Transcript Advanced Intelligent Systems

Advanced Intelligent Systems
By
Dr.S.Sridhar,Ph.D.,
RACI(Paris),RZFM(Germany),RMR(USA),RIEEEProc
.
email : [email protected]
web-site : http://drsridhar.tripod.com
Learning Objectives
• Understand second-generation
intelligent systems.
• Learn the basic concepts and
applications of case-based systems.
• Understand the uses of artificial
neural networks.
• Examine the advantages and
disadvantages of artificial neural
networks.
• Learn about genetic algorithms.
• Examine the theories and
applications of fuzzy knowledge.
Household Financial’s
Vision Speeds Loan
Approvals With Neural
Networks
Vignette
• Loan product regulation varies in each
state
• Develop an object-oriented loan approval
system
• Neural network-based
• Fed risk, interest rate variables, customer data
• Estimates credit worthiness, potential for fraud
• Pattern recognition
• Integrates all loan approval phases
• Uses intelligent underwriting engine
• Reduced training time and administrative
overhead
• Decreased managed basis efficiency ratio
• Upgradeable to web-based architecture
Machine Learning
• Acquisition of knowledge through
historical examples
• Implicitly induces expert knowledge
from history
• Different from the way that humans
learn
• Implications of system success and
failure unclear
• Manipulates of symbols instead of
numbers
Methods
• Supervised learning
• Induce knowledge from known outcomes
•
•
•
•
•
•
New cases used to modify existing theories
Statistical methods
Rule induction
Case based and inference
Neural computing
Genetic algorithms leading to survival of fittest
• Unsupervised learning
• Determine knowledge from data with unknown
outcomes
• Clustering data into similar groups
• Neural computing
• Genetic algorithms leading to survival of fittest
Case Reasoning
• Inductive
• Case base used for decisionmaking
• Effective when rule-based
reasoning is not
• Case
• Primary knowledge element
• Ossified
• Paradigmatic
• Stories
Process
• Features assigned as character indexes
• Indexing rules identify input features
• Indexes used to retrieve similar cases from
memory
• Episodic case memories
• Similarity metrics applied
• Old solution adjusted to fit new case
• Modification rules
• Solution tested
• If successful, assigned value and stored
• If failure, explain, repair, test
• Alter plan to fit situation
• Rules for permissible alterations
Case Reasoning Success
Factors
• Specific business objectives
• Knowledge should directly support
end users
• Appropriate design
• Updatable
• Measurable metrics
• Acceptable ROI
• User accessible
• Expandable across enterprise
Human Brain
• 50 to 150 billion neurons in brain
• Neurons grouped into networks
• Axons send outputs to cells
• Received by dendrites, across synapses
Neural Networks
• Attempts to mimic brain
functions
• Analogy, not accurate model
• Artificial neurons connected in
network
• Organized by topologies
• Structure
• Three or more layers
− Input, intermediate (one or more hidden
layers), output
• Receives modifiable signals
Processing
• Processing elements are neurons
• Allows for parallel processing
• Each input is single attribute
• Connection weight
• Adjustable mathematical value of input
• Summation function
• Weighted sum of input elements
• Internal stimulation
• Transfer function
• Relation between internal activation and output
− Sigmoid/transfer function
− Threshold value
• Outputs are problem solution
Architecture
• Feedforward-backpropogation
• Neurons link output in one layer to input
in next
• No feedback
• Associative memory system
• Correlates input data with stored
information
• May have incomplete inputs
• Detects similarities
• Recurrent structure
• Activities go through network multiple
times to produce output
Network Learning
• Learning algorithms
• Supervised
• Connection weights derived from known cases
• Pattern recognition combined with weighting
changes
• Back error propagation
Easy implementation
Multiple hidden layers
Adjust learning rate and momentum
Known patterns compared to output and allows for
weight adjustment
− Established error tolerance
−
−
−
−
• Unsupervised
• Only stimuli shown to network
• Humans assign meanings and determine
usefulness
• Adaptive resonance theory
• Kohonen self-organizing feature maps
Development of Systems
• Collect data
• The more, the better
• Separate data into training set to adjust weights
• Divide into test sets for network validation
• Select network topology
• Determine input, output, and hidden nodes, and
hidden layers
• Select learning algorithm and connection weights
• Iterative training until network achieves preset
error level
• Black box testing to verify inputs produce
appropriate outputs
• Contains routine and problematic cases
• Implementation
• Integration with other systems
• User training
• Monitoring and feedback
Genetic Algorithms
• Computer programs that apply
processes of evolution
• Viability of candidate solutions
• Self-organized
• Adaptable
• Fitness function
• Measured by objective obtained
• Iterative process
• Candidate solutions combine to produce
generations
• Reproduction, crossover, mutation
Genetic Algorithms
• Establish problem
• Parameters
•
•
•
•
• Number of initial solutions, number of offspring,
number of parents and offspring for each
generation, mutation level, probability distribution
of crossover point occurrence
Generate initial set of solutions
Compute fitness functions
Total all fitness functions
Compare each solution’s fitness function
to total
• Apply crossover
• Apply random mutation
• Repeat until good enough solution or no
improvement
Fuzzy Logic
•
•
•
•
•
Mathematical theory of fuzzy sets
Imprecise thinking
Describes human perception
Continuous logic
Not 100% true or false, black or
white
• Fuzzy neural networks
• Fuzzification
• Fuzzy logic applied to input and output used
to create model
• Defuzzification
• Model converted back to original input, output
scales