UNIT V

APPLICATIONS

Applications of AI
Artificial intelligence is used in a variety of ways in today's society. It is
becoming increasingly important in today's world because it can efficiently
handle complicated problems in a variety of areas, including healthcare,
entertainment, banking, and education. Our daily lives are becoming more
comfortable and efficient as a result of artificial intelligence.
The following are some of the areas where Artificial Intelligence is used:

1. AI (Astronomy)
Artificial Intelligence (AI) can be extremely helpful in resolving complicated
challenges in the universe. AI technology can assist in gaining a better
understanding of the cosmos, including how it operates, its origin, and so on.
2. AI (Healthcare)
In the previous five to ten years, AI has become more beneficial to the
healthcare business and is expected to have a big impact.
AI is being used in the healthcare industry to make better and faster diagnoses
than humans. AI can assist doctors with diagnosis and can alert doctors when a
patient's condition is deteriorating so that medical assistance can be provided
before the patient is admitted to the hospital.
3. AI (Gaming)
AI can be employed in video games. AI machines can play strategic games
like chess, in which the system must consider a vast number of different
options.
4. AI (Finance)
The banking and AI businesses are the ideal complements to each other.
Automation, chatbots, adaptive intelligence, algorithm trading, and machine
learning are all being used in financial activities.
5. AI (Data Security)
Data security is critical for every business, and cyber-attacks are on the rise in
the digital age. AI can help you keep your data safe and secure. Some
examples are the AEG bot and the AI2 Platform, which are used to better
determine software bugs and cyber-attacks.
6. AI (Social Media)
Facebook, Twitter, and Snapchat, for example, have billions of user accounts
that must be kept and handled in a very efficient manner. AI has the ability to
organise and manage large volumes of data. AI can go through a large amount
of data to find the most recent trends, hashtags, and user requirements.
7. AI (Travel & Transport)
For the travel industry, AI is becoming increasingly important. AI is capable
of doing a variety of travel-related tasks, including making travel
arrangements and recommending hotels, flights, and the best routes to
customers. The travel industry is utilising AI-powered chatbots that can
engage with clients in a human-like manner to provide better and faster
service.
8. AI (Automotive Industry)
Some automotive companies are utilising artificial intelligence to provide a
virtual assistant to their users in order to improve performance. Tesla, for
example, has released TeslaBot, an intelligent virtual assistant.
Various industries are presently working on self-driving automobiles that will
make your ride safer and more secure.
9. AI (Robotics)
In Robotics, Artificial Intelligence plays a significant role. Typically,
conventional robots are programmed to execute a repetitive task; but, using
AI, we may construct intelligent robots that can perform tasks based on their
own experiences rather than being pre-programmed.
Humanoid Robots are the best instances of AI in robotics; recently, the
intelligent Humanoid Robots Erica and Sophia were built, and they can
converse and behave like people.
10. AI (Entertainment)
We already use AI-based applications in our daily lives with entertainment
providers like Netflix and Amazon. These services display software or show
recommendations using machine learning/artificial intelligence (ML/AI)
algorithms.
11. AI (Agriculture)
Agriculture is a field that necessitates a variety of resources, including effort,
money, and time, in order to get the greatest results. Agriculture is becoming
more computerised these days, and AI is becoming more prevalent in this
industry. AI is being used in agriculture in the form of agriculture robotics,
solid and crop monitoring, and predictive analysis. AI in agriculture has the
potential to be extremely beneficial to farmers.
12. AI (E-commerce)
 AI is giving the e-commerce industry a competitive advantage, and it is
becoming increasingly demanded in the market. Shoppers can use AI to find
related products in their preferred size, colour, or brand.
13. AI (Education)
Grading can be automated with AI, giving the instructor more time to educate.
As a teaching assistant, an AI chatbot can communicate with students.
In the future, AI could serve as a personal virtual tutor for pupils, available at
any time and from any location.
Machine Learning
"Machine learning" is defined by Simon. "Learning signifies adaptive
modifications in the system that enable the system to perform the same task or
tasks selected from the same population more successfully the next time".

Decision Tree Example

Optimizations
 ACO
 Swarm intelligence
 Genetic Algorithm
Impact Applications
1. IBM built Deep Blue, a chess-playing computer.
2. IBM's DeepQA research generated Watson, an artificially intelligent
computer system capable of answering questions presented in natural
language.
3. Deep learning is a class of machine learning algorithms that aims to learn
layered representations of inputs, such as neural networks.

LANGUAGE MODELS
Language can be defined as a set of strings; “print(2+2)” is a legal program in
the language Python, where “2) + (2 print” is not. Since the are an infinite
number of legal programs, they cannot be enumerated; instead they are
specified by a set of rules called a grammar. Formal languages also have rules
that defined the meaning semantics of a program; for example, the rules say
that the “meaning” of “2 + 2” is 4, and the meaning of “1/0” is that an error is
signated.
1. Natural languages, such as English or Spanish, cannot be described as a set
of predetermined sentences. For example, while everyone agrees that "not
being invited is sad" is an English sentence, opinions differ on the
linguistically correctness of "to be not invited is stated." As a result, rather of
defining a natural language model as a definitive collection, it is more fruitful
to characterise it as a probability distribution over sentences. Instead of asking
whether a string of words is a member of the set defining the language, we
question P(S = word) - what is the chance that a random sentence would
contain words. Natural languages are unclear as well. "He saw her duck" can
refer to either a waterfowl belonging to her or a movement she made to avoid
anything. As a result, rather than a single meaning for a sentence, we must
speak of a probability distribution across all alternative interpretations.
2. Finally, natural language is challenging to deal with due to its huge size and
rapid change. As a result, our language models are only a rough
approximation. We begin with the simplest possible approximation and work
our way up.
Let’s begin with the task of computing P(w|H) — probability of word ‘w’,
given some history ‘H’.
Suppose the ‘H’ is ‘its water is so transparent that’, and we want to know the
probability of next word ‘the’: P(the|its water is so transparent that).
One way to estimate this probability — relative frequency counts. Take a large
corpus, count the number of time ‘its water is so transparent that’ and also
count the number of times it has been followed by ‘the’.

