Emergent Computing with Swarm Intelligent Systems

Ryan  McCune  &  Greg  Madey  

University  of  Notre  Dame,  Computer  Science  &  Engineering  SwarmFest  2014  

Notre  Dame,  IN,  USA  July  1st,  2014  

Problem – Big Data

•  80%  of  world’s  data  from  last  2  years  

•  Increased  volume  challenges  data  analysis  

•  Analysis  as  a  system  •  Problems  with  centralized  computa@on  

  1

Distributed Computing •  Connected  computers  – Nodes  and  edges  

•  Distributed  computa@on  – S@ll  central  coordinator  •  BoFlenecks  – Not  Scalable  – Failure  prone  

•  Global  Informa@on  – Computa@onally  Intractable   2

Solution - Emergent Computation •  Global  behavior  emerges  from              interac@on  of  distributed  computers  –  Global  behavior  also  a  computa@on  

•  Decentralized  –  No  boFlenecks  

•  Scalable  •  Robust  

–  Efficient  •  Each  parallel  computer  executes  simple  program  •  Complex  computa@on  emerges  

3

4

Distributed  Compu@ng  Systems  

Swarm  Intelligent  Systems  

Emergent  Compu@ng  

Swarm Intelligent System

•  Ar@ficial  swarm  inspired  by  biology  

•  Mul@-‐agent  system  opera@ng  in  an  environment  

•  U@lize  emergent  behavior  to  solve  problems  – No  boFlenecks  

•  Scalable  •  Robust  

– All  local  behavior  •  Complex  behavior  emerges   5

Swarm Example - Flocking

6

Separa@on  

Alignment  

Cohesion  

•  Move  with  speed  and  direc@on  •  Sight  radius  to  perceive  neighbors  

•  Adjust  movement  in  3  ways  based  on  neighbors  (leU)  

•  Coordinated  flock  emerges  – From  simple,  local  behaviors  

7

Approach •  Emergent  compu@ng  

–  Poten@al  to  solve  Big  Data  challenges  –  But  few  examples,  if  any  –  So  how?  

•  Look  at  swarms  that  do  computa@on  –  Then  figure  out  how  to  translate  to  distributed  systems  

•  Swarm  example-‐  “Ant  Foraging”  –  Well-‐known  –  Shortest-‐path  emerges  

•  Swarm  example-‐  “Decentralized  Clustering”  –  New,  based  off  “Ant  Foraging”  –  Clustering  emerges  

8

Example – General Ant Foraging

9

•  Ants  search  for  food  to  bring  back  to  nest  

•  Randomly  search  environment  •  Deposit  pheromones  while  searching  – Likely  to  follow  pheromones  – Random  Ac@on  Probability  (RAP)  

•  Shortest  path  emerges  

RAP  =  ρ  

1  –  ρ  Follow  highest  pheromone    

ρ  Random  direc@on  

Ant Foraging - An Implementation[1] •  Ants  deposit  2  pheromones  – Green  lead  to  home,  deposit  while  foraging  – Blue  lead  to  food,  deposit  while  returning  home  

 

10

[1]  Panait,  Liviu,  and  Sean  Luke.  "A  pheromone-‐based  u@lity  model  for  collabora@ve  foraging."  Proceedings  of  the  Third  Interna@onal  Joint  Conference  on  Autonomous  Agents  and  Mul@agent  Systems-‐Volume  1.  IEEE  Computer  Society,  2004.  

•  1  ant  hill  –  Sta@onary  

•  1  food  –  unlimited  

•  Many  ants  

11

[1]  Panait,  Liviu,  and  Sean  Luke.  "A  pheromone-‐based  u@lity  model  for  collabora@ve  foraging."  Proceedings  of  the  Third  Interna@onal  Joint  Conference  on  Autonomous  Agents  and  Mul@agent  Systems-‐Volume  1.  IEEE  Computer  Society,  2004.  

Swarm Clustering •  Adapted  from  ant  foraging  – Many  food  instead  of  1  food  – Many  ant  hills  instead  of  1  ant  hill  •  Ant  hills  can  move  (right)  

– Only  1  pheromone  type,  not  2  •  Deposit  when  looking  for  food  •  Follow  to  return  to  ant  hill  •  No  pheromone  leads  to  food  •  Once  any  food  is  found  randomly,        pheromone  leads  to  nearest  ant  hill  

12

1.  Ant  finds  food  

2.    Ant  returns  to  nest  

3.  Nest  moves  closer                      to  food  

13

Tanker  Moves  

Behavior

14

Swarm Advantages •  Self-‐Organizing  – Autonomous  

•  Robust  –  Failure  of  any  agent  doesn’t  impact  performance  

•  Scalable  –  Can  add  agents  without  burden  

•  Adaptable  – Add  or  remove  sensors  and  system  adapts  

•  Computa@onally  Tractable  –  Simple  behaviors,  complex  result  

In Contrast: Central Control

16

•  BoFlenecks  –  Failure-‐prone  •  System  broken  if  link  fails  

–  Not  scalable  •  Bandwidth  limits  

•  Computa@onally  intractable  –  Test  combina@ons  of  all  sensors  

•  Not  adaptable  –  New  computa@on  if  sensors  added  or  removed  

Applying to Big Data •  Swarm  Clustering  – Applicable  to  spa@al  problems  – Applica@on  to  Big  Data  not  clear  

•  Networks?  – Convert  grid  to  network  – Agents  traverse  network  – Mechanism  for  sensing  pheromones  in  adjacent  nodes?  

17

Vertex-Centric Graph Processing •  Big  Data  includes  networks  – Web  – Facebook  

•  Recent  introduc@on  of  graph-‐parallel  frameworks  

•  Each  ver@ce  executes  func@on  •  Vertex-‐to-‐vertex  message  passing  

18

PageRank Example

19

•  Itera@vely  calculates  a  node’s  importance  

•  Centrally  computed  using  matrices  – Expensive  when  big  

•  Can  be  simply  expressed  as  vertex  program  – Easily  distributed  

Vertex Functions + Agents

20

Vertex Functions + Agents •  Advantages  – Scalable,  robust,  adaptable  – Verifica@on  – Quan@fy  emergence  

•  Future  work  – Develop  ant  foraging  model  – Explore  alterna@ve  configura@ons  – Equivalent  computa@onal  power  

21

Conclusions •  Explore  swarm  intelligent  computa@on  – How  to  translate  to  distributed  compu@ng  

•  Introduce  swarm  intelligent  clustering  –  Further  work  

•  Elaborate  behavior  •  Compare  centralized  clustering  

•  Applica@ons  of  swarms  –  Robust,  scalable,  adaptable,  computa@onally  efficient  

•  Further  explore  Emergence  22

QUESTIONS? 23

Acknowledgments •  Air  Force  Office  of  Scien@fic  Research  DDDAS  program  grant  

•  GAANN  Fellowship  provided  by  the  Department  of  Educa@on  through  the  University  of  Notre  Dame’s  Computer  Science  Department  

24