intelligent agents kriti puniyani – 04305012 neela sawant – 04305811 under the guidance of prof....

34
Intelligent Agents Kriti Puniyani – 04305012 Neela Sawant – 04305811 Under the Guidance of Prof. Pushpak Bhattacharyya

Upload: elvin-damian-cook

Post on 28-Dec-2015

222 views

Category:

Documents


1 download

TRANSCRIPT

Page 1: Intelligent Agents Kriti Puniyani – 04305012 Neela Sawant – 04305811 Under the Guidance of Prof. Pushpak Bhattacharyya

Intelligent Agents

Kriti Puniyani – 04305012Neela Sawant – 04305811

Under the Guidance of Prof. Pushpak Bhattacharyya

Page 2: Intelligent Agents Kriti Puniyani – 04305012 Neela Sawant – 04305811 Under the Guidance of Prof. Pushpak Bhattacharyya

Agenda

What is an agent ? Structure of Intelligent Agents Types of Agents Need for Learning Syskill and Webert – An Intelligent Agent

Page 3: Intelligent Agents Kriti Puniyani – 04305012 Neela Sawant – 04305811 Under the Guidance of Prof. Pushpak Bhattacharyya

What is an agent ?

An agent is anything that perceives its environment through sensors and acts upon that environment through actuators.

PESA concepts :Percepts, Environment, Sensors, Actuators

Page 4: Intelligent Agents Kriti Puniyani – 04305012 Neela Sawant – 04305811 Under the Guidance of Prof. Pushpak Bhattacharyya

Some Examples

robot Camera, sound recorder , etc. for sensors Robotic arms, wheels, speakers, etc. for

actuators software agent (softbot)

functions as sensors information provided as input to functions in

the form of encoded bit strings or symbols functions as actuators

results deliver the output

Page 5: Intelligent Agents Kriti Puniyani – 04305012 Neela Sawant – 04305811 Under the Guidance of Prof. Pushpak Bhattacharyya

How is an “agent” different from a traditional program?

Autonomy Personalizability Pro-activity Cooperation-interactivity

Page 6: Intelligent Agents Kriti Puniyani – 04305012 Neela Sawant – 04305811 Under the Guidance of Prof. Pushpak Bhattacharyya

Structure of Agents Agent = architecture + program Agent program: the implementation of agent's perception-action mapping Architecture: a device that can execute the agent program (e.g., general-purpose computer, specialized device, robot)

Page 7: Intelligent Agents Kriti Puniyani – 04305012 Neela Sawant – 04305811 Under the Guidance of Prof. Pushpak Bhattacharyya

Types of Agents Simple reflexive agents Model-based reflex agents (agents with internal

states) Goal Based Agents Utility Based Agents All can be turned into learning agents.

Page 8: Intelligent Agents Kriti Puniyani – 04305012 Neela Sawant – 04305811 Under the Guidance of Prof. Pushpak Bhattacharyya

Simple reflexive agentsAct on current percept, ignoring percept

history.

Page 9: Intelligent Agents Kriti Puniyani – 04305012 Neela Sawant – 04305811 Under the Guidance of Prof. Pushpak Bhattacharyya

Model-based agentsUse internal states (or models) to deal with

the world that is only partially observable.

Page 10: Intelligent Agents Kriti Puniyani – 04305012 Neela Sawant – 04305811 Under the Guidance of Prof. Pushpak Bhattacharyya

Goal-based agents

Anticipating future … involving planning and search. More flexible behavior.

Page 11: Intelligent Agents Kriti Puniyani – 04305012 Neela Sawant – 04305811 Under the Guidance of Prof. Pushpak Bhattacharyya

Utility Based AgentsImprove upon the goal-based agents by

having high-quality behavior in most environment.

Page 12: Intelligent Agents Kriti Puniyani – 04305012 Neela Sawant – 04305811 Under the Guidance of Prof. Pushpak Bhattacharyya

Learning Agents

Four main components: learning element, performance element, critic and problem generator.

Page 13: Intelligent Agents Kriti Puniyani – 04305012 Neela Sawant – 04305811 Under the Guidance of Prof. Pushpak Bhattacharyya

Syskill & Webert

An Intelligent Agent That Identifies Interesting Websites

Page 14: Intelligent Agents Kriti Puniyani – 04305012 Neela Sawant – 04305811 Under the Guidance of Prof. Pushpak Bhattacharyya

Motivation Information Overload on the Internet. It has been estimated that only 26% of Lycos

results are actually interesting. Syskill & Webert is an intelligent agent that

learns the profile of the user, and uses it to suggest interesting pages to the user.