Counting Conditional Probability with Relative Frequency Counts

While this method of estimating probabilities straight from counts works well
in many circumstances, it turns out that the web isn't large enough to provide us
with accurate predictions in the vast majority of cases. Why? Because language
is dynamic, and new sentences are introduced on a daily basis that we will
never be able to count.
As a result, we need to incorporate more sophisticated methods for calculating
the likelihood of word w given history H.
To represent the probability of a particular random variable Xi taking on the
value “the”, or P(Xi = “the”), we will use the simplification P(the). We’ll
represent a sequence of N words either as w1 . . . wn or wn (so the
expression wn−1 means the string w1,w2,…,wn−1). For the joint probability of
each word in a sequence having a particular value P(X = w1,Y = w2,Z = w3,
…,W = wn) we’ll use P(w1,w2,…,wn).
How to compute probability of entire sequence? — Using chain rule of
probability
Using Chain rule of probability to calculate joint probability distribution

The chain rule demonstrates the relationship between computing the

conditional probability of a word given prior words and determining the joint
probability of a sequence. However, there is a catch --- we don't know how to
calculate the actual likelihood of a term given a long string of words.
The bigram model uses only the conditional probability of the last preceding
word to approximate the probability of a word given all previous words:
Approximation used in Bigram Model
Markov models are a type of probabilistic model that assumes we can
anticipate the likelihood of a future unit without having to look too far back in
time.
The bigram (which looks back one word) can be generalised to the trigram
(which looks back two words) and therefore to the n-gram (which looks back n
- 1 words).
We can compute the probability of a whole word sequence by substituting the
bigram assumption for the probability of an individual word.

Bigram Model
What method do we use to calculate bi-gram (or n-gram) probabilities? —
Maximum likelihood estimation, or MLE, is a simple method for estimating
probabilities. By taking counts from a corpus and normalising them so that they
fall between 0 and 1, we may acquire the MLE estimate for the parameters of
an n-gram model.
To compute a specific bigram probability of a word y given a prior word x, for
example, we'll count the bigram C(xy) and normalise by the sum of all the
bigrams that share the same first word x:

Estimation of bigram probability using MLE

Because the sum of all bigram counts that begin with a specific word wn - 1
must equal the unigram count for that word wn- 1, we can reduce this equation:

Using MLE to estimate bi-gram probability

Maximum likelihood estimation, or MLE, is an example of using relative
frequencies to estimate probabilities. What language phenomena does the bi-
gram statistic capture, despite the fact that it has been calculated?
Some of the bigram probabilities listed above encapsulate truths that we
consider to be strictly syntactic.
We employed bi-gram models for pedagogical purposes, but in practise, we use
tri-gram or 4-gram models. For computing probabilities in language modelling,
we utilise the log format — log probabilities. If estimated for a 4-gram/5-gram,
the numerical underflow may occur because the probability (<1) and while
multiplying together, the product grows smaller.
Evaluating Language Models
The easiest approach to assess a language model's performance is to embed it
in an application and track how much the application improves. Extrinsic
evaluation refers to this type of end-to-end assessment. Unfortunately, running
large NLP systems from start to finish is frequently prohibitively expensive.
Instead, we devised a score that may be used to swiftly assess prospective
language model enhancements. An intrinsic evaluation metric is one that
assesses a model's quality without regard to its application.
A test set is required for an intrinsic evaluation of a language model. The
probabilities of an n-gram model, like many other statistical models in our
field, are determined by the corpus on which it is trained, often known as the
training set or training corpus. The performance of an n-gram model on unseen
data, referred to as the test set or test corpus, can subsequently be used to assess
its quality.
If we have a corpus of text and wish to compare two distinct n-gram models,
we divide the data into training and test sets, train both models' parameters on
the training set, and then compare how well the two trained models match the
test set.
A better model is one that gives a greater probability to the test set — that is,
one that more precisely predicts the test set.
In practise, we usually divide our data into three categories: 80 percent training,
10% development, and 10% test.
Perplexity
The perplexity (abbreviated as PP) of a language model on a test set is the test
set's inverse probability, normalised by the number of words.

General Formula of Perplexity

Expanding the probability distribution using chain rule

As a result, if we use the bi-gram model to compute perplexity:

Perplexity for Bigram Model (LM)

Because of the inverse, the smaller the perplexity, the higher the conditional
probability of the word sequence. According to the language model, reducing
perplexity is comparable to increasing the test set probability.
Smoothing
What do we do with terms that are in our vocabulary (and aren't unfamiliar) yet
appear in an unknown context in a test set? We'll have to take a bit of
probability mass from the more common occurrences and give it to the events
we've never seen to prohibit a language model from assigning 0 probability to
these unseen events. Smoothing or discounting is the term for this type of
adjustment.
Laplace Smoothing
Before we normalise the bigram counts into probabilities, the simplest
approach to smooth them is to add one to all of them. All counts that were
previously zero will now have a count of one, counts of one will have a count
of two, and so on. Laplace smoothing is the name of the algorithm.
Although Laplace smoothing does not perform well enough to be employed in
recent n-gram models, it does provide an excellent introduction to many of the
concepts found in other smoothing algorithms.
The count ci normalised by the total number of word tokens N is the
unsmoothed maximum likelihood estimate of the unigram probability of the
word wi:
 Unigram probability of a word
We must additionally update the denominator to account for the extra V
observations because there are V terms in the lexicon and each one was
incremented:

Laplace Smoothing of Unigram Probability

By establishing an adjusted count c*, it is easy to define how a smoothing
technique impacts the numerator. Because we're simply modifying the
numerator and adding 1, we'll need to multiply by a normalisation factor to
determine this count.

Adjusted Count for Laplace Smoothing

After that, the corrected count can be normalised by N words, yielding a
probabilistic value.
 Formula for Laplace Smoothing on Bigram Model
Smoothing can also be viewed as discounting (reducing) some non-zero counts
to obtain the probability mass that will be allocated to the zero counts. We
might express a smoothing process in terms of a relative discount dc, the ratio
of discounted counts to original counts, rather than discounted counts c*:

Discounting factor
Add-k smoothing
One alternative to add-one smoothing is to move a bit less of the probability
mass from the seen to the unseen events. Instead of adding 1 to each count, we
add a frac- tional count k (.5? .05? .01?). This algorithm is therefore called add-
k smoothing.
Add-k Smoothing Algorithm

Add-k smoothing requires that we have a method for choosing k; this can be
done, for example, by optimizing on a devset. Although add-k is useful for
some tasks (including text classification), it turns out that it still doesn’t work
well for language modeling.
Backoff and Interpolation
If we are trying to compute P(wn|wn−2,wn−1) but we have no examples of a
particular trigram wn−2,wn−1,wn. We can instead estimate its probability by
using the bigram probability P(wn|wn−1). Similarly, if we don’t have counts to
compute P(wn|wn−1), we can look to the unigram P(wn).
Sometimes using less context is a good thing, helping to general- ize more for
contexts that the model hasn’t learned much about.
There are two ways to use this n-gram “hierarchy”.
i. If the evidence is sufficient, we use the trigram; otherwise, we use the
bigram; otherwise, we use the unigram.
We always blend the probability estimates from all the n-gram estimators in
interpolation, weighing and mixing the trigram, bigram, and unigram counts.
We merge different order n-grams by linearly interpolating all the models in
simple linear interpolation. As a result, we estimate the trigram likelihood by
combining the probabilities of the unigram, bigram, and trigram.

