We want the capability to launch a new alphazero run that starts from generation G of some prior run. A simple way to do this is to just cp all directories from the prior run directory to the new run directory, deleting all files that come after generation G. Maybe this is good enough, although if we have a large G, large models, and do N>>1 forks, this might use a lot of disk space. We'd save by doing some sort of soft-linking to the prior run.
The reason we want this is to perform experiments more economically. For example, if we want to test whether a change improves the skill-ceiling, it is uneconomical to repeat the first ~90% of the learning. Better to fork from a prior run to skip the first 90%.
We want the capability to launch a new alphazero run that starts from generation G of some prior run. A simple way to do this is to just
cpall directories from the prior run directory to the new run directory, deleting all files that come after generation G. Maybe this is good enough, although if we have a large G, large models, and do N>>1 forks, this might use a lot of disk space. We'd save by doing some sort of soft-linking to the prior run.The reason we want this is to perform experiments more economically. For example, if we want to test whether a change improves the skill-ceiling, it is uneconomical to repeat the first ~90% of the learning. Better to fork from a prior run to skip the first 90%.