The Parable of the Hare and the Tortoise: How "Small Worlds" Reduce the Long Run Performance of Systems David Lazer Program on Networked Governance Harvard.

Slides:



Advertisements
Similar presentations
Jacob Goldenberg, Barak Libai, and Eitan Muller
Advertisements

1 Panel Data Analysis – Advantages and Challenges Cheng Hsiao.
Particle Swarm Optimization
Optimizing Membrane System Implementation with Multisets and Evolution Rules Compression Workshop on Membrane Computing Eighth page 1 Optimizing Membrane.
Stephen McCray and David Courard-Hauri, Environmental Science and Policy Program, Drake University Introduction References 1.Doran, P. T. & Zimmerman,
Evolving Cutting Horse and Sheepdog Behavior with a Simulated Flock Chris Beacham Computer Systems Research Lab 2009.
Moderation: Assumptions
Zakaria A. Khamis GE 2110 GEOGRAPHICAL STATISTICS GE 2110.
1 Evolution of Networks Notes from Lectures of J.Mendes CNR, Pisa, Italy, December 2007 Eva Jaho Advanced Networking Research Group National and Kapodistrian.
Feedback Effects between Similarity and Social Influence in Online Communities David Crandall, Dan Cosley, Daniel Huttenlocher, Jon Kleinberg, Siddharth.
Non-Equilibrium Industry Dynamics with Knowledge-Based Competition: An Agent-Based Computational Model Myong-Hun Chang Department of Economics Cleveland.
Introduction to Genetic Algorithms Yonatan Shichel.
PSY 307 – Statistics for the Behavioral Sciences
Center for Evolutionary Computation and Automated Design Rich Terrile Symposium on Complex Systems Engineering Rand Corp. January 11, 2007 Rich Terrile.
Simulation Models as a Research Method Professor Alexander Settles.
Basic Logic of Experimentation The design of an Internally valid experimental procedure requires us to: Form Equivalent Groups Treat Groups Identically.
Chapter 2 – Tools of Positive Analysis
Why Globalized Communication may increase Cultural Polarization Paper presented at 2005 International Workshop Games, Networks, and Cascades Cornell Club.
1 ES 314 Advanced Programming Lec 2 Sept 3 Goals: Complete the discussion of problem Review of C++ Object-oriented design Arrays and pointers.
Hilton’s Game of Life (HGL) A theoretical explanation of the phenomenon “life” in real nature. Hilton Tamanaha Goi Ph.D. 1st Year, KAIST, Dept. of EECS.
Analysis of a Yield Management Model for On Demand IT Services Parijat Dube IBM Watson Research Center with Laura Wynter and Yezekael Hayel.
February 11, 2003Ninth International Symposium on High Performance Computer Architecture Memory System Behavior of Java-Based Middleware Martin Karlsson,
Innovation Economics Class 3.
Image Registration of Very Large Images via Genetic Programming Sarit Chicotay Omid E. David Nathan S. Netanyahu CVPR ‘14 Workshop on Registration of Very.
Genetic Algorithms: A Tutorial
1 Reasons for parallelization Can we make GA faster? One of the most promising choices is to use parallel implementations. The reasons for parallelization.
Efficient Model Selection for Support Vector Machines
Evolutionary Algorithms BIOL/CMSC 361: Emergence Lecture 4/03/08.
Introduction to the Management of Technology and Innovation TMGT1006 G. Brophey W’08.
Techniques for Analysis and Calibration of Multi- Agent Simulations Manuel Fehler Franziska Klügl Frank Puppe Universität Würzburg Lehrstuhl für Künstliche.
Week 3a Mechanisms for Adaptation. POLS-GEOG-SOC 495 Spring Lecture Overview Review –CAS –Principles of chaos How do systems “learn”? –“Credit.
Complex Adaptive Systems approach to Economic Development Ivan Garibay Director, Information Systems Group, ORC Joint Faculty, EECS Department Research.
