Re: Event-Oriented Computing

[Marcus G. Daniels]

>>>>> "GR" == glen e p ropella <address@hidden> writes:

MD> The line would have to be above all those things, of course.
MD> Agents can't be allowed to run systems calls or directly or
MD> indirectly control the program counter.  (Unless you have a
MD> workstation that's not connected to the internet and you don't care
MD> if its files are lost.)

GR> Why not?  We already allow system calls.  

The difference is that, currently, the models are handcrafted to 
Do The Right Thing.  In the "throw a snake into a fish tank just to see
what it'll do" scenario, the experimenter doesn't have a understanding
of the consequences of their actions.  If the foreign agent looks
around for features to run, and stumbles across `unlink', `opendir'
and `readdir', it isn't hard to imagine how files would start
disappearing.

GR> Is there no way to build
GR> a kind of virtual machine such that the agent can have access to
GR> functions that in a normal process would be dangerous but in here
GR> would only crash the virtual machine?  Does what I want to do
GR> require the complexity of something like the JVM?

..or some kind of controlled interpreter.

GR> So, is this the kind of thing we could muck with
GR> (e.g. permute, delete items from, switch objects, etc)?  Or is it
GR> a "truth" thing where what you see is what's really there?

There's no notion of security in Swarm, and a model running across
multiple computers has equal access to resources, so yes, I suppose so.

GR> I suppose the next thing for me to worry about is the dynamism of
GR> the protocols.  With defobj, we could generate whole classes
GR> (interface and methods), right?  Can we generate protocols
GR> dynamically?

The `read' end of serialization will generate a class on the fly if
necessary, but it doesn't yet have any support for generating methods
on the fly (other than inheriting them).  The method code has to
come from somewhere.  The best we could do without an interpreter would
be to provide method lookup by name.

GR>  I.e. what i'd like to do is, while the simulation is
GR> running an agent would poke around in the directory, find methods
GR> she'd like to ball up into a class, construct that class and then
GR> construct a protocol for objects to adhere to, then instantiate an
GR> object.  Will that be possible?

To me, constructing variant classes from existing methods isn't very
interesting, although I could see how it could be used.  Ideally, it
should be possible to have a simulation write new methods by dynamically
generating code (and testing it in a sandbox).


                  ==================================
   Swarm-Modelling is for discussion of Simulation and Modelling techniques
   esp. using Swarm.  For list administration needs (esp. [un]subscribing),
   please send a message to <address@hidden> with "help" in the
   body of the message.
                  ==================================