gbml
DESCRIPTION
Genetics Based Machine LearningTRANSCRIPT
Introduction to Genetics-based
Machine Learning
Mahalingam.P.R
Semester III, M.Tech CSESIS, RSET
Contents
Background of Machine Learning
About Machine Learning
GA in Machine Learning
Evolution of GBML
Classifier System
Rule and Message System
Apportionment of Credit System
Genetic Algorithm
A simple classifier system in Pascal
Conclusion
2 Introduction to GBML
Introduction
Introduction to GBML3
Machine Learning
A machine learns whenever it changes its structure,
program, or data (based on its inputs or in response
to external information) in such a manner that its
expected future performance improves.
Machine learning usually refers to the changes in
systems that perform tasks associated with artificial
intelligence (AI).
Such tasks involve recognition, diagnosis, planning,
robot control, prediction, etc. The “changes" might be
either enhancements to already performing systems
or ab-initio synthesis of new systems.
4 Introduction to GBML
A typical AI system
Here, the model progressively learns from the experience it earns from the environment
Introduction to GBML5
Need for Machine Learning
Some tasks cannot be defined well except byexample. We might be able to specify input/output pairs but not a
concise relationship between inputs and desired outputs.
We would like machines to be able to adjust their internalstructure to produce correct outputs for a large number ofsample inputs
suitably constrain their input/output function to approximate therelationship implicit in the examples.
It is possible that hidden among large piles of dataare important relationships and correlations. Machine learning methods can often be used to extract
these relationships (data mining).
6 Introduction to GBML
Machine learning methods can be used for on-the-jobimprovement of existing machine designs.
The amount of knowledge available about certain tasksmight be too large for explicit encoding by humans. Machines that learn this knowledge gradually might be able to
capture more of it than humans would want to write down.
Environments change over time. Machines that can adapt to a changing environment would
reduce the need for constant redesign.
New knowledge about tasks is constantly beingdiscovered by humans (Vocabulary changes). There is a constant stream of new events in the world.
Continuing redesign of AI systems to conform to newknowledge is impractical.
Machine learning methods might be able to track much of it.
7 Introduction to GBML
More on Machine Learning
Contributions to Machine Learning
Varieties of Machine Learning
Introduction to GBML8
Statistics
Brain Models
Adaptive Control Theory
Psychological Models
Artificial Intelligence
Evolutionary Models
Functions
Logic Programs and
Rule sets
Finite-state machines
Grammars
Problem solving
systems
GA in Machine Learning
Introduction to GBML9
Conventional GA systems work with the followingproperties: Probabilistic
Random
Enumerative
The structures adapted cater to the human-likemechanism adopted in the methodology.
This adaptation itself is the biggest hindrance whenincorporating GA into more complex, less completelydefined environments. Too much “guesses” when it comes to searching
Methodology safe under the “sandbox” of searching
Coming Up
Origins of GBML
Systems
Classifier System
Operation of Classifier
Systems
Implementation of a
Simple Classifier System
in Pascal
Testing classifier in a
problem domain –
Learning a Boolean
function
Introduction to GBML10
Study Machine Learning systems
that use genetic search as their
primary discovery heuristic.
Adapting GA structures to work in
complex environments like Machine
Learning
Origin of GBML systems
Introduction to GBML11
Introduction to GBML12
In nature, not only do individual animals learn to
perform better, but species evolve to be better fit in
their individual niches.
Since the distinction between evolving and learning
can be blurred in computer systems, techniques that
model certain aspects of biological evolution have
been proposed as learning methods to improve the
performance of computer programs.
Genetic algorithms [Holland, 1975] and genetic
programming [Koza, 1992, Koza, 1994] are the most
prominent computational techniques for evolution.
Theoretical Foundation for GBML
Introduction to GBML13
Laid by Holland (1962)
Outline for Adaptive Systems Theory
role of program replication as a method of emphasizing
past programs.
Fundamental role of recombination
Holland (1965)
Schemata Processors
Holland (1968-1971)
Application of a classifier system
Holland and Reitman (1978)
Introduction to GBML14
Modern classifier systems resemble schemata
processors in both outline and detail
Holland suggested four prototypes in the initial
proposal.
No experiments or implementation have been reported
yet!
Proposal coincided with the development of the
theory of schemata.
