When generating new training data we have to find positions for rollout and then roll them out. This is currently done in one step.
In the future finding positions will be more complicated (for example we might only want to find positions where the current neural nets are bad at evaluating them). So this will take more time. Also currently if we interrupt the rollouts, the information about which positions have been found is lost.
So we should split finding positions and rollout. The result of finding positions should be written into a CSV file with just a single column: the position ID.
When generating new training data we have to find positions for rollout and then roll them out. This is currently done in one step.
In the future finding positions will be more complicated (for example we might only want to find positions where the current neural nets are bad at evaluating them). So this will take more time. Also currently if we interrupt the rollouts, the information about which positions have been found is lost.
So we should split finding positions and rollout. The result of finding positions should be written into a CSV file with just a single column: the position ID.
Also see #13