What is BISON doing? Let us try to be precise - Geoff Canright Telenor R&D Norway

Page created by Charlotte Robles
 
CONTINUE READING
What is BISON doing?
   Let us try to be precise…
Geoff Canright
Telenor R&D
Norway

     Project funded by the Future and Emerging Technologies arm of the IST Programme
What is BISON about?

We want to implement a number of functions on
  a variety of network structures, using complex
  adaptive systems (CAS) which have nice
  properties
„ Functions, structures, CAS: we have some
  progress towards precision in definitions
„ Nice properties (adaptive, robust, etc): to be
  rendered more precise in 1st half 2004

1st Review Meeting 23 February 2004                      2
Structures
„   Our original list: P2P, Grid, Virtual, Ad-Hoc
„   Improved: Overlay (logical) networks (OL) vs. mobile
    (wireless) ad-hoc networks (MANETs)
„   More precise: structure = topology
          • We have a growing list: lattice, random, scale-free, small-
            worlds, etc.
          • Some arise spontaneously; others must be managed (more
            on this soon)
          • In the OL case, routing and connectivity are cheap and
            easy Î many possible topologies
          • An (unmanaged) MANET has a (dynamic) “spatial
            topology”

1st Review Meeting 23 February 2004                                   3
Functions

„   Basic functions: routing, search, monitoring,
    aggregation, topology management
„   Advanced functions: distributed content,
    distributed storage, distributed processing
„   The basic/advanced distinction is not so precise
„   We can write all basic functions (except topology
    management) as flow diagrams
„   We use load balancing in two of our advanced
    functions, but haven’t placed it in our scheme

1st Review Meeting 23 February 2004                4
Some flow diagrams for
                                              basic functions
               goodness   routing 
  ( S , D) +            ⇒        ⇒ ( path(s))
               criteria   function 
    resource   search   locations , 
                    ⇒        ⇒        
     specificat ion   function   resources 

        raw               monitoring             health     
              ⇒                         ⇒              
         data              function               indices    

1st Review Meeting 23 February 2004                                   5
Advanced functions

„   Distributed processing: work to be done is
    shared — try to use capacity of all participating
    nodes (load balancing)
„   Distributed storage: a distributed warehouse.
    Can I fetch my files when I want them? (load
    balancing; also, need redundancy/reliability)
„   Content sharing: a distributed library. No
    ownership of content. How to find stuff?

1st Review Meeting 23 February 2004                     6
The significance of topology
                                         management
„   Some topologies support a given function better (even
    much better!) than others
„   At least in the OL case, we can aim for techniques for
    maintaining a desired topology to support a desired
    function
„   Close the loop: we use (other) functions to support the
    target topology
„   Î “layered” approach (see Bologna talk):
    given topology supports functions, which support new
    topologies, and so on
„   MANETs: we do not rule out interesting possibilities
    here either

1st Review Meeting 23 February 2004                           7
CAS = COW (can of worms)

„   Need to avoid ambiguous and confusing
    language
„   We have a “BISON working definition”:
          •   many simple agents
          •   sensitive to their environment
          •   interacting via local roles
          •   the interaction is not “brute parallel”
     „   NB1: we view our agents as finite-state machines
         (FSMs); this gives an implicit definition of “simple”
     „   NB2: brute parallelism fits easily in our scheme,
         but we don’t wish to call it a CAS
1st Review Meeting 23 February 2004                         8
Building blocks for agent behavior
Agents are FSMs, augmented with other things. They
   have types (state structures) A,B,C,… and states
   a,b,c,…
We call our building blocks microscopic mechanisms:
     1.    Aa Î Aa   (survival, memory)
     2.    Aa Î 0     (death)
     3.    A Î AA    (proliferation)
     4.    Aa Î Ab   (response, interaction)
     5.    A Î AB    (signaling)
     6.    A Î     A (mobility)
     7.    A ÎB       (mutation)

1st Review Meeting 23 February 2004                   9
The biological inspiration

„   We have looked at (to date) five biological
    systems with interesting behavior:
          • Ants: path finding using pheromone; gathering
          • Immune cells: search, recognition, and response
            to antigens
          • Slime mold amoebae: physical aggregation as a
            response to collective hunger, using chemotaxis
          • Neurons: collective memory using synapse
            modification
          • Viruses: epidemic spreading

1st Review Meeting 23 February 2004                       10
Analysis of the 5 biological systems
           using our basic building blocks
                                Immune
                     Ants                   Amoebae   Neurons   Viruses
                                  cells
Memory
(state)                x              x       x         x
Selection                             x                           x
Proliferation                         x                           x
Response               x              x       x         x
Signaling
(diffusive)                           (x)     x
Signaling
(nondiffusive)         x                                x
Mobility               x              x       x                   x
Mutation                              x                          (x)
1st Review Meeting 23 February 2004                                       11
What BISON is about (II)

 „   “Engineering with emergence”
 set of     
              abstract               collective 
microscopic  ≡         emergence
                               →                 = ( function )
mechanisms   CAS                     behavior 
            
      „   Collective behavior is easy to predict when the
          interaction is trivial
      „   We also want to predict collective behavior for
          nontrivial interactions, which exhibit emergence

 1st Review Meeting 23 February 2004                                12
A BISON goal:
                               systematic understanding

„   This is hard!
„   That is: at the end of BISON, we can seek to
    extract some systematic understanding from
    our experience
„   We envision using two approaches:
     „   Collection of performance results—focus on
         effects of topology, and of CAS/mechanism
     „   Collection of heuristics for “design” of CAS to
         perform a given function

1st Review Meeting 23 February 2004                        13
One type of heuristic
                           for CAS design: “synthesis”
„   Look at known examples—analyze them in
    terms of our microscopic mechanisms—and
    compare them
„   Generate new candidate CAS by “tweaking”
    the known sets of microscopic mechanisms
„   EX: slime mold colonies perform collective
    computation of the colony’s level of hunger;
    and they move (physically aggregate) as part
    of the process
„   Candidate: the same set of mechanisms can
    perform collective computation, but without
    mobility
1st Review Meeting 23 February 2004                 14
Another heuristic: “operators”
„   An “operator” takes (CAS) ⇒ (CAS)’
•   That is, it creates a new CAS from a given one
•   EX1: “inversion”. Given a CAS with known collective
    behavior, invert one or more microscopic mechanisms
    to get inverted collective behavior
     •   Ants: inverted gathering = load balancing
     •   Chemotaxis: inverted aggregation = load balancing
•   EX2: “exaptation”. Put a CAS in a new environment,
    and it can display enhanced functioning
     •   Immune system: RW + proliferation appear to give
         efficient search on a network
•   How far can one go with this approach??

1st Review Meeting 23 February 2004                          15
You can also read