Prototype 1 Prototype 2
Introduction to GBML15
Stimulus-response (SR)
processor that would
link environmental
schemata (conditions)
with particular action
effectors.
Extend type 1 by adding
internal effectors
(internal states).
Prototype 3 Prototype 4
Introduction to GBML16
Build upon types 1 and
2 by including explicit
environmental state
prediction (a model of
the real world), and an
internal evaluation
mechanism.
Extend the other types
by incorporating the
capability to modify its
own effectors and
detectors, permitting
greater (or lesser) range
of data detection and a
larger behavioural
response.
Broadcast language
Introduction to GBML17
Derived from early proposals.
Broad
Unimplemented
Creating broadcast units (production rules) from a 10-letter alphabet.
The alphabet added wild card (single and multiple match)characters to an underlying binary alphabet.
If included, the following would have given sufficientpower for computational completeness andrepresentational convenience.
A fundamental punctuation mark, a persistence symbol forcontinued broadcast of a message.
A quotation character for taking the next character literally.
Introduction to GBML18
The proposal for broadcast language was
instrumental in unifying the earlier suggestions for
schemata operators by:
theoretically permitting a consistent representation of all
operators, data, and rules or instructions.
The generality gained in theory has not been
realized in practice.
First practical implementation
Introduction to GBML19
Three years following the broadcast languageproposal (Holland and Reitman, 1978).
Cognitive System Level One (CS-1) Trained to learn two maze-running tasks.
Performance system with message list and simple stringrules (classifiers).
GA composed of reproduction, crossover, mutation andcrowding.
Epochal learning mechanism where reward isapportioned to all classifiers active between successivepayoff events.
Learning mechanism has largely been supplanted byanother mechanism – a bucket brigade.
GBML Applications
Introduction to GBML20
Introduction to GBML21
Introduction to GBML22
Introduction to GBML23
Introduction to GBML24
Classifier System
Introduction to GBML25
Introduction to GBML26
A classifier system is a machine learning system that
learns syntactically simple string rules (called
classifiers) to guide its performance in an arbitrary
environment.
A classifier system consists of three main
components:
Rule and message system
Apportionment of credit system
Genetic algorithm
Introduction to GBML27
The rule and message system of a classifier system
is a special kind of production system.
A production system is a computational scheme that
uses rules as its only algorithmic device. The general
syntax of such systems are as follows:
if <condition> then <action>
The production means that the action may be taken
(rule is “fired”) when the condition is satisfied.
Introduction to GBML28
It has been shown that production systems are
computationally complete, as well as convenient.
A simple rule or set of rules can represent a complex
set of thoughts compactly.
But when it comes to situations in need of learning,
this is not advisable due to the complex rule syntax.
Many production systems permit involved
grammatical constructions for the condition and
action part of the rule.
Introduction to GBML29
Classifier systems depart from the mainstream by
restricting a rule to a fixed length representation.
This restriction has two benefits.
All strings under the permissible alphabet are syntactically
meaningful
A fixed string representation permits string operators of
the genetic kind. This leaves the door open for a Genetic
Algorithm search of the space of permissible rules.
Introduction to GBML30
Classifier systems use parallel rule activation,
whereas traditional systems use serial rule
activation.
They permit multiple activities to be coordinated
simultaneously. When choices must be made
between mutually exclusive environmental actions,
or size of matched rule set must be pruned to
accommodate the fixed length message list, the
choices are postponed to the last possible moment,
and the arbitration is then performed competitively.
Introduction to GBML31
In traditional expert systems, the value or rating of a rule relative to other is fixed by the programmer in conjunction with the expert or group of experts to be emulated. But this is not possible in rule learning system.
In such cases, the relative importance has to be “learned”. So, classifiers are forced to coexist in an information-based service economy. A competition is held among classifiers where the right to answer relevant messages goes to the highest bidder. Subsequent bids serve as the income for previously successful message senders.
This competitive nature ensures that good rules (profitable ones) survive and the bad rules (unprofitable) die off.
Internal Currency
Introduction to GBML32
The exchange and accumulation of an internal
currency provides a natural figure of merit for
applying GA.
Using the “bank balance” as a fitness function,
classifiers may be reproduced, crossed, and
mutated.
So, it can rank extant rules, and discover new,
possibly better rules by innovative combinations of
old ones.
Here, the stress is on “who gets replaced”, not
“replace entire populations”.