Spacetime Constraints Revisited Joe Marks J. Thomas Ngo Using genetic algorithms to find solutions to spacetime constraint problems in 2D.
Requirements To Design--Iteratively Chapter 12 Applying UML and Patterns Craig Larman.
GATree: Genetically Evolved Decision Trees 전자전기컴퓨터공학과 데이터베이스 연구실 G 김태종.
Authors: Ioannis Komnios Sotirios Diamantopoulos Vassilis Tsaoussidis ComNet Group.
Planned AlltoAllv a clustered approach Stephen Booth (EPCC) Adrian Jackson (EPCC)
Genetic Algorithms Genetic algorithms imitate a natural optimization process: natural selection in evolution. Developed by John Holland at the University.
The Particle Swarm Optimization Algorithm Nebojša Trpković 10 th Dec 2010.
1 IE 607 Heuristic Optimization Particle Swarm Optimization.
A Queueing Model for Yield Management of Computing Centers Parijat Dube IBM Research, NY, USA Yezekael Hayel IRISA, Rennes, France INFORMS Annual Meeting,
Robin McDougall Scott Nokleby Mechatronic and Robotic Systems Laboratory 1.
Lecture 1 Introduction Figures from Lewis, “C# Software Solutions”, Addison Wesley Richard Gesick.
Evolving Virtual Creatures & Evolving 3D Morphology and Behavior by Competition Papers by Karl Sims Presented by Sarah Waziruddin.
1. Survey- obtain information by asking many individuals to answer a fixed set of questions 2. Case Study- an in depth analysis of the of a single individual.
Siddhartha Shakya1 Estimation Of Distribution Algorithm based on Markov Random Fields Siddhartha Shakya School Of Computing The Robert Gordon.
Gathering Useful Data. 2 Principle Idea: The knowledge of how the data were generated is one of the key ingredients for translating data intelligently.
Agent Based Modeling (ABM) in Complex Systems George Kampis ETSU, 2007 Spring Semester.
Outline The role of information What is information? Different types of information Controlling information.
This is unchecked growth:
SwinTop: Optimizing Memory Efficiency of Packet Classification in Network Author: Chen, Chang; Cai, Liangwei; Xiang, Yang; Li, Jun Conference: Communication.
Classification Ensemble Methods 1
Agent-Based Modeling in ArcGIS Kevin M. Johnston.
5. Implementing a GA 4 학습목표 GA 를 사용해 실제 문제를 해결할 때 고려해야 하는 사항에 대해 이해한다 Huge number of choices with little theoretical guidance Implementation issues + sophisticated.
New Product Development Page 1 Teddy Concurrent Engineering by Teddy Sjafrizal.
Particle Swarm Optimization (PSO)
Learning Behavioral Parameterization Using Spatio-Temporal Case-Based Reasoning Maxim Likhachev, Michael Kaess, and Ronald C. Arkin Mobile Robot Laboratory.
HORIZONTAL OFFENSE some principles apply to this offense so remember the golden rules! 1. Don’t throw over a players head, or in other words stay wide.
Learning in Complex Networks Christina Fang New York University Jeho Lee Korea Advanced Institute of Science and Technology Melissa A. Schilling New York.
1 KAIST Graduate School of Management 27 September, 2003 한상필 Organizational Learning Dynamics From managerial perspectives.
` Question: How do immune systems achieve such remarkable scalability? Approach: Simulate lymphoid compartments, fixed circulatory networks, cytokine communication.
An Evolutionary Algorithm for Neural Network Learning using Direct Encoding Paul Batchis Department of Computer Science Rutgers University.
CEng 713, Evolutionary Computation, Lecture Notes parallel Evolutionary Computation.
A Viewpoint-based Approach for Interaction Graph Analysis
Particle Swarm Optimization (2)
USING MICROBIAL GENETIC ALGORITHM TO SOLVE CARD SPLITTING PROBLEM.
Particle Swarm Optimization
AP Computer Science Principals Course Importance and Overview
Chapter 2 The Process of Design.
AP Computer Science Principals Course Importance and Overview
Presentation transcript:

