0% found this document useful (0 votes)
7 views6 pages

Unit 5 Long ML

Uploaded by

krgptuse123
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
7 views6 pages

Unit 5 Long ML

Uploaded by

krgptuse123
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 6

MACHINE LEARNING [JNTU-HYDERABADI

5.12 control can be formulated ha


Theproblem oflearning search prok
This domain theory do notentail any sort of conclusions example, consider a general search
byapplying EBL. For legal s
bascd on instances of the target concept play cricket. If learier search operators, s is set of possible a
ohserves a positive example day where humidity = 30, then lem where O is setof legalpredicate defined over s that denotes rul
search states and Gis a
domain theory entails the following hypothesis problem is to find a sequence of
states that are goal states. The
state i.e., S, to final state ie
(Playcricket yes) Humidity<30) operators that transform initial
This example represents a situation where B Vh, BAD satisfies predicate G. One approach to formulate learnin
that understand about the taroet
-h. The learned hypothesis in this scenario entails the predic problemn is to make the system
tions that are not done by domain theory. concept for every operator in O.
enhance its
Knowkedge level learming is sometimes referred to this The system that employs the EBL must independent
related
type of vhere learned hypothesis entails predictions searching known as prodigy. It is a domain domain problem
learningWICewhich
and conclusions that go be
domain theory. Th
are entailed by the
of all the:predret
dictions entailed by a set
planning system that accept the definition of
sed on operators 0 and state space S. For eXample, Consider
based
of assertions is called deductive closure. In knowledge level a problem space finding a sequence of of operators
opera that Cus on
learning, the deductive closure of B is an appropriate subset of identifying initial state s to a state that satisfies goal predicate
deductive closure of B + h. means-ends planner that
G. The prodigy system generally uses Here
Another example of knowledge level learning of ana breaks the problems into sub problems and solves them.
1o
Iytical methods is provided by a type of assertions called as all the sub solutions are combined into one solution inorder
determinations. They assert few attributes of instance that are find the initial problem. This technique, raises questions like
determined by specific attributes without denoting th» nature of Which operator need to be considered for solving sub
the dependence. Consider a target concept "People who speak (1
goals?
french
is given ase"the language spoken by the person is determined
h" and the domain theory for single determination assertion
(11) Which subgoal is solved next?
by their respective nationality". This target concept of domain
theory do not enable to classification of instances as either posi The rule learned by prodigy for the given target concept
tive or negative. Although, it can be observed that if "Rose, a in a simple block stacking domain problem is given as
20 year old left handed Canadian speaks french". Then it can follows,
be concluded that "all Canadians speak french".
Both examples depicts that the deductive learning pro If one subgoal is to be solved, on (x, y) and another
vides output hypothesis h that are not entailed by domain theory. subgoal must be solved on (y, z) then solve the subgoal on (y.
In both scenarios h satisfies BA DEh and do not satisfy B z) before (x, y).
+h. The learner deduces a justified hypothesis in both cases Consider an example where the goal is to stack the blocks
which do not follow either training data or domain theory: such that word "universal" is spelled. The prodigy system di
vides this problem into various sub goals involving on (v, ),on
5.1.4 Explanation Based Learning of Search (N, ), on (1, V) and so on. The above mentioned rule is similar
Control Knowledge to the goals of on(U, N) and on (N, I), Thus, on (U, N) need to
Q25. Explain in detail about the EBL of search con be solved before solving on (N, I).
trol knowledge.
The main usage of EBL 0s to obtain control knowledge for
Answer : prodigy been specified in several problem domains with ssimple
block stacking problems, planning problems and complex
Explanation Based Learning of Search Control Knowledge
scheduling. The researcher minton had reported experiments in