Next…
Each component is
discussed in detail and
the interconnections
studied.
•Rule and message
system
•Apportionment of credit
•Genetic Algorithm
Implementation in Pascal
Testing on real world
problem – learning a
Boolean function
Introduction to GBML33
Thus, apportionment of credit via
competition and rule discovery
using GA form a reasonable basis
for constructing a machine learning
system atop the computationally
convenient and complete
framework of classifiers.
Rule and Message System
Introduction to GBML34
This is the schematic of a
complete classifier
system.
Introduction to GBML35
Introduction to GBML36
The rule and message system forms the backbone of the silicon beast.
Information flows from the environment through the detectors, where it is decoded to one or more finite length messages.
The environmental messages are fed into a finite length message list where the messages activate string rules called classifiers.
When activated, the classifier posts a message on to the message list, which may in-turn invoke other classifiers, or cause an action to be taken through the system’s action triggers called effectors.
Introduction to GBML37
Classifiers combine environmental cues and internal
thoughts to determine what the system should do
and think next.
Coordinates the flow of information from where it is
sensed (detectors) to where it is processed
(message list and classifier store) to where it is
called to action (effectors).
Informational units
Messages
Classifiers
Introduction to GBML38
A message within a classifier system is simply a
finite-length string over some finite alphabet. If we
limit to the binary alphabet, we get
<message> ::= {0, 1}L
This means taking the concatenation of 0s or 1s for
“L” times.
Messages are the basic token of information
exchange in a classifier system.
The messages on the message list may match one
or more classifiers or string rules.
Introduction to GBML39
A classifier is a production rule with the syntax
<classifier> ::= <condition>:<message>
The condition is a simple pattern recognition device
where a wild card (#) is added to the underlying alphabet.
<condition> ::= {0, 1, #}L
So, the condition matches a message if at every position,
a 0 in the message matches a 0 in the condition, 1
matches a 1, or a # matches either a 0 or a 1.
For example, #01# matches 0010, 0011, 1010 and 1011. But it
doesn’t match 0000.
Similar to schema in GA
Introduction to GBML40
Once the classifier’s condition is matched, it
becomes a candidate to post its message to the
message list in the next time step.
Whether the candidate classifier posts its message
is determined by the outcome of an activation
auction, which in turn depends on evaluation of the
classifier’s value or weighting.
Sample simulation by hand
Introduction to GBML41
Initial message list is as follows.
Introduction to GBML42
Introduction to GBML43
Introduction to GBML44
Introduction to GBML45
Introduction to GBML46
Introduction to GBML47
Introduction to GBML48
Introduction to GBML49
Apportionment of Credit
Algorithm
THE BUCKET BRIGADE
Introduction to GBML50
Introduction to GBML51
Many classifiers attempt to rank or rate the individual
classifiers according to a classifier’s role in achieving
reward from the environment.
Most prevalent method
Bucket Brigade Algorithm
John Holland
About the algorithm
Introduction to GBML52
An information economy where the right to trade
information is bought and sold by classifiers.
Classifiers form a chain of middlemen from
information manufacturer (the environment) to
information consumer (the effectors).
Components of service economy
Auction
ClearingHouse
Introduction to GBML53
When classifiers are matched, they don’t directly
post their messages.
A matching message entitles a classifier to
participate in an activation auction.
Each classifier maintains a record of its net worth.
Called Strength (S).
Each matched classifier makes a bid B.
Bid proportional to its strength.
In this way, rules that are highly fit (have
accumulated a net worth) are given preference over
other rules.
Introduction to GBML54
Once a classifier is selected for activation, it must clear its payment through the clearinghouse, paying its bid to other classifiers for matching messages rendered.
A matched and activated classifier sends its bid B to those classifiers responsible for sending the messages that matched the bidder classifier’s condition.
Bid amount divided among the matching classifiers.
Division of payoff among contributing classifiers helps ensure the formation of an appropriately sized subpopulation of rules.
Different types of rules can cover different types of behavioral requirements without undue interspecies competition.
Introduction to GBML55
In a rule-learning system of any experience, we
cannot search for one master rule.
We must instead search for a co-adapted set of rules
that together cover a range of behavior that provides
ample payoff to the learning system.
Consider the
classifiers as before.
Introduction to GBML56
Now, lets follow the payments, with initial strength of
200.
