The algorithm calculates all possible specular reflections up to some order. I really don't think I can explain it any better than this paper: https://mediatech.aalto.fi/~samuli/publications/laine2009aa_...

It doesn't calculate all beams that originate from one point. It calculates only the beams which originate from one of the points and eventually reach the other point. It only looks random because the walls are placed at odd angles.