I've been doing some simulations where the set of particles involved
changes quite frequently. This has raised some issues with how IMP
should handle particles being created and destroyed
1) When can you change the set of particles being optimized? My
current opinion is that OptimizerStates should be allowed to change
the set of particles, but ScoreStates should not. With that rule, the
space being optimized can change between optimizer steps, but not
during one. I added a constructor to BrownianDynamics which allows you
to explicitly provide the set of the particles to optimize (via a
SingletonList), but I think simply rescanning the model might not be
noticeably less efficient.
2) How to handle saving and restoring state? The model_io functions
don't work when the number of particles changes. The reason for this
is that there is no way to insert extra particles read into restraints
and things. One possible solution to this is to have the
List{Singleton,Pair}Container objects use a Particle as their
underlying storage mechanism. Then their sets of particles can be
easily saved and restored by the model_io functions (or anything else
which looks at all the particles). Since restraints (should) use these
containers as their underlying storage, then restraints could have
their state restored too. Making such a change would result in much
more data being stored as particle attributes, so we would have to be
careful that it didn't result in a performance penalty.
Anyone have any thoughts on these questions?