The Parable of the Hare and the Tortoise: How "Small Worlds" Reduce the Long Run Performance of Systems David Lazer Program on Networked Governance Harvard University

Acknowledgements… Allan Friedman NSF grant

Living in the (self-consciously) networked age Growth of research on networks across disciplines We live in an “smaller world” with ever- accelerating flows of information Explosion of consultants, software, etc to make organizations “smaller”

Does connecting people help an organization solve problems?

The problem of parallel problem solving in human systems Many agents working on same problem simultaneously How is that problem solving aggregated?

Brainstorming

“Laboratories of democracy”

Global diffusion…

(Not) Re-inventing the wheel

Roadmap The role of informational diversity in systemic performance Networks as architecture for experimentation Description of model Results Conclusion

Role of informational diversity Sunstein, Nemeth, etc.  Informational diversity provides the menu of options in the system However: pressures toward homogeneity, some of which may increase system performance (e.g., the elimination of bad solutions)

Processes of emulation Neo-institutionalism– strong pressures for conformity (DiMaggio and Powell) Networks play a key conduit for those pressures (Lazarsfeld, Friedkin, Lazer) Convergence often not on system “optimum”, even when emulation is driven by success (Bikhchandani, Hirshleifer, and Welch; Strang and Macy)

Network structure Cliquish Small world– “six degrees of separation” (Milgram, Watts) Birds of a feather (Lazarsfeld and Merton) “Scale free” (Barabasi)  how does the architecture of the network affect balance between exploration and exploitation?

Cliques

Small worlds (Milgram, Watts and Strogatts) Big worldSmall world

Birds of a feather…

Scale Free networks (Barabasi)

Network structure Cliquish Small world– “six degrees of separation” (Milgram, Watts) Birds of a feather (Lazarsfeld and Merton) “Scale free” (Barabasi)  how does the architecture of the network affect balance between exploration and exploitation?

Computational model KISS principle– simplest possible model that captures some essence of reality Agent-based– decision rules dictating agent behavior based on local conditions (not analytically tractable) “Experimentally” manipulate parameters, test for robustness Key question: what systemic patterns emerge?

Model Problem space– what’s the problem agents are trying to solve? Agent decision rules– how do agents seek improvements in performance? Agent neighborhood– who do agents see (and emulate)?

Problem space Key attribute of problem space is its ruggedness

Easy to find optimum…

Less easy to find optimum…

Problem space NK model (Kauffman) N dimensions (19 in these simulations) The marginal contribution of each dimension to performance is contingent on K other dimensions K determines the ruggedness of the problem space (5 in most of these simulations) Scores are calculated using a rank-preserving monotonic transformation

Decision rule Capacity of agents to search problem space must be very limited

Decision rule If someone agent can see is doing better than agent at time t, copy best alternative. Otherwise, look at impact of randomly changing one dimension. If this is an improvement, move there. If not an improvement, stay at previous solution.

Informational velocity Always looking at others? If not: –Is communication synchronous (e.g., group meetings)? –Is communication asynchronous?

Network– determines neighborhood Linear (max degrees of separation = population size – 1) Fully connected (max degrees of separation = 1)

Basic model parameters 100 agents 200 time steps 1000 simulations of each experiment –20 NK spaces (N = 19, K = 5) –50 randomly seeded starting points Vary size, network structure, velocity, and synchronicity Code written in Java using the Repast libraries

Findings Size Network structure Velocity Synchronicity

Bigger is better

The hare and the tortoise: Small worlds are good for a quick fix…

…but not so good in the long haul

Small worlds drive out variety

LR Performance of random graphs

Small worlds

Impact of structure is contingent on problem space

Velocity increases exploitation and decreases exploration

Synchronicity.

.

.

.

Heterogeneity

The Social Structure of Exploration and Exploitation (March 1991) Exploration– looking for new solutions (experimentation) Exploitation– taking advantage of what the system knows (emulation)

Illustrations Agricultural diffusion Creative groups

Technological diffusion Diamond, Guns, Germs, and Steel “…[G]eographic connectedness has exerted both positive and negative effects on the evolution of technology. As a result, in the very long run, technology may have developed most rapidly in regions with moderate connectedness, neither too high nor too low. Technology’s course over the last 1,000 years in China, Europe, and possibly the Indian subcontinent exemplifies those net effects of high, moderate, and low connectedness, respectively.” (p. 416)

Creative groups Field work on creative groups suggests curvilinear relationship between performance and connectedness (Leenders) Experimental work on problem solving groups (Goldstone) Broadway (Uzzi and Spiro) Project teams (Binz-Scharf)

Conclusions Trade-off between networks that perform well in the short run vs long run –Small, high bandwidth, worlds good for SR, bad in LR Tragedy of the network: Trade-off between interests of individuals and system Are some networks better than others in both SR and LR? Are some networks good “compromises”?

Extensions Vary problem space Error in copying (crossover) Timing of “velocity” Assume some heterogeneity in problem space Make network endogenous Have landscape change

Genetic programming Holland, Koza, solution “breeding” Performs much better if there are multiple (largely) isolated populations, within which there is great intermixing and competition, between which there is little (> 0) Speciation