Page 15: Intelligent Agents Kriti Puniyani – 04305012 Neela Sawant – 04305811 Under the Guidance of Prof. Pushpak Bhattacharyya

Functionality

Learns a profile for every topic of interest of the user.

Uses an index page or Lycos to find other web-pages on the topic of interest, and rate them according to user profile.

Allows the user to give an initial set of good or bad web-pages.

Page 16: Intelligent Agents Kriti Puniyani – 04305012 Neela Sawant – 04305811 Under the Guidance of Prof. Pushpak Bhattacharyya

Rating Of Pages

Page 17: Intelligent Agents Kriti Puniyani – 04305012 Neela Sawant – 04305811 Under the Guidance of Prof. Pushpak Bhattacharyya

Suggested Web Sites

Page 18: Intelligent Agents Kriti Puniyani – 04305012 Neela Sawant – 04305811 Under the Guidance of Prof. Pushpak Bhattacharyya

Overall Block Diagram Of the System

Page 19: Intelligent Agents Kriti Puniyani – 04305012 Neela Sawant – 04305811 Under the Guidance of Prof. Pushpak Bhattacharyya

How is the classification done? Naive Bayes Classifier TF-IDF & Clustering Algorithms Neural Networks Self Organising Maps

Page 20: Intelligent Agents Kriti Puniyani – 04305012 Neela Sawant – 04305811 Under the Guidance of Prof. Pushpak Bhattacharyya

Is Bayes Theorem Applicable?

Bayes Theorem : P(Y|X) = P(X|Y) * P(Y) / P(X)

Y= Website Intersting (+) or not (-) X = Vector of words Question : Is it simpler to calculate

P(+|X) or P(X|+) Calculate P(+|X) - data sparsity

problem

Page 21: Intelligent Agents Kriti Puniyani – 04305012 Neela Sawant – 04305811 Under the Guidance of Prof. Pushpak Bhattacharyya

Naïve Bayes Classification Web page to be classified is converted into a

Boolean feature vector – X. P(+ | X) = P(X | +). P(+) (P(Denom)

ignored)= P(X1, X2, X3, ....Xk | +). P(+)=P(X1|+). P(X2|+). P(X3|

+). .....P(Xk|+).P(+)(Naive Bayes Assumption)

Similarly, P(- | X) = P(X1|-). P(X2|-). P(X3|-). .....P(Xk|-).

P(-) If P(+ | X) > P(- | X) significantly, then X is an

interesting web page.

Page 22: Intelligent Agents Kriti Puniyani – 04305012 Neela Sawant – 04305811 Under the Guidance of Prof. Pushpak Bhattacharyya

Construction of Feature Vectors

User rates html pages as either hot or cold, which gives us the set of positive and negative examples.

Each example is analysed to find set of k most informative words, and a Boolean feature vector is constructed, in which 1 indicates the presence of word i in the document and 0 indicates absence of the word.

Words from stop list are not considered.

Page 23: Intelligent Agents Kriti Puniyani – 04305012 Neela Sawant – 04305811 Under the Guidance of Prof. Pushpak Bhattacharyya

Example Interesting Web Pages:

<title>Supervised Learning requires a set of positive and negative examples.</title>

<body>Rote learning is not a very efficient learning technique.

Uninteresting Web Pages: Does there exist an efficient method of

learning? 4 most informative words:

learning, supervised, rote, technique

Page 24: Intelligent Agents Kriti Puniyani – 04305012 Neela Sawant – 04305811 Under the Guidance of Prof. Pushpak Bhattacharyya

Probability Calculation P(learning | +)=3/22 P(supervised |

+)=1/22P(rote | +)=1/22 P(technique | +)=1/22

P(learning | -)=1 / 8 P(supervised | -)=0P(rote | -)=0 P(technique | -)=0

P(+) = 2 / 3 P(-) = 1 / 3 Due to presence of 0 probabilities, we need

to smoothen the data.

Page 25: Intelligent Agents Kriti Puniyani – 04305012 Neela Sawant – 04305811 Under the Guidance of Prof. Pushpak Bhattacharyya

Smoothing Add 1 to the Numerator & k to the

Denominator for every conditional probability calculated.

