opencog-dev team mailing list archive

[Ari Heljakka <ari.heljakka@xxxxxxxxxxxxxx>]

> 1)
> When PLN wants to set fresh=true, then the AtomTableWrapper
> creates a new version for Atom A.
>
> PLN will have N different "dummy contexts" associated with it:
> C1, C2,..., C_N
>
> The first version of A created by PLN will have a VH associated with
> context C1... the second version of A created by PLN will have a VH
> associated with context C2 ... etc.

This is fine, as long as PLN gets a separate valid Handle to each 
version of A, such that it can use that Handle as if it pointed directly 
to the actual A. So if you have a unique node for each context, that 
should work. Then, ATW just translates between the contextual form and 
the simpler form which PLN sees.

> 2)
> Then, when appropriate, PLN can merge all the versions of A associated
> with the different PLN dummy contexts into a single version (which in most
> cases will just be the default version for the Atom A ... except in odd
> cases
> where the TV estimated for A by PLN is radically different from the one
> estimated
> by another method like direct evaluation, in which case we might still want
> to keep
> around one PLN-based version of A, along with other versions...).  After
> this merger
> the dummy-context versions are all removed.

Well, the current PLN will never actually want to merge/delete those 
wrong versions. The AtomSpace "garbage collection" was never implemented 
due to being low priority at the time. (The "managed" parameter involved 
some attempt at it, but I don't remember the details.)

Essentially, PLN treats the "fresh=true" atoms as VALUES, not as
OBJECTS. So, it doesn't really care if multiple atoms with the same 
structure exist or not. It has its own data structures which track those 
different atom versions (with distinct TVs) and then merges them by 
creating a yet another atom version.

So there's no need to implement your #2.

IIRC, the GC happens currently by simply resetting the whole AtomSpace, 
and then re-initializing it with the premise set.

...How then it was possible to create a dynamic continously running 
system like the Fetch Loop without memory overflow? I probably did some 
ingenious engineering there, but it'll take me some code-reviewing to 
bring it back to mind...

> I guess this would all be done in the AtomTableWrapper ...

Probably.

--
Ari Heljakka
CTO
Dream Broker Ltd		

Tekniikantie 14			ari.heljakka@xxxxxxxxxxxxxx
02150 Espoo			+358 40 568 2420
Finland				www.dreambroker.fi