Rogue MindAgents

From OpenCog
Jump to: navigation, search


    What happens if you have a MindAgent that spawns many MindAgents (that
    spawn MindAgents)? Would the system get completely bogged down?

    It is those kinds of things I worry about.

     Will


Yes, it could...

But, I really think that effectively protecting against arbitrary "rogue 
MindAgents" would make the system infeasibly slow and cumbersome
 
I think that

1)
we have to trust human MA developers not to do insanely stupid stuff ... and if 
they do, we don't use their MAs in our OpenCog configurations

2)
when we get to automatic learning of MA's, we need to direct the **learning 
process** itself not to produce rogue MindAgents, rather than assuming it will
 rampantly produce them and then needing to filter them out



The system should be robust against being *crashed* by rogue MA's, but IMO it 
shouldn't be required to protect against them in some way that doesn't disrupt 
its ordinary functions.  I think that it is OK to assume that if a rogue MA 
arises, the system may have to stop what it is doing while the problem is 
solved...


ben

...



      I think that it is OK to assume that if a
     rogue MA arises, the system may have to stop what it is doing while the
     problem is solved...
    

    I wouldn't want it flying my plane, driving my car or running my world
    economy in this case....


What the OpenCogPrime design specifies is that learning of new MindAgents 
should be done in special "experimental" lobes of the system.  Once the new MA 
has proved its value in these experimental lobes, it can then be moved into the
 main lobe.

So, you wouldn't use the experimental lobe of an OCP system to do these 

practical things, no.

We have not implemented this stuff because, as I said, it just doesn't seem the
 most pressing problem.

If you have an AI design in which creation of new MA's occurs routinely as part 
of ordinary cognition, then you would need this stuff early on.

But instead, in the OCP design, we have "Schema", i.e. small Combo programs 
that can act more like Stan Franklin's or Hofstadter's "codelets", or Minsky'
 "agents".  MindAgents are not supposed to play that sort of role.  Schema can 
be created, destroyed and morphed freely now, unlike MindAgents.

The CogServer doesn't support Schema right now, but the OpenPetBrain does, and
 they'll be moved into the CogServer "any day now"

-- Ben G