emergent computing with swarm intelligent systems
TRANSCRIPT
Emergent Computing with Swarm Intelligent Systems
R Ryan McCune Advised by Greg Madey
Department of Computer Science & Engineering, University of Notre Dame 2014 Center for Advanced Modeling Graduate Workshop
Johns Hopkins University, BalKmore, Maryland, USA July 10-‐11, 2014
Overview OBJECTIVE: Agent-‐based modeling for Big Data • What is Big Data? • Why agent-‐based modeling? • Emergent behavior in distributed compuKng • Ant foraging applicaKon • Future direcKon
1
What is Big Data? • ExponenKal increase in data volume and velocity
• 80% of world’s data from last 2 years
• Challenges centralized computaKonal methods
• Data analysis is now a system – Not a program
2
“Think Like a Vertex” Programs • Graphs (networks) prevalent in Big Data • Typical graph processing not easily parallelized – E.g. graph-‐as-‐matrix operaKons
• Instead, execute funcKon on each vertex – UKlize vertex-‐local informaKon – Interact w/adjacent verKces
• Easily distributed (right) – Low inter-‐machine communicaKon
3
Machine 1 2
3 4
4
3 6 2 1
3 2 1 1
2 1 1 1
1 1 1 1
1 1 1 1
STEP 0
1
2
3
4
AcKve
Halt
Min Value
Agents for Big Data?
Big Data Challenges • Scalability • Inflexible • ComputaKonally expensive
Agent-‐based Strengths • Distributed • Adaptable • Simple behaviors generate global result
5
Mobile Agent Computing • Autonomous agent(s) traverse a network • Local behavior achieves global result – Vertex-‐oriented perspecKve
• Network environment – Problems
• ExploraKon, mapping, rendezvous
– ProperKes • Vertex labeling, agent communicaKon, synchronicity
• QuanKfiable and verifiable 6
Swarm Intelligence • Biologically inspired – Ant colonies, flocks of birds, fish
• Emergent phenomena – Simple, local behavior of agents – Complex, global behavior of system
• Self-‐organizaKon – Decentralized
• Emergent behavior solves problems
7
8
Distributed CompuKng Systems
Swarm Intelligent Systems
Emergent CompuKng
Ant Foraging
9
• Ants search for food to bring back to nest
• Randomly search environment • Deposit pheromones while searching – Likely to follow high pheromone – Random AcKon Probability (RAP)
• Shortest path emerges
RAP = ρ
1 – ρ Follow highest pheromone
ρ Random direcKon
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 uKlity model for collaboraKve foraging." Proceedings of the Third InternaKonal Joint Conference on Autonomous Agents and MulKagent Systems-‐Volume 1. IEEE Computer Society, 2004.
11
[1] Panait, Liviu, and Sean Luke. "A pheromone-‐based uKlity model for collaboraKve foraging." Proceedings of the Third InternaKonal Joint Conference on Autonomous Agents and MulKagent Systems-‐Volume 1. IEEE Computer Society, 2004.
Ant Hill
Food
How to Apply? • Ant foraging computes shortest path • Network environment – Convert grid to network – Sojware “ant” agents
• How to implement pheromones? – Each node stores value – Can’t read adjacent pheromone values – Use “think-‐like-‐a-‐vertex” (TLAV) program! • Share pheromone value with neighbors 12
13
.7 .6
.5
.3 .2 .1
vID pheromone 2 .6 4 .1
vID pheromone 1 .7 3 .5
vID 4
vID 3
vID 1 vID 2
RAP = ρ
Implications • Shortest path emerges from two simple programs
• Verifiable – But expensive and requires whole network in memory
• QuanKfiable Emergent bvr. = Agent bvr. + TLAV bvr. O(n log n) = O( 1 ) + O( 1 ) + m agents
+ k rounds
• Not necessarily faster – But adaptable!
14
MACADAM • Mobile agents traverse a network where nodes execute “think-‐like-‐a-‐vertex” program – Share variables
• Feasible validaKon • QuanKfiable emergence • Trade-‐Offs • HypercomputaKon
15
Future Work • Development of model – TLAV framework – Graph streaming capabiliKes
• Explore other behaviors – MACADAM framework
• Empirical MAC results • TheoreKcal foundaKon • WSN applicaKon 16
Conclusions • Big Data and Swarm Intelligence • Emergent computaKon for Big Data • Ant foraging applicaKon for networks • Mobile agents traverse a network that runs “think-‐like-‐a-‐vertex” programs
17
Acknowledgements • Air Force Office of ScienKfic Research (AFOSR) DDDAS program award # FA9550-‐11-‐1-‐0351
• GAANN Fellowship provided by the Department of EducaKon – Managed by the University of Notre Dame Computer Science Department
18