In PROLOG-EBG approach, practical applicability 3problem domains that enhance the efficiency of problem solv
can be controlled based on the requirement. This methods opt ing. Moreover, the performance of learned rules is compared to
for domain theory which is complete and correct. This type of three problem domains. The total number of extensions to EBL
. learning focused on speeding of complicated search programs. enhance the effectiveness of the learning control knowledge.
As it is one of the essential class of learning problems. How
ever, several attempts that are applied on EBL had addressed Another example of general problem solving architec
about the issue of learningto control search which is also called ture that include aform of EBL is SOAR System. This system
as speedup learning. For instance, games like chess involves supports several problem solving strategies which specifies
searching of positions within a huge space with respect to all that prodigy denotes end planning strategy. Similar to prodigy,
possible moves. In this, the several optimization problems and SOAR systems understands by analyzing the situations in which
practical scheduling problems are formulated easily as large current search strategy leads to inefficiencies. SOAR System
search problems. In this, an activity is to find a possible move adopts searching through weak methods like generate and test if
towards the goal. The correct and complete domain theory for it encounters a search choice for
determining the proper action.
SOAR also uses a EBL called chunking
learning search control knowledge is obtained by combining for obtaining conditions
the definitions of legal search operators and search that is applied to problem domains.
objectives.
WARNING: Xerox/Photocopying of this book is a CRIMINAL act. Anyone
found guilty is LIABLE to face LEGAL proceedings.
learns a variable set of hom clauses and EBNN learns a fixedset This issue is avosded in EBNN as the fxed size tarzgt metwork
054sI needs constant time for classifying other new instances. Moreover, the fixed size neural network has disadvantage that it camnot
most term calculates the ahle to denote complex functions but avariable size neural network can denote thecomplex functions
atives that are extracted
atíves of target network 5.2.2 Using Prior Knowledge to Augment Search Operators
the sum of squared er Q31. Explain about using prior knowledge to augment search operators with FOCL algorithm.
Answer :
target network f(x)
the hypothesis space search. In this, the search
contributes the induc Prior knowledge can be used to augment search operator and also to alter
ncontributes the search process using hypothesis space. This type of alteration are done using FOCL
ana operator which denotes the legal steps in the
and ML-SMART algorithms gets altered. The FOCL algorithm is as follows,
e case of expression
FOCL Algorith m
system. Both FOCL and FOIL employe a sequèntial covering
algo
mation that replaces This algorithmis an extension of pure inductive FOIL positive examples that are covered by new single hom clause, the
the
rithmthat understands a single horn clause which removes
example. Both systems learns about the set of first
order horn clauses inorder
procedure is then iterated over remaining training clause is created by specific search that begins with most general pos
to give coverage to the training examples. Every new hom are then generated with information
importanceof both specializations of the current hom clauses
ents by constant H sible horn clause. After this, multiple candidate process is iterated by generating other candidate specializations and choosing
among the f(r) and related to the selected training examples. Then, the obtained. The main difierence between the systems is that they lhe in the way
clause is
ning is suppressed the best tillefficient performance of a horn during search process of single horn clause. It is known that, FOCL system gener
cted properly. This where candidate specializations are generated literal to the preconditions of clause. The FOCL also uses the same approach for
adding a domain theory. The following
ates the candidate specializations by
the derivatíves ob
producing candidate specializations and also generates additional specializations depending on the
bility to be correct process in FOCL system.
figuredepicts the hypothesis space search
ENGINEERING STUDENTS
SPECTRUM ALL-IN-ONE JOURNAL FOR
ceedings.
JN
MACHINE LEARNING [JNTU-HYDERABAD)
5.18