Linear Interpolation of Trigram Proabability

Such that the estimate’s add up to 1.
N-GRAM WORD MODEL
1. n-gram models used to words instead of characters
2. The word and character models are both subject to the same procedure.
3. The most significant difference is that the vocabulary—the set of symbols
that make up the corpus and model—is greater.
4. Most languages have just roughly 100 letters, and we sometimes create
character models that are even more restricted, such as recognising "A" and
"a" as the same symbol or treating all punctuation as the same symbol. Word
models, on the other hand, have tens of thousands, if not millions, of symbols.
The vast range is due to the ambiguity of what constitutes a word.
5. Word n-gram models must cope with terms that are not in the vocabulary.
6. Because there is always the possibility of a new word that was not observed
in the training corpus using word models, we must explicitly reflect this in our
language model.
7. You may achieve this by simply adding one new term to your vocabulary:,
which stands for the unknown word.
8. For different classes, numerous unknown-word symbols are sometimes
employed. Any string of digits, for example, might be changed with, or any
email address with.

INFORMATION RETRIEVAL
The task of retrieving materials that are relevant to a user's desire for
information is known as information retrieval. Search engines on the World
Wide Web are the most well-known instances of information retrieval
systems. When a Web user types (AI book) into a search engine, a list of
relevant pages appears. We'll look at how such systems are put together in this
part. A system for retrieving information (henceforth referred to as IR) can be
defined as follows:
1. A corpus of documents. Each system must decide what it wants to treat as a
document: a paragraph, a page, or a multipage text.
2. Queries posed in a query language. A query is a statement that expresses
what the user wants to know. The query language can be just a list of terms,
like [AI book]; or it can define a phrase of words that must be adjacent, like
["AI book"]; it can include Boolean operators, like [AI AND book]; or it can
include non-Boolean operators, like [AI NEAR book] or [AI book
site:www.aaai.org].
3. A result set: This is the subset of documents deemed relevant to the query
by the IR system. By relevant, we mean material that is likely to be useful to
the individual who asked the question for the specific information requirement
specified in the inquiry.
4. A presentation of the result set: This can be as simple as a ranked list of
document titles or as complicated as a rotating colour map of the result set
projected into a three-dimensional space and displayed in two dimensions. A
Boolean keyword model was used in the first IR systems. Each word in the
document collection is handled as a Boolean feature, which is true if the term
appears in the document and false otherwise.
Advantage
 It's easy to explain and put into practise.
Disadvantages
 There is no direction on how to organise the relevant materials for
presentation because the degree of relevance of a document is a single
bit.
 Users who are not programmers or logicians are unfamiliar with
Boolean expressions. Users find it inconvenient that in order to learn
about farming in Kansas and Nebraska, they must use the query
[farming (Kansas OR Nebraska)].
 Even for a seasoned user, formulating a good query might be difficult.
Let's say we try [information AND retrieval AND models AND
optimization] and get nothing. We could try [information OR retrieval
OR models OR optimization], but if it yields too many outcomes, it's
hard to determine where to go next.

IR SCORING FUNCTIONS
1. Most IR systems have abandoned the Boolean paradigm in favour of
models based on word count data.
2. A scoring function takes a document and a query and produces a numerical
score; the papers with the highest scores are the most relevant.
3. The score in the BM25 scoring function is a linear weighted sum of the
scores for each of the terms in the question.
4. A query term's weight is influenced by three factors:
• First, the frequency with which a query term appears in a document (also
known as TF for term frequency). For the query [farming in Kansas],
documents that mention “farming” frequently will have higher scores.
• In result set Not in
result set
Relevant 30 20
Not relevant 10 40
Second, the term's inverse document frequency, or IDF. • Because the term
"in" appears in practically every document, it has a high document frequency
and hence a low inverse document frequency, making it less essential to the
query than "farming" or "Kansas."
• Finally, the document's length. A million-word paper will almost certainly
have all of the query words, although it may not be about the query. A short
document that includes all of the terms is a far better option.
All three of them are taken into account by the BM25 function.

where |dj| is the length of document dj in words, and L is the average

document length i |di| N. We have two parameters, k and b, that can be tuned
by crossvalidation; typical values are k = 2.0 and b = 0.75. IDF (qi) is the
inverse document.in the corpus, L = Systems create an index ahead of time
that lists, for each vocabulary word, the documents that contain the word. This
is called the hit list for the word. Then when given an query, we intersect the
hit lists of the query words and only score the documents in the intersection.
IR SYSTEM EVALUATION
Consider the case when an IR system has given a result set for a single query
for which we know which documents are relevant and which are not from a
corpus of 100 documents. The following table shows the number of
documents in each category.
Precision measures the proportion of documents in the result set that are
actually relevant. In our example, the precision is 30/(30 + 10)=.75. The false
positive rate is 1 − .75=.25. 2. Recall measures the proportion of all the
relevant documents in the collection that are in the result set. In our example,
recall is 30/(30 + 20)=.60. The false negative rate is 1
−.60=.40. 3. Recall is difficult to compute in a big document collection, such
as the WorldWideWeb, because there is no practical way to review every page
on the Web for relevance.
IR REFINEMENTS
1. A better model of the effect of document length on relevance is a common
refinement.
2. Although the BM25 scoring system utilises a word model that treats all
terms as fully independent, we know that some words are connected: "couch"
is related to both "couches" and "sofa." Many IR systems try to account for
correlations like these.
3. For example, if the query is [couch], it would be a shame to eliminate pages
that mention "COUCH" or "couches" but not "couch" from the result set. In
both the query and the documents, most IR systems case fold "COUCH" to
"couch," and others utilise a stemming method to reduce "couches" to the stem
form "couch."
4. Next, look for synonyms for words like "couch," such as "sofa." This, like
stemming, has the potential to improve recall but at the expense of precision.
In dictionaries, or by looking for correlations in documents or searches, you
can find synonyms and related words.
5. As a last refinement, metadata—data outside of the document's text—can be
used to improve IR. Human-supplied keywords and publication data are two
examples. Hypertext linkages between documents are an important source of
information on the Internet.
PAGERANKALGORITHM
PageRank (PR) is a Google Search algorithm that ranks web pages in search
engine results. Larry Page, one of Google's founders, was the inspiration for
Page Rank. PageRank is a metric for determining how important a website's
pages are. According to Google, PageRank calculates the importance of a
website by counting the quantity and quality of links that point to it. The basic
premise is that more important websites are more likely to gain links from
other websites.

