Transcript Agent-Based Hybrid Intelligent Systems and Their Dynamic

Agent-Based Hybrid Intelligent
Systems and Their Dynamic
Reconfiguration
Zili Zhang
Faculty of Computer and Information Science
Southwest University
[email protected]
Acknowledgement: Pengyi Yang, Li Tao
Roadmap
• Background and Motivation
• The Genetic Ensemble (GE) Hybrid System
–
–
–
–
Problem Definition
Overview of the GE System
Experimental Results
Advantages of the GE System
• Dynamic Reconfiguration
– Dynamic Reconfiguration Model
– The Algorithm for Dynamic Reconfiguration
– Experiments and Evaluation
• Conclusions
Background and Motivation
• A few years ago, we proposed an agentbased framework for complex problem
solving
(Z. Zhang & C. Zhang, Agent-Based Hybrid
Intelligent Systems: An Agent-Based
Framework for Complex Problem Solving,
LNAI2938, Springer, 2004.)
• This framework was applied to:
– Financial Investment Planning (PRICAI’02)
– Data Mining etc (Applied AI, 2003)
Our Understanding
• Hybrid approaches are required for many
complex problems
• Better results can be achieved when these
techniques are combined in hybrid
intelligent systems
• Agent technology is well suited to model
the manifold interactions among the many
different components of hybrid intelligent
systems
In Bioinformatics
• Hybrid algorithms can be used to solve a variety
of problems in bioinformatics
• The hybrid algorithm often improves on either a
single algorithm, in terms of performance, or in
the transparency of its results
• There are numerous ways in which algorithms
can be combined
(E. Keedwell and A. Narayannan: Intelligent
Bioinformatics – The application of AI
techniques to bioinformatics problems, Wiley,
2005 [Ch. 11])
The GE Hybrid System
• Z. Zhang. P. Yang, X. Wu, and C. Zhang, An
Agent-based Hybrid System for Microarray
Data Analysis, IEEE Intelligent Systems,
Sept./Oct. 2009.
Problem Definition
• To identify biologically important genes and to improve
data classification accuracy, we need to develop
effective gene selection measures.
• Currently, there are a lot of gene selection strategies
available. Several studies set out to compare the
strength and the weakness of each method. The
comparison results tell us that with different datasets
different methods often perform unevenly.
• Instead of choosing one method for one situation, we
combine different methods to create a more accurate
and general hybrid system—the GE system.
Classifiers
• Integrated classifiers:
–
–
–
–
–
Decision Tree (DT)
Random Forest (RF)
7-Nearest Neighbors (7NN)
Naïve Bayes (NB)
3-Nearest Neighbors (3NN)
Blocking Committee
• Suppose a total of n classifiers each creates a
different hypothesis denoted as hi (s),(i  1, , n)
while classifying the data using gene subset s.
The fitness function derived from blocking
integration strategy can be defined as follows:
n
fitnessb ( s )  arg min  C (hi ( s ), y )
yY
i 1
Where y is the class label and C(.) is the accuracy
Evaluation function which can be calculated by cross
Validation etc.
Voting Committee
• Suppose a total of n classifiers each creates a different
hypothesis denoted as hi ( s),(i  1, , n) while
classifying the data using gene subset s. The fitness
function derived from voting integration strategy can
be defined as follows:
k
fitnessv ( s )  arg max  V (hi ( s ), y )
yY
i 1
Where y is the class label, k is the number of classifier
in voting, and V(.) is the voting function of a given
classifier.
Filters
• Gain Ratio
d
Gain( g j )
j 1
Split ( g j )
fitnessg ( s )  
• ReliefF
d
fitnessr ( s)   ReliefF ( g j )
j 1
where d denotes the number of genes in
subset s, and g j is the jth gene in subset s.
Why Choosing these algorithms?
• We based the selection of classifiers for
GE hybrids on their sample classification
accuracy and diversity from other
classifiers.
• For the filter components, we favored
those that are consistent with classifier
components.
• By Implementing different algorithms in terms of
agents, we can test different combinations by
forming different multiagent systems at runtime
rather than at design time.
• We no longer need to modify the code each time
we test a new combination.
• Therefore, in a given time frame we can
investigate many more algorithm combinations
than we could using traditional methods.
Experimental Results
• Datasets
• Gene Subset Size Determination
• Classification Results
• Reproducibility
Advantages of the GE System
• Evaluate gene subsets from multiple aspects.
• Select better generalization gene profiles.
• Avoid selection bias of certain feature evaluation method.
• Allow different selection and evaluation methods to be
integrated at ease.
• Achieve higher classification accuracy while also obtain
better selection stability
Dynamic Reconfiguration
• Given the scale and complexity of real-world
applications, it is increasingly important that
complex software systems operating in dynamic
operational environments call for dynamic
reconfiguration (self-configuration, selforganization) property.
• Such software systems should reorganise
automatically with environments and tasks
changes as we can not pre-define everything at
the design time.
Observations
• Today’s complex problem solving systems
often fail because of the wrong
configuration or static organization.
• In this context, to fail means to produce
incorrect solutions, solutions of
unacceptable bad quality or to spend
unreasonable time and resources.
How to Model Dynamic
Configuration
• Autonomic computing perspective
• Self-organization perspective
• Ecological perspective
(F. Zambonelli, Self-Management and Many
Facets of “Nonself”, IEEE IS, March/April,
2006, pp. 53-55.)
Autonomic computing perspective
Self-organization perspective
Ecological perspective
Using AOC to Model Dynamic
Reconfiguration
• Autonomy-Oriented Computing (AOC) is an
emerging computational paradigm that draws on
the principles of self-organization and complex
systems [Jiming Liu et al]
• A formal framework of AOC consists of a
population of autonomous entities and the rest
of the system referred to as the environment.
• An autonomous entity consists of a detector (or
a set of such), an effector (again, there can be a
set of such) and a repository of local behavior
rules.
• Specifically, using AOC framework to
model dynamic reconfiguration of agentbased systems, we need to clearly
describe what are the environment,
primitive behaviors and behavioral rules of
autonomous entities (here agents), and
the interactions between agents and their
environment.
• Once we have done these, the dynamic
reconfiguration of agent-based systems is
then reduced to the self-organization of
AOC systems.
• A heuristic algorithm (HIERA) has been
developed to support the organization
formation behavior in dynamic
reconfiguration.
The HIERA algorithm
• Initialization: initialize the parameters of
CAgents, including searching steps, primitive
behaviors probability, and so on.
• Search: use local-searching algorithm to search
appropriate MAgent.
• Behavior Selection: use Roulette method to
select primitive behaviors according to different
behavior probability, and use heuristic rules to
modify behavior probabilities
• Based on the experiments conducted,
HIERA algorithm can converge in an
acceptable time
• It can change its searching strategy
adaptively according to different
environments.
Conclusions
• Hybrid approaches are required for many
complex problems
• Agent-based approaches are suitable for
building hybrid systems in general, and
that a genetic ensemble system is
appropriate for microarray data analysis in
particular.
• Dynamic reconfiguration is crucial for
complex software systems
Future Work
• Trustworthiness of agent-based systems
with dynamic reconfiguration capability
– Trusted computing initiative
– Trusted Software initiative in China
• Questions and Comments?