Generated by the con


domaln theory

Cup5HasHandle

Qp Fragile
[2+4-1
Cup Has handle BottomlsFlat,
Light,
[2 +3 HasConcaib
ConcavityPoints Up [4 +2-1

Cup Bottoml
omls Flat,
ish
HasConcavity.
Concaviry PointsUp BottoisFlat,
HandleOn Top Cup
LUgh,
(O+,2-) HasConcavity,
Cup Bottonls Flat,
Ught, ConcavityPointsUp.
HasConcavity, HandleOnSide
ConcavityPointsUp. 12+.0-1
HandleOnTop [4+. 0-1

Figure: Hypothesis Space Search in F0CL


In the above figure, solid lines specify the general to specific search steps in search process by FOIL system, the dashed
lines denotes additional specialization by FOCL relying on domain theory.
The FOCL and FOlL operations are used to determine simple domain of propositional horn clauses. In target concept
"cup", the operation of FOCL is described by drawing an distinction between the two kinds of literals that appear in hypothesis
representation and domain theory. The literal is said to be operational when it is allowed to be used in output hypothesis. Gener
ally, the output hypothesis is allowed to refer 12 attributes that depicts training examples. And, the literals that are based on these
12 attributes are considered as operational literals.
FOCL expands the hypothesis hat every point of search by using two operators mentioned as follows,

(1) Create e a specialization for each operational literal by adding single literal to preconditions. This approach is also used by
FOIL inorderto produce candidate successor. This type of specialization is denoted by solid arrows.
(11) Create an logical and operationally suficient condition for target concept based on domain theory by adding the set of
literals to curent preconditions of hand then they are pruned by removing the literals that are not reqired. This type of
specialization is denoted by dashed arrows.
The procedure for second operator is as follows,
Initially, FOCL selects a domain theory clauses whose post-condition matches the target concept. If there are multiple
lauses,then it try to select asinglea clause whose post-condition has the maximum information regarding the examples of target
oncept. Consider the following clause,
Cup Stable, Liftable, Open vessel
Here, the preconditions of selected clause generally form a condition which is logically sufficient for the target concept.
very non-operational operator is replaced using clause preconditions and domain theory. If stable BottomlsFlat is used for
ibstituting the BottomlsFlat unoperational stable and this process of "unfolding" in the domain theory is continued till the condi
Dns are restated based on operational literals. If there are other alternative domain theory that generate different results, highest
formation gain is selected at each step of unfolding. The finaloperational sufficient condition can be verified by the reader with
ven dormain theory and data in example as follows,

WARNING: Xerox/Photocopying of this book is a CRIMINAL act. Anyone found guilty is LIABLE to face LEGAL proceedings.
5.22 MACHINE LEARNING [JNTU-HYDERABe

5.3.3 Using Prior Knowledge to Initialize the Hypothesis


Q36. Explain about using prior knowledge to initialize the hypothesis.
Answer :

The usage of prior knowledge is one of the most important approaches to initialize the hypothesis that lits perfcctly
in domain theory. Here, the initial hypothesis is refined as per requirement in such away that it lits in the lraining data,
method is used by Knowledge Based Artificial Neural Network ie. (KBANN) algorithm to understand the neural network
this algorithmthe initial network is initially build such that for every possible instance the classilication is allocated by tlhe nety
is similar to that allocated by the domain theory. Then, the backpropagation algorithm is employed for adjusting initial nety
weights as per requirement inorder to fit in the training examples,
The basic goal of this method is that when the domain theory is correct then the initial hypothesis can accurately cla
the training examples without revising i.c., moreover, the initial hypothesis gets refined inductively by improving it to trai
examples only when it is found for inmperfectly classifving the training examples, Whereas, in backpropagationalgorithm
weights are initialized to smallrandom values. This approach can be illustrated using KBANN algorithm.
KBANNAlgorithm
This algorithm initializes the hypothesis using domaintheories and assunes that a set of non-recursive and propositi
horm clausesspecifies adomain theory. Where, a horn clause is said to be proportional ifit contains zero variables. The KB,
algorithm is given as,
Algorithm
KBANN (Domain-theory, Training Examples)
Domain - Theory: It defines set of non-recursive and propositional horn clauses.
Training - Examples: It represents set of target function inputs and outputs.
Step-1: Initially a network which is equivalent to the domain thcory is created.
Step-2: Create anetwork input for each instance attribute.
Step-3: Next create a network unit for each horn clause in domain theory as follows,
(a) All the input units are connected to the attributes and then tested by clause antecedents.