We will see that the recursion bottom out property. We assume page A has
pages T1…Tn which point to it (i.e., are citations). The parameter d is a
damping factor which can be set between 0 and 1. We usually set d to 0.85.
There are more details about d in the next section. Also C(A) is defined as the
number of links going out of page A. The PageRank of a page A is given as
follows:

THE HITS ALGORITHM

1. Another prominent link-analysis method is the Hyperlink-Induced Topic
Search algorithm, also known as "Hubs and Authorities" or HITS.
2. There are various ways in which HITS differs from PageRank. To begin
with, it is a query-dependent metric: it ranks pages in relation to a query.
3. HITS finds a set of pages relevant to the query first. It accomplishes this by
combining query hit lists.
4. Words, followed by pages in the link neighbourhood of these pages—pages
that link to or are linked from one of the relevant set's pages.
5. To the extent that other pages in the relevant set refer to it, each page in this
set is regarded an authority on the query. When a page points to other
authoritative pages in the relevant set, it is considered a hub.
6. Similar to PageRank, we don't just want to measure the quantity of links;
we want to give high-quality hubs and authorities greater weight.
7. We iterate a procedure, similar to PageRank, that changes a page's authority
score to be the total of the hub scores of the pages that point to it, and the hub
score to be the sum of the authority scores of the pages it points to.
8. PageRank and HITS both contributed to our growing understanding of Web
information retrieval. As search engines improve their methods of collecting
finer indications of search relevance, these algorithms and their expansions are
utilised to score billions of queries every day.
IMAGE EXTRACTION
The practise of collecting information knowledge by skimming a text and
looking for occurrences of a certain class of item and links between objects is
known as information extraction. Extracting addresses from Web pages, with
database columns for street, city, state, and zip code; or extracting storms from
weather reports, with fields for temperature, wind speed, and precipitation, is a
common activity. This can be done with high accuracy in a small domain.
FINITE-STATE AUTOMATA FOR INFORMATION EXTRACTION
1. An attribute-based information extraction system assumes that the entire
text belongs to a single item, and the job is to extract attributes of that object.
2. Relational extraction systems, which deal with many objects and their
relationships, are a step up from attribute-based extraction methods.
3.FASTUS, which handles news reports about corporate mergers and
acquisitions, is an example of a relational-based extraction system.
4. A sequence of cascaded finite-state transducers can be used to create a
relational extraction system.
5. That is, the system is made up of a sequence of small, efficient finite-state
automata (FSAs), each of which takes in text as input, converts it to a new
format, and then sends it on to the next automaton. FASTUS is divided into
five stages: 1.Tokenization 2. Handling of complex words 3. Handling of basic
groups 4. Handling of complex situations 5. Merging of structures
6. Tokenization is the initial stage of FASTUS, which divides the stream of
characters into tokens (words, numbers, and punctuation). Tokenization in
English is rather straightforward; simply separating characters at white space
or punctuation suffices. Some tokenizers can handle markup languages like
HTML, SGML, and XML as well.
7. The second stage deals with more complicated phrases like "set up" and
"joint venture," as well as formal names like "Bridgestone Sports Co." A
combination of lexical elements and finite-state grammar rules is used to
identify these.
8.The third level deals with core groups, such as noun and verb groups. The
objective is to break them down into manageable chunks for later phases.
9. In the fourth level, the fundamental groupings are combined to form
complex phrases. The goal is to have finite-state rules that can be processed
rapidly and that produce unambiguous (or nearly unambiguous) output
phrases. Domain-specific events are dealt with by one type of combination
rule.
10.The tenth and last stage combines the structures created in the preceding
stage. If the next line says, "The joint venture will begin production in
January," this step will note that there are two references to a joint venture,
which should be combined into one. This is an example of the problem of
identity doubt.
PROBABILISTIC MODELS FOR INFORMATION EXTRACTION
 The hidden Markov model, or HMM, is the most basic probabilistic
model for sequences containing hidden states.
 In terms of extraction, HMMs have two significant benefits over FSAs.
 For starters, HMMs are probabilistic and thus noise-tolerant. When a
single expected character is absent in a regular expression, the regex
fails to match; with HMMs, missing characters/words are gracefully
degraded, and we get a probability indicating the degree of match, not
just a Boolean match/fail.
 Second, HMMs may be trained using data rather than costly template
engineering, making it easier to keep them up to date as text changes
over time.

Hidden Markov model for the speaker of a talk announcement

 After learning the HMMs, we may use the Viterbi method to discover
the most likely path through the HMM states and apply them to a text.
One method is to apply each attribute HMM independently; in this
scenario, you can expect the majority of the HMMs to be in the
background. When the extraction is sparse - when the number of
extracted words is modest in comparison to the length of the text - this
method is appropriate.
 The alternative is to merge all of the individual qualities into a single
large HMM, which would then search for a path that travels over
several target attributes, first locating a speaker target, then a date
 target, and so on. When we predict only one of each characteristic in a
text, several HMMs are better, and when the texts are more freeform
and rich with attributes, a single large HMM is better.
 HMMs have the advantage of providing probability numbers to aid in
decision-making. If any targets are missing, we must determine
whether or not this is a valid instance of the desired relation, or
whether the targets discovered are false positives. This decision can be
made using a machine learning system.