P(learning | +)=4/26 P(supervised | +)=2/26P(rote | +)=2/26 P(technique | +)=2/26

P(learning | -)=2/12 P(supervised | -)=1/12P(rote | -)=1/12 P(technique | -)=1/12

P(+) = 2/3 P(-) = 1/3

Page 26: Intelligent Agents Kriti Puniyani – 04305012 Neela Sawant – 04305811 Under the Guidance of Prof. Pushpak Bhattacharyya

Classification

Incoming Web Site:<u>I Like Machine Learning

Techniques</u> X = < l, s, r, t> X = <1, 0, 0, 1> P(+ | X) = 4/26 * 2/26 * 2/3 =

0.0079 P(- | X) = 2/12 * 1/12 * 1/3 =

0.0046 P(+ | X) > P(- | X) => Interesting

Web Site

Page 27: Intelligent Agents Kriti Puniyani – 04305012 Neela Sawant – 04305811 Under the Guidance of Prof. Pushpak Bhattacharyya

Till Now... We discussed how the user rates sites

as hot or cold. Analysed them, & extracted most

informative features Got other html pages from the net,

classified them, and presented them to the user

Issues: Where is the learning of the profile? Can we make this work without the

initial set of pages?

Page 28: Intelligent Agents Kriti Puniyani – 04305012 Neela Sawant – 04305811 Under the Guidance of Prof. Pushpak Bhattacharyya

Learning of Profile

User can also rate pages suggested by the agent, and the agent then re-calculates the k-most informative words and their probabilities – thus learning takes place.

k-most informative words & their probabilities are stored as user profile for the particular topic.

User can specify keywords (and optionally probabilities ) instead of rating pages as initial input.

Page 29: Intelligent Agents Kriti Puniyani – 04305012 Neela Sawant – 04305811 Under the Guidance of Prof. Pushpak Bhattacharyya

Advantages & Disadvantages Simple, efficient, good performance Even if only 20 pages is required as

training data, expecting a user to rate 20 web-pages before he can see results is not a very good idea.

Every html page has to be prefetched before it can be rated.

Choice of k-most informative words determines the effectiveness of the results observed.

Page 30: Intelligent Agents Kriti Puniyani – 04305012 Neela Sawant – 04305811 Under the Guidance of Prof. Pushpak Bhattacharyya

Current Status & Future Work User does not need to specify the

topic of interest – the agent is able to learn new interest topics by keeping track of web sites being accessed, and even scanning mail.

It is also able to collaborate with agents of other users to be able to recommend sites, movies etc. - although not very efficiently as of now.

Current work going on is to maintain a common profile across mutiple agents of the same user.

Page 31: Intelligent Agents Kriti Puniyani – 04305012 Neela Sawant – 04305811 Under the Guidance of Prof. Pushpak Bhattacharyya

Conclusion

Agents encompass almost all fields of AI from knowledge bases to learning, and are one of the most widely researched areas in core AI today.

Syskill and Webert is the “classic personal agent”, which though not practically very efficient, brings out the basic issues in agent design, in a comprehensible and logical manner.

Page 32: Intelligent Agents Kriti Puniyani – 04305012 Neela Sawant – 04305811 Under the Guidance of Prof. Pushpak Bhattacharyya

References

Russel, S. J., and Norvig P. (1995). Artificial Intelligence. Prentice Hall, New Jersey.

http://more.btexact.com/projects/agents.htm

Syskill & Webert : Identifying Interesting Web sites – Michael Pazzani, Jack Muramatsu, & Daniel Billsus (1998)

Living with Agents : Jaron Collis, Stuart Soltysiak, Divine Ndumu and Nader Azarmi - BT Technology Journal, 18(1), pp. 66-7, 2000.

Agents of Change : Patrick Thibodeau - SEPTEMBER 06, 2004 (COMPUTERWORLD)

Knowing Me, Knowing You – Practical Issues in the personalistaion of Agent Technology : Stuart Soltysiak and Barry Crabtree(2000)

Page 33: Intelligent Agents Kriti Puniyani – 04305012 Neela Sawant – 04305811 Under the Guidance of Prof. Pushpak Bhattacharyya

Thank You

Page 34: Intelligent Agents Kriti Puniyani – 04305012 Neela Sawant – 04305811 Under the Guidance of Prof. Pushpak Bhattacharyya

k-most Informative Words Words that appear frequently in the hot

list, and not in the cold list, or vice versa.

Expected GainE(W, S)=I(S) - [P(W= present)* I(S|

W=present)+ P(W = absent)* I | (S | W =

absent) Information Content

I(S) = - P(Sc) * log P(Sc)c { hot, cold }