Open cmaclell opened 5 years ago
A possible middle ground could be the use some of the non-incremental mechanisms that SimStudent used to use but with some kind of memory retrieval model so we don't have the perfect memory problem. I suppose that raises questions around whether the non-incremental processes still work if they aren't guaranteed to get all of the previous examples back.
Another possibility on this question is we could do the ModularAgent, and the IncrementalModularAgent, where one makes a stronger commitment than the other. It would really help if we could understand the exact process differences between the two.
This is related to #27 and #26
So Ken brought up the point that AL is not strictly more general than SimStudent I think this is right and the main reason is that AL makes a commitment to incremental learning where as SimStudent uses many non-incremental mechanisms.
I guess this issue is just a question, it is possible to commit to both non-incremental and incremental learning at the architecture level?
For example, to support non-incremental learning we might need to call the learning mechanisms with all the past examples. However, this might force us to commit to maintaining a history of all examples and always providing the learning mechanisms with all the examples-- effectively committing to a non-incremental approach at the architecture level.
Maybe i'm overthinking it and we can easily support both, but just wanted to put this out there.