ONTOLOGY EXTRACTION FROM LARGE CORPORA

1. Building a vast knowledge base or ontology of facts from a corpus is
another application of extraction technology. In three ways, this is
distinct:
• For starters, it's open-ended—we're looking for information on a
variety of areas, not just one.
• Second, with a large corpus, this task is dominated by precision
rather than recall—similar to how question answering on the Web is
dominated by statistical aggregates gathered from multiple sources. 
• Third, rather than being extracted from one specific text, the results
can be statistical aggregates gathered from multiple sources.
2. Here is one of the most productive templates: NP such as NP (, NP)* (,)?
((and | or) NP)?.
3. The strong words and commas must appear in the text exactly as written,
although the parentheses are for grouping, the asterisk indicates zero or more
repetitions, and the question mark indicates that anything is optional.
4. A noun phrase is represented by the variable NP.
5. This template combines the phrases "diseases like rabies impact your dog"
and "supports network protocols like DNS," implying that rabies is a disease
and DNS is a network protocol.
6. The key words "including," "particularly," and "or other" can be used to
create similar templates. Of course, many crucial passages, such as "Rabies is
a sickness," will be missed by these templates. This is done on purpose.
7. The "NP is an NP" template does suggest a subclass relationship on
sometimes, but it frequently implies something different, such as "There is a
God" or "She is a little fatigued." We can afford to be choosy with a huge
corpus, using only high-precision templates.
8. We'll miss a lot of subcategory relationship statements, but we'll almost
certainly discover a paraphrase of the statement someplace else in the corpus
that we can utilise.
AUTOMATED TEMPLATE CONSTRUCTION
Clearly, these are examples of the author–title relationship, but the learning
system had no idea who the writers were or what their titles were. The words
in these samples were used in a Web corpus search, which yielded 199 hits.
Each match is defined as a tuple of seven strings (Author, Title, Order, Prefix,
Middle, Postfix, URL), with Order being true if the author came first and false
if the title came first, Middle being the characters between the author and title,
Prefix being the 10 characters before the match, Suffix being the 10 characters
after the match, and URL being the Web address where the match was made.
1. As a match, each template contains the same seven components.
2. The Author and Title are regexes with any characters (but beginning and
ending in letters) and a length ranging from half the minimum to twice the
maximum length of the samples.
3. Only literal strings, not regexes, are allowed in the prefix, middle, and
postfix.
4. The middle is the most straightforward to grasp: each individual middle
string in the collection of matches is a separate candidate template. The
template's Prefix is then defined as the longest common suffix of all the
prefixes in the matches, and the template's Postfix is defined as the longest
common prefix of all the postfixes in the matches for each such candidate.
5. The template is rejected if one of these has a length of zero.
6. The template's URL is defined as the longest prefix of the matches' URLs.
The sensitivity to noise is the major flaw with this method. Errors can spread
quickly if one of the initial few templates is faulty. One method to mitigate
this issue is to require many templates to verify a new example, and to require
a new template to discover numerous examples that are also discovered by
existing templates.
MACHINE READING
1. A traditional information extraction system that is targeted at a few
relationships and more like a human reader who learns from the text itself; the
field is known as machine reading because of this.
2. TEXTRUNNER is an example of a machine-reading system.
TEXTRUNNER employs cotraining to improve its performance, but it
requires a starting point.
3. Because TEXTRUNNER is domain-agnostic, it can't rely on established
noun and verb lists.

On a huge Web corpus, TEXTRUNNER obtains an accuracy of 88 percent

and a recall of 45 percent (F1 of 60 percent). From a corpus of half a billion
Web pages, TEXTRUNNER has extracted hundreds of millions of facts.

Natural Language Processing

Natural Language Processing (NLP) is an area of AI that allows machines to
interpret human language. Its purpose is to create systems that can understand
text and conduct activities like translation, spell check, and topic classification
automatically.
Natural Language Processing (NLP) is a technique that allows computers to
comprehend human language. NLP analyses the grammatical structure of
phrases and the particular meanings of words behind the scenes, then applies
algorithms to extract meaning and produce results. In other words, it
understands human language so that it can accomplish various activities
automatically.
Virtual assistants like Google Assist, Siri, and Alexa are probably the most
well-known instances of NLP in action. NLP recognises written and spoken
text such as "Hey Siri, where is the nearest gas station?" and converts it to
numbers that machines can comprehend.
Chatbots are another well-known application of NLP. They assist support
teams in resolving issues by automatically interpreting and responding to
typical language queries.
You've undoubtedly come across NLP in many other programmes that you use
on a daily basis without even realising it. When drafting an email, offering to
translate a Facebook post written in another language, or filtering undesirable
advertising emails into your junk bin, you can use text recommendations.
In a word, the goal of Natural Language Processing is to make machines
understand human language, which is complex, ambiguous, and immensely
diverse.

Vauquois Triangle : Schematic diagram for Machine Translation Systems

Difference - NLP, AI, Machine Learning
Natural language processing (NLP), artificial intelligence (AI), and machine
learning (ML) are all terms that are sometimes used interchangeably, so it's
easy to get them mixed up.
The first thing to understand is that natural language processing and machine
learning are both subsets of AI.
Artificial intelligence (AI) is a broad phrase that refers to machines that can
mimic human intelligence. Systems that simulate cognitive abilities, such as
learning from examples and solving problems, are included in AI. From self-
driving automobiles to predictive systems, this covers a wide spectrum of
applications.
Natural Language Processing (NLP) is the study of how computers
comprehend and translate human speech. Machines can understand written or
spoken material and execute tasks such as translation, keyword extraction,
topic classification, and more using natural language processing (NLP).
However, machine learning will be required to automate these procedures and
provide reliable results. Machine learning is the process of teaching machines
how to learn and develop without being explicitly programmed through the
use of algorithms.
NLP is used by AI-powered chatbots to interpret what users say and what they
mean to accomplish, while machine learning is used by AI-powered chatbots
to automatically offer more correct responses by learning from previous
interactions.
NLP Techniques
Syntactic and semantic analysis are two techniques used in Natural Language
Processing (NLP) to assist computers interpret text.
Syntactic Analysis
Syntactic analysis, often known as parsing, examines text using basic
grammatical rules to determine sentence structure, word organisation, and
word relationships.
The following are some of the important sub-tasks:
Tokenization is the process of breaking down a text into smaller pieces called
tokens (which can be phrases or words) in order to make it easier to work
with.
Tokens are labelled as verb, adverb, adjective, noun, and so on using part of
speech tagging (PoS tagging). This aids in deducing a word's meaning (for
example, the word "book" has multiple meanings depending on whether it is
employed as a verb or a noun).
Lemmatization and stemming are techniques for simplifying the analysis of
inflected words by reducing them to their base form.
Stop-word elimination eliminates often used terms that have no semantic
value, such as I, they, have, like, yours, and so on.
Semantic Analysis
The goal of semantic analysis is to capture the meaning of text. It begins by
looking at the meaning of each individual word (lexical semantics). The
programme then examines the word combination and what it means in context.
The following are the primary sub-tasks of semantic analysis:
The goal of word sense disambiguation is to figure out which sense a word is
being used in a given situation.
Relationship extraction tries to figure out how entities (places, people,
organisations, and so on) in a text relate to one another.