(b) A weight wis assigned to the corresponding sigmoid input unit for each clause negated autocedent.
(c) The threshold weight w, is set to -(n-.5)w.
Here. nis the total number of non-negated clause antecedent.
Step-4: Add the additional components among the network units by
to network units at depth (i+ 1). Then a random zero
connecting cach network unít at depthi) from the input
near weights are assigned to these components.
Step-5: Refine initial network finally, apply backpropagation algorithm inorder to
fit in the training examples. maintain the ínitial weights of the netwo

In the above algorithm the input is, a set of


training exarnples and a domain theory with propositional and
horn clauses and the output will be an artificial neural non-recum
network which fits the training examples.
There aretwo stages of KBANN algorithm.
1. First is to create an artificial neural network that
fits the training examples based by the
dormain theory.
2. Second is to apply backpropagation algorithm for refining the
initial network to fit training examples.
WARNING: XeroxPhotocopying of this book is a CRIMINAL act. Anyone found
quitty is LIABLE to face LEGAL proceedings
5.22 MACHINE LEARNING [JNTU-HYDERABe

5.3.3 Using Prior Knowledge to Initialize the Hypothesis


Q36. Explain about using prior knowledge to initialize the hypothesis.
Answer :

The usage of prior knowledge is one of the most important approaches to initialize the hypothesis that lits perfcctly
in domain theory. Here, the initial hypothesis is refined as per requirement in such away that it lits in the lraining data,
method is used by Knowledge Based Artificial Neural Network ie. (KBANN) algorithm to understand the neural network
this algorithmthe initial network is initially build such that for every possible instance the classilication is allocated by tlhe nety
is similar to that allocated by the domain theory. Then, the backpropagation algorithm is employed for adjusting initial nety
weights as per requirement inorder to fit in the training examples,
The basic goal of this method is that when the domain theory is correct then the initial hypothesis can accurately cla
the training examples without revising i.c., moreover, the initial hypothesis gets refined inductively by improving it to trai
examples only when it is found for inmperfectly classifving the training examples, Whereas, in backpropagationalgorithm
weights are initialized to smallrandom values. This approach can be illustrated using KBANN algorithm.
KBANNAlgorithm
This algorithm initializes the hypothesis using domaintheories and assunes that a set of non-recursive and propositi
horm clausesspecifies adomain theory. Where, a horn clause is said to be proportional ifit contains zero variables. The KB,
algorithm is given as,
Algorithm
KBANN (Domain-theory, Training Examples)
Domain - Theory: It defines set of non-recursive and propositional horn clauses.
Training - Examples: It represents set of target function inputs and outputs.
Step-1: Initially a network which is equivalent to the domain thcory is created.
Step-2: Create anetwork input for each instance attribute.
Step-3: Next create a network unit for each horn clause in domain theory as follows,
(a) All the input units are connected to the attributes and then tested by clause antecedents.

(b) A weight wis assigned to the corresponding sigmoid input unit for each clause negated autocedent.
(c) The threshold weight w, is set to -(n-.5)w.
Here. nis the total number of non-negated clause antecedent.
Step-4: Add the additional components among the network units by
to network units at depth (i+ 1). Then a random zero
connecting cach network unít at depthi) from the input
near weights are assigned to these components.
Step-5: Refine initial network finally, apply backpropagation algorithm inorder to
fit in the training examples. maintain the ínitial weights of the netwo

