evaluating resilience strategies based on an evolutionary multi agent system
DESCRIPTION
Evaluating Resilience Strategies Based on an Evolutionary Multi agent System. Kazuhiro Minami, Tomoya Tanjo , and Hiroshi Maruyama Institute of Statistical Mathematics, Japan December 4 , 2013 CyberneticsCom 2013. We sometimes have an unexpected event. 9.11 - PowerPoint PPT PresentationTRANSCRIPT
Evaluating Resilience Strategies Based on an Evolutionary Multi agent System
Kazuhiro Minami, Tomoya Tanjo, and Hiroshi Maruyama
Institute of Statistical Mathematics, Japan
December 4, 2013CyberneticsCom 2013
We sometimes have an unexpected event
• 9.11• Lehman financial shock
in 2008
• 3.11 earthquake and tunami
7/31/2012 Kazuhiro Minami 2
• We cannot completely prevent such disasters• Instead, we should aim to design a system that contains a damage
and is readily recoverable to an acceptable level
Resilience: Definition
“Capacity of a (social-ecological) system to absorb a spectrum of shocks or perturbations and to sustain and develop its fundamental function, structure, identity, and feedbacks as a result of recovery or reorganization in a new context.”
-- by Buzz Holling (1973)
7/31/2012 3Kazuhiro Minami
Resilience = Resistance + Recovery
Taoi-cho, Miyagi Pref.http://www.bousaihaku.com/cgi-bin/hp/index2.cgi?ac1=B742&ac2=&ac3=1574&Page=hpd2_view http://fullload.jp/blog/2011/04/post-265.php
+
Logstaff et al., “Building Resilient Communities,” Homeland Security Affairs, Vol VI, No.3, 2010
7/31/2012 Kazuhiro Minami 4
Goal: How to make our systems more resilient against large unexpected events?
5Financial Systems
Civil Infrastructure
Engineering Systems
Society
Organizations
Natural Disasters
Financial Crisis
New Technologies
Malicious Attackers
Biological science might be a major source of wisdom for resilience engineering
6
Redundancy
Diversity Adaptability
Multiple pathwaysfor metabolism
Redundancy and diversity are heavily used techniques in Computer Science
• Maintain a backup system in a cloud service– Financial companies was able to continue their services
after 9.11 event– Many web sites maintain multiple copies of the server
• Software diversity makes it difficult for hackers to compromise multiple servers of the same service– Change compiler options or use different algorithms
• Ethernet uses a randomization technique to avoid message collision
7
However, applying those techniques to real-world systems is NOT so trivial
• Cost for replication would be high in NON-ICT systems
• Replication sometimes decreases the quality of service– Inconsistency of data– Timely monitoring of a system is more difficult; thus
need to sacrifice the adaptability of a system• Toyota’s supply chain system put precedence
on adaptability over redundancy8
Multi-agent simulationsbased on a population genetics model
Colony of n agents Each robot has ten binary features (e.g., 2-leg/4-leg, flying/non-flying, …)E.g., <0110111011>
C: “fit” configurationsResource
• Resource Reserve R– Fit robots contribute to build up R – A robot consumes one unit for reconfiguring its one feature
• The colony is resilient if robots can survive a series of changing constraints C1, C2, …, Ct, …
Constraint CA Subset of 2(set of all 1,024
configurations)
A robot is fit if its configuration is in C
9
Represent a changing environment as a sequence of dynamic constraints
10
Ct
`
Ct+1
Time t Time t+1
fit
fit
fit fit
fit
unfit
unfit
unfit
Need to pay a cost for adaptation
11
Resource
Adaptation10110010 10110011 10110011System
bitstring
Unfit fit
Remove Add
An adaptation in our model is much faster than that in biological systems
Adaptation
A robot could produce a clone or die
• Make a clone– when the amount of the resource is doubled
• Die – when the resource is used up
12
Metrics of resilience in our model
• Redundancy– How much resource does a robot maintain?
• Diversity – Diversity index
• Adaptability– How many bits a robot can flip at a time?
13
Multi-agent Simulations• Define initial parameters
– Population size– Bit length of a robot– Size and type of constraints– Initial amount of each robot’s resource– Initial diversity index– Adaptation strategy
• Random or intelligent• #flips at a time
• Run the system at 100 time steps• Examine how a population size, the diversity index vary
over time
14
Diversity at the beginning helps a population survive longer
15
Parameter Value
Initial population size
100
Agent bit length
8
Constraint size 26
Constraint transition
continuous
Adaptation strategy
random
Adaptation speed
1
Time
#Age
nts
Two adaptation Strategies
1. Random strategy (flip one bit randomly)
2. Intelligent strategy (flip one bit to be closer to the constraint)
16
10110110
Constraint
If robots adapt intelligently, the population grows much faster
17
Time
#Age
nts
Time
If agents share the common resource, the sustainability of a system can be greatly improved
18
Sharedresource
Individualresources
Sudden changes of the constraint
Sudden changes of the constraint
Summary• Explore design space parameterized by three
resilience properties based on an evolutionary multi-agent system– Redundancy– Diversity– Adaptability
• Obtain quantitative initial results regarding design strategies for building resilient systems
19
Future work: Further possibilities for adaptation strategies
• Local vs Global– Local: Each robot makes its own decision independently
from others– Global: There is a global coordination. Every robot must
follow the order– Mixed
• Complete vs Incomplete knowledge on C– Complete knowledge: max 10 steps to become fit again– Incomplete knowledge: probabilistic (max 1023 steps if
the landscape is stable)
20
21
Backup
We consider three types of constraints
22
1. Disruptive changes: a new constraint Ct is generated randomly at each time t
2. Small changes: a new constraint Ct is generated from Ct-1 by adding a neighbor configuration into Ct-1 or removing a configuration in Ct-1
T = tT = t-1 T = t+1
T = tT = t-1 T = t+1
3. Small changes with continuous topology: Same as case 2, but all configurations in Ct are connected
T = tT = t-1 T = t+1
Measure diversity considers population abundance of each type
23
where N is the size of a population and pi is the size of an individual i
Example 1: if N=5, Pr(`1101’) = 5, then D = 52/52 = 1
Example 2: if N=5, size(`1101’) = 3, and size(`1111’) = 2, then D = 52/32+22 = 25/13 = 1.92