It can be done in four steps:
- identifying four points lying on the same plane and finding its equation (we know that we have 15 planes so it is enough to check 15*3+1 points to find them);
- applying the equation to all points in randu dataset to divide them into 15 classes;
- veryfying that the classes separate points into 15 parallel planes;
- plotting the solution changing colors for points on different planes.
library(rgl)
library(caTools)
# STEP 1
all.combs <- combs(1:46, 4)
i <- 1
repeat {
model <- lm(z ~ x + y, data = randu[all.combs[i, ], ])
if (summary(model)$r.squared > 0.99999) {
break
}
i <- i + 1
}
# STEP 2
line.class <- predict(model, randu) - randu$z
line.class <- factor(round(line.class) + 10)
# STEP 3
summary(lm(z ~ x + y + line.class, data = randu))
# STEP 4
with(randu, plot3d(x, y, z, axes = FALSE, col = line.class,
xlab = "", ylab = "", zlab = ""))
rgl.viewpoint(theta = -3.8, phi = 3.8, fov = 0, zoom = 0.7)
At step three we can see that the model obtains perfect fit. The final figure is plotted below:
It would be a cool follow-up to show analytically that some rotation of randu generated data lie on a series of planes. Maybe this had already been shown...
ReplyDeleteYou can find it at http://en.wikipedia.org/wiki/RANDU.
ReplyDeleteHowever, I thought that the code is a nice example of how one can numerically categorize and plot data located on parallel planes:
1) without knowing the exact relationship formula;
2) taking into consideration rounding of observations.