5 Use Cases of NLP in Business

Companies are using natural language processing (NLP) techniques to better
understand how their customers perceive them across all channels of
communication, including emails, product reviews, social media posts,
surveys, and more.
AI solutions may be used to automate monotonous and time-consuming
operations, boost efficiency, and allow people to focus on more meaningful
duties in addition to understanding online interactions and how customers
communicate about firms.
Here are some of the most common business applications of NLP:
Sentiment Analysis
 Sentiment analysis is a technique for detecting emotions in text and
categorising them as positive, negative, or neutral. By pasting words
into our free sentiment analysis tool, you can observe how it works.
 Companies can acquire insight into how customers feel about brands or
products by monitoring social media posts, product reviews, or online
polls. You could, for example, examine tweets mentioning your
business in real time to spot furious consumer remarks straight away.
 Perhaps you might conduct a survey to learn how people feel about
your customer service. You may discover which areas of your
customer service earn good or negative feedback by evaluating open-
ended replies to NPS questionnaires.
Language Translation
 Over the last few years, machine translation technology has advanced
significantly, with Facebook's translations approaching superhuman
performance in 2019.
 Businesses can use translation technologies to communicate in a
variety of languages, which can help them strengthen their worldwide
communication or break into new markets.
 Translation systems can also be trained to grasp specialised language
in any area, such as finance or health. As a result, you won't have to
worry about the inaccuracies that come with generic translation
software.
Text Extraction
 You can extract pre-defined information from text using text
extraction. This programme assists you in recognising and extracting
relevant keywords and attributes (such as product codes, colours, and
 specifications) and named entities from enormous amounts of data
(like names of people, locations, company names, emails, etc).
 Text extraction can be used to find key terms in legal documents,
identify the essential words cited in customer service tickets, and
extract product specs from a paragraph of text, among other things.
Isn't it intriguing? You can use this keyword extraction tool.
Chatbots
 Chatbots are artificial intelligence systems that communicate with
humans via text or speech.
 Chatbots are increasingly being used for customer service because of
their capacity to provide 24/7 support (shortening response times),
manage many requests at once, and free up human agents from
answering repetitive questions.
 Chatbots actively learn from each contact and improve their grasp of
human intent, allowing you to trust them with routine and easy tasks. If
they come across a client question they can't answer, they'll forward it
to a human representative.
Topic Classification
 Topic classification aids in the categorization of unstructured text. It's a
terrific approach for businesses to learn from customer feedback.
 Assume you want to examine hundreds of open-ended NPS survey
responses. How many people mention your customer service in their
responses? What percentage of clients bring up the topic of "pricing"?
You'll have all your data categorised in seconds with this topic
classifier for NPS feedback.
 Topic classification can also be used to automate the process of
labelling incoming support issues and routing them to the appropriate
individual.
Closing
 The branch of AI known as Natural Language Processing (NLP)
 explores how machines interact with human language. NLP is used to
improve technologies that we use every day, such as chatbots, spell-
checkers, and language translators.
 NLP, when combined with machine learning algorithms, results in
systems that learn to do tasks on their own and improve over time.
Among other things, NLP-powered solutions can help you identify
social media posts by emotion or extract identified entities from
business correspondence.

Machine Translation

Automated translation is known as machine translation (MT). It is the process

of converting a text from one natural language (such as English) to another
using computer software (such as Spanish).

The meaning of a text in the original (source) language must be fully restored
in the target language, i.e. the translation, in order to process any translation,
whether human or automated. While it appears simple on the surface, it is
significantly more complicated. Translation is more than just a word-for-word
replacement. A translator must be able to evaluate and analyse all of the text's
aspects, as well as understand how each word influences the others. This
necessitates considerable knowledge of the source and target languages'
grammar, syntax (sentence structure), semantics (meanings), and other
aspects, as well as familiarity with each local region.
Both human and machine translation have their own set of difficulties. For
example, no two translators can create identical translations of the same
content in the same language pair, and customer satisfaction may require
numerous rounds of revisions. The major challenge, though, is determining
how machine translation can produce translations of publishable quality.
Rule-Based Machine Translation Technology
For each language pair, rule-based machine translation relies on a large
number of built-in linguistic rules and millions of bilingual dictionaries.
The software parses text and generates a transitional representation from
which the target language text can be generated. This approach necessitates
enormous sets of rules and extensive lexicons with morphological, syntactic,
and semantic information. These sophisticated rule sets are used by the
programme, which subsequently translates the source language's grammatical
structure to the target language.
Large dictionaries and complex language rules are used to create translations.
By including their terminology into the translation process, users can increase
the quality of the out-of-the-box translation. They generate custom dictionaries
that override the system's default options.
In most circumstances, there are two steps: an initial investment that improves
quality considerably at a low cost, and a continuous investment that improves
quality progressively. While rule-based MT can help firms get to and beyond
the quality threshold, the quality improvement process can be lengthy and
costly.
Statistical Machine Translation Technology
Statistical machine translation makes use of statistical translation models
whose parameters are derived through monolingual and bilingual corpora
analysis. The process of creating statistical translation models is swift, but the
technique is strongly reliant on existing multilingual corpora. For a specific
domain, a minimum of 2 million words is necessary, with considerably more
for general language. Although it is theoretically possible to meet the quality
criterion, most companies lack the appropriate multilingual corpora to create
the necessary translation models. Furthermore, statistical machine translation
uses a lot of CPU power and necessitates a lot of hardware to run translation
models at average performance levels.
Rule-Based MT vs. Statistical MT
 Rule-based MT has good out-of-domain quality and is predictable by nature.
Customization based on dictionaries ensures higher quality and adherence to
corporate terminology. However, the fluency that readers expect from a
translation may be lacking. The customising cycle required to attain the
quality criteria might be lengthy and costly in terms of expenditure. Even on
ordinary hardware, the performance is excellent.
When big and qualified corpora are available, statistical MT produces good
results. The translation is fluent, which means it reads smoothly and so
satisfies the needs of the user. The translation, on the other hand, is neither
predictable nor consistent. Good corpora training is automated and less
expensive. However, training on broad language corpora, or text outside of the
defined domain, is ineffective. Statistical machine translation also necessitates
a lot of hardware to construct and manage huge translation models.

Rule-Based MT Statistical MT

+ Consistent and predictable

– Unpredictable translation quality
quality

+ Out-of-domain translation
– Poor out-of-domain quality
quality

+ Knows grammatical rules – Does not know grammar

+ High performance and – High CPU and disk space

robustness requirements

+ Consistency between versions – Inconsistency between versions

– Lack of fluency + Good fluency

– Hard to handle exceptions to + Good for catching exceptions to
rules rules

+ Rapid and cost-effective

– High development and
development costs provided the
customization costs
required corpus exists

