3 ms·
In general, simulated annealing is hard to parallelize. But in this case we also tend to get stuck in local minima a lot, because they are very deep troughs in
by fogleman 9y ago
In general, simulated annealing is hard to parallelize.
But in this case we also tend to get stuck in local minima a lot, because they are very deep troughs in the search space. (Annealing does still perform much better than hill climbing though, so it's still worthwhile.) So we can run multiple annealings in parallel and pick the best result.
- sitkack 9y agoHow about tossing a bunch of boats into the build volume and shaking it? Use a sand/particle simulation to fit them together, seems amenable to the GPU.
- vortico 9y agoIf passing through one another is allowed during the physical simulation, this is just a CPU-heavy simulated annealing. If not, you won't reach solutions that are "far away" from your initial condition in terms of number of passes required to reach it.
- sitkack 9y agoThe process could be made continuous by building in supports to the volume of parts being printed, so it wouldn't need to be supported from the bottom for the whole duration of the build. Continuous Selective Laser Sintering via periodic support structures and powder dams. The interlocking parts and periodic build support structures could be made so that eventually lower parts could fall away and the powder recycled. One could also build in dams, so that the powder didn't fall away as the lower parts fell away. Those dams would allow for powder retention.
- sitkack 9y agoIn parallelization of annealing, do you think it would be worthwhile to broadcast the paths that are A) already taken B) not productive ? So that, each parallel annealer is working in a unique solution space?