Smart Home Application

CRESCENT TCU Dept. of Computer Science

Smart Home Application

Intelligent TV ViewingVince Guerin


Glorified House Controller

• NSF funded research project on “Smart Home” technologies

• UTA / TCU Smart Home Project:- “Glorified House Controller (GHC), a remote control system, will be able to operate any electronic device in a home. It will also be able to change the status of different appliances, save settings of all devices for a quick change, and have the ability to learn television viewing habits.”


Smart TV Recommender Goal

• “Intelligent” program(s) will predict, according to a person’s likes and dislikes, whether it should record a television program or not.

• This will be similar to what “Amazon.com” does for books.


TV Recommender

Product must be:• Accurate• Easy to use• Able to build trust in the

recommendations delivered


Agenda


Data Needs

• Learning Algorithms

http://tvlistings2.zap2it.com

- Online TV Guide - Current online guides lack info needed for some learning methods (keywords, etc…)


TV Recommender – Two Ways to Learn

• Program reads various “keywords” inputted by the user (such as ‘comedy,’ ‘horses,’ ‘horror,’ etc..). Program then picks out television shows that contain those words in the description

• Program monitors how often the user watches certain types of shows; decides based on past viewings.


AI Project – Keyword Matching


Other types of Keywords


Scenario 1

1 – User watches at least 2 hours of TV per night.

2 – Program monitors viewing and gathers keywords and names of programs most watched.

3 – After 3 weeks of viewing, user takes vacation and turns on program to record shows most watched.

4 – User returns from vacation and views recorded shows.


Scenario 2

1 – User watches at least 2 hours of TV per night.

2 – Program monitors viewing and gathers keywords and names of programs most watched.

3 – After 3 weeks of viewing, user takes vacation and turns on program to record shows most watched, as well as programs he/she might enjoy.

4 – User returns from vacation and views recorded shows.


Scenario 3

• User manually inputs keywords, channels, and television programs to guide the system as to which programs to record.

• User lets system run all day.• According to specifications, system

records appropriate programs.• User returns and watches pre-selected

viewing material.


Java Expert System Shell (JESS)

• What is JESS?- Java rule based expert system from Sandia National Laboratories (http://herzberg.ca.sandia.gov/jess)- Stores rules and facts- Ability to reason given rules, and assert actions based on facts- Similar to a relational database


JESS Cont…

• Why is JESS important to Smart Home Technologies?– Continuously changing data – Unambiguous language to represent

rules – References and method invocations of

Java object – Seamless interaction between rule

evaluation and framework


JESS in Action – KM Project

• 2 Phases – Rank & Record• Rank

– rules with decreasing salience fire, with each rule looking for something different each time

– The highest salience rules fire first, and they assign the highest rankings based on the criteria for which they check

– When none of those rules can fire any more, then the phase change rule fires and changes the phase from ranking shows to recording shows


JESS in Action cont…

• Record Phase– Iterates through the rankings in the

same fashion (using decreasing salience)

– It will keep recording shows with decreasing rank, so long as there isn't a time conflict, and there is enough tape left.


Phillips USA solution

Kaushal Kurapati’s ideas for capturing preferences:

• Using “stereotypes” from which the user can choose (clusters of TV shows that are similar to one another)

• Create a user “Profile” according to the stereotypes


Phillips USA solution cont…

• Calculating the “distance” between networks/shows

Example:

Calculating the “distance” between FOX and NBC


Cont…

Computing Distances:


Cont…

Deriving Stereotypes from Clustering Algorithm:


Phillips Solution conclusion

• Tested in Manor, New York area on 10 users

• Users contributed TV viewing histories for periods ranging from 5 months to 2 years

• Average initial “error rate” was around 40% (best was 30%, worst was 62.6%)• Need to improve “out-of-box” error rates• Future work – deeper pool of user data


Summary

• 2 solutions presented somewhat solve the problem, but for the final product, we need more.

• Java, perhaps, as the language of choice• Implementation of keywords for online

TV guides• Overall, these ideas are a good start on

working toward a useful, functional product


References

• http://www.cs.umbc.edu/~skaush1/IASTED_2002.pdf

TV-Learning paper #1• http://www.csee.umbc.edu/~skaush1/TV02_Ea

se_of_Use_Trust_Accuracy.pdf TV-Learning paper #2• http://tvlistings2.zap2it.com/

television guide site

• http://www.captions.org/ closed captioning information site


References Cont…

• http://red.cs.tcu.edu/crescent.html#_Work_InformationCrescent Home

Smart Home Application

Technology

clusters of tv shows

user profile

television shows

hours of tv

tv viewing histories

tvlearning paper

tv recommender product

jess cont