Machine Translation Paradigms

SPEECH RECOGNITION
Given an auditory signal, speech recognition is the challenge of detecting a
series of SPEECH words pronounced by a speaker. It has become one of the
most widely used AI applications.
1. Example: When stated quickly, the word "recognise speech" sounds almost
identical to "wreak a nice beach." Even this brief example demonstrates a few
of the challenges that make communication difficult.
2. First segmentation: written words in English have spaces between them, but
in fast speech there are no pauses in “wreck a nice” that would distinguish it as
a multiword phrase as opposed to the single word “recognize”.
3. Second, coarticulation: when speaking quickly the “s” sound at the end of
“nice” merges with the “b” sound at the beginning of “beach” yielding
something that is close to a “sp”. Another problem that does not show up in
this example is homophones – words like “to”, “too” and “two” that sound the
same but differe in meaning argmax P(word1:t | sound1:t) = argmax
P(sound1:t | word1:t) P(word1:t). word1:t word1:t Heere P (sound1:t |
sound1:t) is the acrostic model. It describes the sound of words – that “ceiling”
begins with a soft “c” and sounds the same as “sealing”. P (word1:t) is known
as the language model. It specifies the prior probability of each utterance – for
example, that “ceiling fan” is about 500 times more likely as a word sequence
than “sealing fan”.
4. Once we define the acoustic and language models, we can solve for the
most likely sequence of words using the Viterbi algorithm.
Acoustic Model
1. The magnitude of the current – which approximates the amplitude of
the sound wave – is measured by an analog-to-digital converter at
discrete intervals called sampling rate.
2. The quantization factor determines the precision of each measurement;
speech recognizers commonly keep 8 to 12 bits. A low-end system
sampling at 8 kHz with 8-bit quantization would require roughly half a
megabyte of speech every minute.
3. A phoneme is the smallest unit of sound that has a specific meaning for
language speakers. The "t" in "stick," for example, sounds similar
enough to the "t" in "tick" that English people assume them to be the
same phoneme. A vector of features summarises each frame. The
phone model is depicted in the image below.
 Translating the acoustic signal into a sequence of frames. In this
diagram each frame is described by the discretized values of three
acoustic features; a real system would have dozens of features

Building a speech recognizer

1. A voice recognition system's quality is determined by the quality of all of its
components, including the language model, word-pronunciation models,
phone models, and the signal processing techniques used to extract spectral
information from acoustic signals.
2. The most accurate systems train different models for each speaker,
incorporating variances in dialect, gender, and other factors. Because this type
of training can take several hours of contact with the speaker, the most widely
used systems do not develop speaker-specific models.
3. A system's accuracy is determined by a variety of elements. First, the signal
quality is important: a high-quality directional microphone focused at a
stationary mouth in a cushioned room will perform far better than a cheap
microphone broadcasting a signal across phone lines from a car stuck in traffic
with the radio on. The word mistake rate is below 0.5 percent when
recognising digit strings with a vocabulary of 11 words (1-9 plus "oh" and
"zero"), but it jumps to about 10% on news items with a 20,000-word
vocabulary and 20% on a corpus with a 64,000-word vocabulary. The task
matters too: when the system is trying to accomplish a specific task – book a
flight or give directions to a restaurant – the task can often be a accomplished
perfectly even with a word error rate of 109% or more.
A diagram depicting the various options for a machine translation system. At
the top, we have English text. An interlingua-based system parses English first
into a syntactic form, then into a semantic representation and an interlingua
representation, and last into a semantic, syntactic, and lexical form in French
through generation. The dashed lines are used as a shortcut in a transfer-based
system. Different systems convey data at different points; some even do so
several times.

ROBOT
1. Robots are physical agents who manipulate the physical world to fulfil
tasks.
2. They are outfitted with effectors like as legs, wheels, joints, and grippers in
order to accomplish this.
3. The sole goal of effectors is to exert physical pressures on the environment.
4. Robots also have sensors that enable them to perceive their surroundings.
5. Modern robotics uses a variety of sensors, such as cameras and lasers to
monitor the environment and gyroscopes and accelerometers to track the
robot's own movements.
6. The majority of today's robots can be classified into one of three groups.
Manipulators, sometimes known as robot arms, are physically attached to their
work environment, such as a factory assembly line or the International Space
Station.
Robot Hardware
1. Sensors serve as a perceptual link between the robot and its surroundings.
2. Passive sensors, such as cameras, are actual environmental observers,
capturing signals created by other sources in the environment.
3.Active sensors, like sonar, emit energy into the surroundings. The fact that
this radiation is reflected back to the sensor is what they rely on. Active
sensors provide more information than passive sensors, but at the cost of
higher power consumption and the risk of interference when numerous active
sensors are employed simultaneously. Sensors can be classified into three sorts
based on whether they sense the environment, the robot's location, or the
robot's internal setup, whether active or passive.
4. Range finders are sensors that measure the distance between things in the
immediate vicinity. Robots were widely equipped with sonar sensors in the
early days of robotics. Sonar sensors produce directed sound waves that are
reflected by objects, with part of the sound making it to the listener.
5. Stereo vision uses many cameras to capture the surroundings from slightly
different perspectives, then analyses the parallax in the images to compute the
range of nearby objects. Sonar and stereo vision are no longer commonly
employed in mobile ground robots due to their inaccuracy.
6. Laser beams and special 1-pixel cameras are used in other range sensors,
which can be guided using sophisticated mirror configurations or spinning
parts. Scanning lidars are the name for these sensors (short for light detection
and ranging).
7. Radar, which is generally the sensor of choice for UAVs, is another
common range sensor. Radar sensors are capable of measuring distances of
several kilometres. Tactile sensors, such as whiskers, bump panels, and touch-
sensitive skin, are on the other end of the range sensing spectrum. These
sensors use physical contact to determine range and can only be used to detect
items that are very close to the robot.
8. Location sensors are a second significant type of sensor. To detect position,
most location sensors use range sensing as a main component. Outside, the
most popular solution to the problem of location is the Global Positioning
System (GPS).
9.Proprioceptive sensors, which notify the robot of its own motion, are the
third major class. Motors are frequently equipped with shaft decoders that
count the revolution of motors in minute increments to measure the exact
configuration of a robotic joint.
10. Inertial sensors, such as gyroscopes, rely on mass resistance to velocity
change. They can aid in the reduction of uncertainty.
11.Force and torque sensors are used to measure other crucial aspects of the
robot's status. When robots handle fragile goods or objects whose exact shape
and placement are unknown, these are essential.
Robotic Perception
1. Perception is the conversion of sensor data into internal representations of
the environment by robots. Sensors are noisy, and the environment is partially
viewable, unexpected, and frequently dynamic, making perception difficult. In
other words, robots face all of the issues associated with state estimation (or
filtering)
2. Good internal representations for robots have three properties as a rule of
thumb: they contain enough information for the robot to make good decisions,
they are structured so that they can be updated efficiently, and they are natural
in the sense that internal variables correspond to natural state variables in the
physical world.

Robot perception – Dynamic Bayes Network

3. Another machine learning technique allows robots to adapt to large changes
in sensor measurements in real time.
4. Robots can adjust to such changes using adaptive perception approaches.
Self-supervised methods are those that allow robots to acquire their own
training data (complete with labels!). In this case, the robot is using machine
learning to transform a short-range sensor that is good for terrain
categorization into a sensor that can see considerably beyond.

