The game Elite Dangerous has a proceduraly generated galaxy of some 400 billion star systems.
Each star system in the game can be uniquely identified bu a 64bit number (id64) which is used as a seed for building the system but can also be decoded down to give some information about the system.
- Boxel coordinates: x y z coordinates of a cube in which the system can be located.
- Masscode: A number that determines the size of the boxel and the composition of the system
What we cannot decode from the id64 is the actual x,y,z coordinates and the boxel coordinates are too imprecise to be useful. Being able to preduct the x,y,z coordinates would allow us to make star maps of unvisited stars.
However the actual coordinates must be derived from the id64, some pseudo random number function and /or some rules grounded in real world astronomy.
Through the game journal files we can get the id64 and the actual coordinates of the system in 3D space and there are databases of player captured data containing about 70million systems we can use for test and train.
I took this data and created a file containing id64,bx,by,bz,masscode and x,y,z
(bx,by,bz is the boxel coordinates)
Using the features id64,bx,by,bz,masscode I want to be able to predict to within a light year, the x,y,z coordinates of the system.
I tried a few different Models from scikit-learn to see if I could train one to get coordinates but so far had little luck using various models. I had hoped that having a training set in the training/test set in the millions might help.
Is there a model or approach that might be better suited to this kind of problem? I guess if there is a psuedo-random element to it, then it might not be possible?
