We're #150 @ Rosetta@Home, improving the world

[VJ]

Float precision not high (or precise) enough for you?

The values get rounded, so do not reflect the ideal mathematical values. The values in my case are coordinates that delimit areas. As the values get stored slightly differently from the real values, the program incorrectly identifies overlap or intersection between areas. And then the whole thing fails... Last 8 months I spent trying to find a way to circumvent it, and now I think I cracked it.

[Umfriend]

Now I only dabble a bit with VBA and SQL but how do values get rounded other than through float/double limitations? I would be terrified if a variable declared as float becomes an integer all of a sudden.

[VJ]

Yep... Imagine that you have the mathematical equations of 2 straight lines. You calculate the intersection point between them, but due to the float limitations its coordinates are rounded slightly. It can then occur that a test to check if the point is one of these lines fails.

Chapter 4 on this page is a nice read: https://www.cs.purdue.edu/homes/cmh/...books/geo.html
(but even the illustrations can give some ideas)

Introducing a normal epsilon check (if the point is within a small enough distance of the line, we consider it on the line) may work, but is not always adequate. My algorithm is highly dependent on accurate geometry calculations, so such small errors can have big consequences...

[Umfriend]

Principally, this seems to me to be an almost unsolvable issue as it may require infinite precision. Say you have two coordinates:
A = {pi, pi, pi}
B = {pi, pi, pi + x.10^-y)

There will always be a y for which an equality is found at any precision < 10^-y. Guess you can not forego an approximation within which an error may fall. I am ignorant on what it is you are trying to do but in cases where an equality or near equality is found, perhaps you could sort of zoom in? E.g, if you find
ABS(A - B) (in a dimension) < 10^-9 (or whereever precision is perfect) then look at
ABS(10^9 * (A-INT(A)) - 10^9 * (B-INT(B))) or somesuch? Not sure how large coordinates can get and it may be relevant as even large integers can not be represented perfectly in floating point notation.

Anyway, what do I know...

[VJ]

You are right that it is unsolvable in the general case. But for specific applications, one can come up with solutions/approximations that are good enough. The zoom idea might work for some applications, but not for others.
In my case, the main issue was the wrong detection of intersection between areas that *should* have a shared boundary, but should not intersect more than that. It took a bit to come up with something, but I think I have a workaround that works for me. I also had some issues with the calculation of intersections, but could avoid them by rewriting the algorithm.

[Umfriend]

Is it 2-dimensional and are the areas defined by continuous functions?

[VJ]

2 dimensional, defined by polygons. But the coordinates are georeferenced, so many calculations happen in the background to transform them to e.g. correct projections. In addition, I don't know how the data was processed before it came to me. These calculations cause that some areas which should share a boundary turn out to have an intersection or not share a boundary at all.

[J1NG]

Due to changes in real life here, after I hit the 1 mil mark, likely in the next few days, I won't be continuing crunching until life restabilises a level of stability. Sorry guys.

J1NG

[Umfriend]

You will be missed but before all: good luck with what ever it is that is going on!

[VJ]

Best of luck J1NG. I hope all is well.

Meanwhile, I can confirm that the new version of BOINC really makes a difference. I don't have the issues of delays when swapping, it properly detects when the computer is in use and releases resources much faster.

[Umfriend]

Also, SitFlyer has Resurrected! Welcome back!

Unfortunately, this may well mean we also stand to lose Marshmallowman soon. It's all a matter of seasons.

[Marshmallowman]

yeah...heatwave about to hit, will stop for while

[Umfriend]

Ah! October 28th another of our team made Predictor of the Day, J1NG! Congrats to you and our team! Nice one before you leave us for a (hopefully short) while.

https://boinc.bakerlab.org/rosetta/rah_old_potd.php

[Umfriend]

OK, apparantly there are only three left crunching for the team

What's up guys?

[VJ]

I forgot to launch BOINC after the last reboot for updates.... I started it now...