Monte Carlo Algorithm

There may be times when no map of the environment is available. The robot
will then need to obtain a map. This is a bit of a chicken-and-egg situation: the
navigating robot will have to figure out where it is in relation to a map it
doesn't know while also generating the map while not knowing where it is.
This subject has been extensively explored under the name simultaneous
localization and mapping, abbreviated as SLAM, and it is vital for many robot
applications.
Many alternative probabilistic approaches, like the extended Kalman filter
outlined above, are used to tackle SLAM difficulties.
 Localization using extended Kalman Filter

Machine learning in robot perception

Robot perception relies heavily on machine learning. This is especially true
when the most appropriate internal representation is unknown. Unsupervised
machine learning algorithms are commonly used to map high-dimensional
sensor streams into lower-dimensional areas. Low-dimensional embedding is a
term for this method.
Another machine learning technique allows robots to adapt to large changes in
sensor measurements in real time.
Robots can adjust to such changes thanks to adaptive perception systems.
Self-supervised methods are those that allow robots to acquire their own
training data (complete with labels!). In this case, the robot is using machine
learning to transform a short-range sensor that is good for terrain
categorization into a sensor that can see considerably beyond.
PLANNING TO MOVE
1. A robot's entire reasoning boils down to deciding how to move effectors.
2. The goal of point-to-point motion is to get the robot or its end effector to a
specific spot.
3. The compliant motion problem, in which a robot moves while in physical
touch with an impediment, poses a higher challenge.
4. A robot manipulator screwing in a light bulb or a robot pushing a box over a
table top are examples of compliant motion. We start by determining an
appropriate representation for describing and solving motion-planning
problems. The configuration space—the space of robot states defined by
location, orientation, and joint angles—turns out to be a better place to operate
than the original 3D workspace.
5. In configuration space, the path planning problem is to find a path from one
configuration to the next. Throughout this book, we've seen many variations of
the pathplanning problem; the challenge introduced by robotics is that path
planning encompasses continuous spaces. T 
6. Cell breakdown and skeletonization are the two basic techniques. The
continuous path-planning problem is reduced to a discrete graph-search
problem in each case. In this section, we assume that the robot's motion is
deterministic and that its localization is precise. The assumptions in the
following sections will be relaxed.
7. The principle of skeletonization underpins the second main family of path-
planning algorithms. These algorithms simplify the planning problem by
reducing the robot's free area to a one-dimensional representation. A skeleton
of the configuration space is a lower-dimensional representation of the
configuration space.
Configuration Space
1. It has two independently moving joints. The (x, y) coordinates of the elbow
and gripper are changed as the joints are moved. (The arm is unable to move
in the z axis.) This shows that the robot's configuration may be represented
using a four-dimensional coordinate system: (xe, ye) for the elbow's placement
in relation to the environment, and (xg, yg) for the gripper's location. Clearly,
these four coordinates represent the robot's whole condition. They make up
what is referred to as workspace representation.
2. Configuration spaces come with their own set of issues. A robot's task is
frequently specified in workspace coordinates rather than configuration space
coordinates. The topic of how to convert workspace coordinates to
configuration space arises as a result.
3. For prismatic joints, these transformations are linear; for revolute joints,
they are trigonometric. Kinematics is the name for this series of coordinate
transformations.
Inverse kinematics is the inverse problem of computing the configuration of a
robot whose effector location is defined in workspace coordinates. The
configuration space can be divided into two subspaces: the space of all
possible robot configurations (often referred to as free space) and the space of
unattainable configurations (usually referred to as occupied space).
Cell Decomposition Methods
1. A uniformly spaced grid is the simplest cell decomposition.
2. The value of each free-space grid cell—that is, the cost of the shortest path
from that cell to the goal—is indicated by grayscale shading.
3. There are a variety of ways to improve cell decomposition methods in order
to relieve some of these issues. The first method permits the mixed cells to be
further subdivided—perhaps utilising cells half the original size. This can be
repeated until a path is determined that is fully made up of free cells. (Of
course, the method only works if it is possible to determine whether a given
cell is a mixed cell, which is simple only if the configuration space boundaries
are defined.)
4. Have relatively simple mathematical descriptions.) This method is complete
provided there is a bound on the smallest passageway through which a
solution must pass. One HYBRID A* algorithm that implements this is hybrid
A*.
Modified Cost Functions
1. A potential field is a function defined over state space whose value
increases as the distance to the nearest obstacle increases.
2. In the shortest-path calculation, the potential field can be used as an
additional cost term.
3. This results in an intriguing trade-off. On the one hand, the robot aims to
reduce the length of the path to the destination. On the other side, it strives to
avoid stumbling blocks by reducing the possible function.
4. The cost function can be changed in a variety of ways. It may be useful, for
example, to smooth the control settings across time.
Skeletonization methods
1. The skeletonization concept underpins the second main family of path-
planning algorithms.
2. These techniques simplify the planning problem by reducing the robot's free
area to a one-dimensional representation.
3. A skeleton of the configuration space is a lower-dimensional representation
of the configuration space.
4. It's a free-space Voronoi graph, or the collection of all points that are
equidistant from two or more obstacles. To use a Voronoi graph for path
planning, the robot must first transform its current configuration to a point on
the graph.
5. It's simple to demonstrate that a straight-line motion in configuration space
can always do this. Second, the robot moves along the Voronoi graph until it
reaches the configuration that is closest to the target. After that, the robot
moves away from the Voronoi graph and toward the destination. In
configuration space, this final phase requires straight-line mobility once more.

A repelling potential field pushes the robot away from obstacles, (b) Path
found by simultaneously minimizing path length and the potential
The Voronoi graph is the set of points equidistance to two or more
obstacles in configuration space (b) A probabilistic moodmap, composed
of 100 randomly chosen points in free space.

Robust methods
1. A robust method assumes a bounded level of uncertainty in each component
of a problem, but assigns no probabilities to values inside the allowable
interval.
2. A robust solution is one that works regardless of the actual values as long as
they stay within the assumed ranges.
3.The conformant planning technique is an extreme kind of resilient strategy.

A two-dimensional environment, velocity uncertainty cone, , butand

envelope of possible robot motions. The intended velocity is , resulting
inwith uncertainty the actual velocity could be anywhere in C a final
configuration somewhere in the motion envelope, which means we
wouldn’t know if we hit the hole or not
The first motion command and the resulting envelope of possible robot
motions. No matter what the error, we know the final configuration will
be to the left of the hole.