I've thought of a possible optimization for protein-folding (its a chemistry and DNA thing) simulations, which can be summarized as: voxel probability fields (of possible atom being at that 3d voxel location or not) of holographic (spheres around each 3d point) bayesian (for each point, the probability of each sphere, distance being a bayesian chance, same bayesian network used for the whole software, reused for every point thats calculated) supervised learning of current atom position mapped to final folded atom position relative to the spheres of bayesian chances, with some acoustic time-curved pattern matching to curve the sphere radiuses (using sound matching algorithms) to adjust for variations in the distances between each pair of atoms in the protein being folded while still using the fuzzy information of the distances relative to eachother. The recursive part is some kind of fast holographic statistical recursion bouncing between the spheres (summing the chances on the surface, but approximating it statistically using less data points) and updating the matrix of 3d voxels representing chances. Theres only 1 3d matrix, and it starts with 0 chance where there are no atoms (or do it at the amino acid level instead at the cost of adding types (which type of amino acid) to the 3d chance field). In a much smaller number of iterations than protein folding normally takes, this should converge to the folded state of almost any protein, using a little of the normal protein folding algorithms when it gets stuck, and using a little genetic algorithms (as is normally done in some protein folding simulations) to optimize parts of my algorithm. Theres a lot of possible variations of this, but I've described a general direction of research that appears to be missing from protein folding research and could possibly simulate it in a smaller Big-O on average. For "dimensionality reduction", for example, you could add 50 bayesian variables to the probability field and sum the surface of each sphere using the normal way of combining bayesian networks, many times, one entry in the sum for each voxel, and use that to get more accuracy, starting with random values (between 0 and 1) for all voxels for all 50 bayesian variables, and converging those voxel-bayesian-networks the same way as the other calculations are done. Written a shorter way: acoustic variable time curved pattern matching bayesian supervised learning of protein folding recursive holographic voxel probability fields. Just one more thing for me to experiment with, in this case using the FOLDING@HOME data... later when I have time. When I finish more of my Human AI Net framework, it should be advanced enough to do protein folding calculations without the core being specificly designed for protein folding, only as a plugin. --Ben F Rayfield