Introduction to GBML57
Introduction to GBML58
Introduction to GBML59
Introduction to GBML60
Introduction to GBML61
Introduction to GBML62
Introduction to GBML63
For steady receipts, the bid value approaches the
receipt. For time-varying receipt values, we see that
the bid is a geometrically weighted average of the
input.
As such, it acts as a filter of the possibly intermittent
and noisy receipt values.
Genetic Algorithm
Introduction to GBML64
Introduction to GBML65
Bucket brigade
Clean procedure
Evaluation of rules
Decide among competing alternatives
But we have to devise a way of injecting new rules
into the system.
Similar to SGA, we can inject new rules using the
tripartite rules
Reproduction
Crossover
Mutation
Introduction to GBML66
The rules are placed in the population and processed by the auction, payment and reinforcement mechanism to properly evaluate their role in the system.
Pay attention to “who replaces whom” Not replacing the entire population
GA in classifier systems strongly resemble those used in search and optimization.
Main difference in Machine Learning
Non-overlapping population model not acceptable here.
In non-overlapping generations, complete generations are selected and replaced by a new population at every run.
Machine Learning Search and optimization
Introduction to GBML67
High level of on-line
performance.
Learn to perform more
proficiently.
Convergence
Offline performance
De Jong’s experiments Machine Learning
Introduction to GBML68
Conventional system
GA Parameter
Generation Gap (G)
Implement and test
overlapping population
Genetic Algorithms.
Whole population
should not be replaced.
Quantity
Selection Proportion
(proportion)
Replace that proportion
of the population at a
given algorithm run.
Coupled with a number
of other parameters.
Other Parameters
Introduction to GBML69
GA Period
Represented as Tga
Specifies number of time steps (rule and message cycles)
between GA calls.
Period can be treated deterministically
GA is called every Tga cycles
Or stochastically
GA is called probabilistically with average period Tga
Invocation of GA learning may be conditioned on
particular events such as:
Lack of match
Poor performance
Selection
Introduction to GBML70
Roulette Wheel Selection
Classifier’s strength S used as the fitness.
No longer generating entire populations
Careful when choosing population members for
replacement.
De Jong’s crowding procedure
Encourage replacement of similar population memebers
Mutation
Introduction to GBML71
Modified procedure
Here, ternary alphabet is used
SGA used a binary alphabet
Mutation probability pm defined as before
When a mutation is called for, we change the
mutated character to one of the other two with equal
probability.
0 { 1, # }
1 { 0, # }
# { 0, 1 }
Next-
A simple Classifier
System in Pascal
•Simple Classifier System
Data Structure
•The Performance System
•Apportionment of credit
algorithm
•Genetic search within the
Simple Classifier System
•Real-world testing
•Results
•Comparison with and
without GA
Introduction to GBML72
With all these changes to the
normal SGA routine, GA may be
dropped into the classifier system
and used in a manner not too
different from normal search and
optimization applications.
A Simple Classifier System in
Pascal
Introduction to GBML73
Introduction to GBML74
Construct a system designed to learn a boolean
function
A multiplexer
Collapse the finite-length message list to a single
message (the environmental message)
Immediate feedback
Simple payoff
Components
Introduction to GBML75
Simple Classifier System Data Structure Adapt to learning strategies
Performance System Heart of SCS
Matching procedures are the heart of the performance system
Apportionment of Credit Procedures
Auction
Clearinghouse
Taxcollector
Genetic Search within Simple Classifier System Similar to SGA
Learning the multiplexing system
Main procedure Reinforcement
Six – bit multiplexing system
Introduction to GBML76
Results using the Simple Classifier System
Without GA With GA
Introduction to GBML77
Conclusion
Introduction to GBML78
The following were discussed
Machine Learning
Role of GA in Machine Learning
The evolution of GA concepts in Machine Learning
Some applications of GBML
Classifier System
Components of Classifier System
Rule and Message System
Apportionment of Credit (The Bucket Brigade)
Genetic Algorithm
A practical implementation in Pascal
Increase in output efficiency when using GA
References
“Introduction to Machine Learning”, Nils J. Nilsson,
Robotics Laboratory, Department of Computer
Science, Stanford University
“Genetic Algorithms in Search, Optimization and
Machine Learning”, David E. Goldberg, pp 217-260
79 Introduction to GBML
THANK YOU
80 Introduction to GBML