In the above algorithm the input is, a set of


training exarnples and a domain theory with propositional and
horn clauses and the output will be an artificial neural non-recum
network which fits the training examples.
There aretwo stages of KBANN algorithm.
1. First is to create an artificial neural network that
fits the training examples based by the
dormain theory.
2. Second is to apply backpropagation algorithm for refining the
initial network to fit training examples.
WARNING: XeroxPhotocopying of this book is a CRIMINAL act. Anyone found
quitty is LIABLE to face LEGAL proceedings
UNIT-5 Analytical
bat EBLBoth prodigy andLearning. Combining lnductive and
can SOAR systems are used to
search control inbe successfully applied analytical Learning 5.13
to gain demonstrate Consider another learning task for developing an
various
search programs uses problems knowledge on institu
ales numerical domains. Howvever, tion for the pros of
various during learning. Theproducing training derivatives, training values
aised obtained
while
by
EBL. This is dueevaluation
to the functions instead of
scenario is shown in the following figure,
in applying
various cases, EBL to leaming practical issues that are
the number of search
(a)

ssnderstood is very vast.


ngy huge cost at The systemscontrol rules control.
that needInitially
to be
Purrent search state. improve
every step to match the is to
set of rulessearch and f(x,) +
for any of This issue is not gains the f(x,)+
ules. The system that denotes its specific to EBL but occurs
the
f(x,) +
and cannot algorithms for knowledge by
be
completelymatching
5.2 ANALYTICAL LEARNING-2 eliminated
rules can alleviate enhancing
this issue

5.2.1 Using Prior


Searchin Objective
026. Explain
Knowledge to Alter the (b)

to alter the detail about using prior


Answer :
search objective. knowledge
The process of using prior knowledge to
engine begins the gradient
which fits the domain descent search throughalter the searchh
distributed as per theory perfectly. hypothesis
Then, the hypothesis
the requirement. This type of is
At torequirements
the training
that, inorder to
maximize thedistribution isdone (c) 4
knowledge error examples. Moreover, by requirementsprior
can be minimized incorporating
that
such that networkcriterion
must fit into
by gradient
descent
However, it is known that the domain prior
theory and training data.
the form of known
derivatives of atargetknowledge considered in
Drior knowledge is denoted function few
in this form. In training a types of
work, fevw target functions are neural net
knowledge specified for expressing the prior
such that the character
identifying is
of rotations and small
called
transactions ofFtthe image. Anindependent
algorithm
Figure: Fitting Derivatives and Values Using
ft bothTANGENTPROP
is used to train the
training derivatives and values.
neural network to
Algorithm
TANGENTPROP
TANGENTPROP Algorithm It can be inferred from
above figure (a) that, the target
For answer refer Unit-V, Q27. function f depending on examples (x, f(x,)), (X,, f(x,)) and
D27. Discribe the (x,, f(x,))) with these the algorithm
TANGENTPROP
aneural network algorithm to train
to fit both training values and
BACKPROPAGATON
expected to hypothesize function shows in fig(b). is
The fig(c)
specifies the effect of issuing training slopes or
training derivates. every training example as an additional data. The derivatives for
nswer: Dec.-19, 11
have a chance to generalize properly from the learner will
data by fitting training derivatives and sparse training
The TANGENTPROP algorithm holds the domain values. The main impact
is tooverride the syntactic inductive bias of
aowledge which is represented as derivates of the target func
PROPAGATION that is used for interpolation algorithm BACK
between the
n depending on the input transformation. For example, con
er a learning task with target function fand instance space x. points and replacing it by the explicit input information. Thus,
the hypothesis h shown in fig(c) provides an accurate estimate
Sassumed that every training example has a pair of (x,, f(x)) of target function f. The derivations
ich defines instance x, and training value f(x). This algorithm which are simple are only
are considered by the target concept. This
vides several training derivatives of the target function. algorithm accepts
huge information of input x. Consider an example of learning
en an instance x is defined by single real value then every task for recognizing the handwritten characters. Assume that x
ing example might be in the form of x . f•n)) a).. denotes an image with asingle handwritten character and task
classify the character. For expressing the prior knowledge a
f )
1e al) represents the derivative of target function f transformation s(, x) is defined. It rotates the image x by c
degrees. For every training instance x,, the training derivative
respect to x and is evaluated at X
x= N h can be asserted as follows.

SPECTRUM ALL-HN-ONE JOURNAL FOR ENGINEERING STUDENTS

You might also like