19 December 2013

If an enemy bot never went outside of the range between -0.1 and 0.3, you're right, it probably wouldn't affect accuracy much. However, say an enemy bot has a huge spike at 1.0, and you decide to fire at it, even though the enemy cannot possibly reach 1.0 with its current heading. The shot would be wasted.

04:22, 19 December 2013

So, yeah, I'm always thinking in terms of precise MEA, and thinking that I will never fire at a GF the enemy never visited. So I don't see how it's assuming anything about how much of the escape range they might actually cover - the MEA is effectively however much of the range they ever cover. There is the scaling, which I think only assumes they're not moving in a fixed pattern.

And you're right that that is a problem with a certain MEA approximation formula some people use with their GuessFactors. :-)

04:29, 19 December 2013

My micros don't even have a formula for MEA, they just use the constant 0.7272...

I suppose the problem with imprecise MEA is kind of moot, as we were talking about what DrussGT assumes, and DrussGT has precise MEA.

05:32, 19 December 2013

GFs adjust angles on bullet flight time. If it saw GF 1.0 previously with a low powered bullet, when shooting with a high powered bullet, GF assumes the target will continue moving until it reaches the adjusted GF 1.0. This holds true only with orbital movement, or linear movement if close to walls.

If it was a pattern movement, like MyFirstRobot or SpinBot, the assumption would be wrong. But since pattern movement is clearly weaker than orbital movement, it is quite safe to assume opponents are orbiting. Segmenting by bullet power, distance and time-since-direction-change fixes the lack of accuracy in the long run.

But if you are trying to maximize APS by crushing weak movements, then PIT might be better in some situations.

MN (talk)05:19, 19 December 2013

If the enemy pays no attention to where I am and just moves randomly, the lateral displacement will still scale by BFT and GFs will still make sense, unless their algorithm is hard-coded to only move a fixed distance. That movement isn't orbital. Am I wrong?

05:35, 19 December 2013

Ok, any movement which scales with BFT works with GF. But which movements are these?

- Orbital movement

- Random movement normalized to move to all reachable angles (random orbital movement)

It doesn't work as well on other movements(slower learning):

- Pattern movement (including linear movement and oscillators)

- Random movements, which change heading/velocity based on anything not related to MEA

In practice, GF works against all movements above, but thanks to segmentation. Segmenting by distance and bullet power fixes most assumptions in GF which happen to be wrong in the long run.

Assuming wrongly that a movement is orbital is not that bad. Movements which do not try to maximize MEA are more predictable in the long run. If it is orbital, GF is right from the beginning and do well. If it is not orbital, then any statistical gun will have a higher hit rate in the long run.

And a worst case scenario. What if someone uses a pattern movement in the beginning and switches to orbital later on? DrussGT is doing this since it incorporated EnergyDome's bullet shielding.

MN (talk)17:02, 19 December 2013