Fuzzy Transfer Learning for Intelligent Environments                       Jethro Shell, Student Member, IEEE             ...
Intelligent Environments   What is an Intelligent Environment?        Applications that use sensors to model and make deci...
Simple Intelligent Environment Example                         Flat A                                               Flat B...
Fuzzy Transfer Learning framework      Propose a Fuzzy Transfer Learning (FuzzyTL) framework.      Addresses the issue of ...
Fuzzy Ad Hoc Data Driven System  Fuzzy Ad Hoc Data Driven Learning                 Automated process for producing rules f...
Fuzzy Ad Hoc Data Driven System  The ad hoc system is made up of four stages:  Construction of fuzzy sets       Basis of t...
Fuzzy Ad Hoc Data Driven System  Construction of fuzzy rules       Each instance of data is used to construct a rule.     ...
Fuzzy Ad Hoc Data Driven System  Rule base reduction       The initial production of rules is equal to the size of the dat...
Fuzzy Ad Hoc Data Driven System  Defuzzification       Mapping of crisp input value to crisp output value.       Many metho...
Transfer Process      Transfer Learning utilises a humanistic style of knowledge exchange.      Crisp unlabelled data is t...
Context Adaptation                To create an effective classification system, FIS needs to be adapted.                Doma...
Application to an intelligent environment      FuzzyTL was applied to Intel Berkeley Research Laboratory[5] sensor      ne...
Application to an intelligent environment   Baseline measurement of predictive values using comparison of the same sensor....
Application to an intelligent environment                                                           1                     ...
Conclusion and Future Work   Conclusion       Constructed a novel fuzzy transfer learning framework applicable to the issu...
Bibliography   Links   Fuzzy Logic:An Introduction http://tinyurl.com/fuzzylogicvideo   Bibliography   [1] T. Bokareva, W....
Upcoming SlideShare
Loading in …5
×

Fuzzy Transfer Learning for Intelligent Environments

688 views
623 views

Published on

The dynamic nature of Intelligent Environments (IE’s) present a challenging problem when attempting to model or learn a model of such environments. By their very nature, IE’s are infused with complexity, unreliability and uncertainty due to a combination of sensor noise and the human element. As a result of this, the quantity, type and availability of data to model such environments is a major issue. Each situation is contextually different and constantly changing. To model each application, training data must be gained that is within the same feature space and has the same distribution, however this is often highly costly and time consuming. There can even be occurrences were a complete lack of target labelled data occurs. It is within these situations that our study is focussed. Within this research we propose a framework to dynamically model IE’s through the use of data sets from differing feature spaces and domains. The framework is constructed using a Fuzzy Transfer Learning process.

This presentation is taken from the winning entry at the Leicester British Computer Society Postgraduate Research Competition 2012.

Published in: Technology, Business
0 Comments
0 Likes
Statistics
Notes
  • Be the first to comment

  • Be the first to like this

No Downloads
Views
Total views
688
On SlideShare
0
From Embeds
0
Number of Embeds
1
Actions
Shares
0
Downloads
16
Comments
0
Likes
0
Embeds 0
No embeds

No notes for slide

Fuzzy Transfer Learning for Intelligent Environments

  1. 1. Fuzzy Transfer Learning for Intelligent Environments Jethro Shell, Student Member, IEEE March 28, 2012Centre for ComputationalIntelligenceDe Montfort UniversityLeicester, United KingdomEmail: JethroS@dmu.ac.uk
  2. 2. Intelligent Environments What is an Intelligent Environment? Applications that use sensors to model and make decisions about their surroundings. Constructed using a large number of varying types of sensor ranging from temperature sensors to Passive Infra-red (PIR) Sensor. Multiple applications: Environmental[3], Healthcare[6], Domestic[4], Military[1], etc. What are the issues? IE’s require models to make decisions but each environment is different:contextually, spatially and temporally. Sensor construction and application domains produce uncertain and dynamic environments. Collecting of data is expensive in terms of time, resource and analysis. Majority of learning processes require source and target data to come from the same feature and domain space.
  3. 3. Simple Intelligent Environment Example Flat A Flat B *⊕ ⊗ * * *⊕ Bedroom Bathroom ⊗ Lounge ⊗ Lounge / Kitchen * ⊗ ⊗ *⊕ * ⊗ * Bedroom Bathroom Hall Kitchen ⊗ ⊗ * ⊕ - Temperature Sensor - Occupancy Sensor ⊗ - Heating Activation Sensor A simple Ambient Intelligent Assisted Living environment. Flat A is occupied by a single individual with data (Occupancy, Temperature, and Heating Activation) monitored during March. Flat B is occupied by a couple with data (Occupancy and Temperature) monitored during October and November.
  4. 4. Fuzzy Transfer Learning framework Propose a Fuzzy Transfer Learning (FuzzyTL) framework. Addresses the issue of the lack of knowledge in target contexts by using a combination of fuzzy logic and transfer learning. Fuzzy logic allows for the incorporation of approximation and uncertainty within the environment. Transfer Learning is a humanistic style approach using knowledge from one task to assist another. Source Task Input Output Fuzzy Rules Ad Hoc Transferable Labelled Data Data Driven FIS Learning Fuzzy Sets Target Task Transferable FIS Input Output Unlabelled Predictive Processing Value Data Context Adapation
  5. 5. Fuzzy Ad Hoc Data Driven System Fuzzy Ad Hoc Data Driven Learning Automated process for producing rules from numerical labelled data. Computationally swift and low resource use. Construction based upon a Fuzzy Inference System (FIS). Characterised by short time requirement of iterative processes for rule production[2]. Outline of FIS Fuzzy Sets Data Input Defuzzification Data Output Fuzzy Rules m(t) m(o) m(h) VL L M H VH VL L M H VH VL L M H VH 1.0 1.0 1.0 0.0 t 0.0 o 0.0 h Temperature (t) Occupancy (o) Heating (h) IF Temperature is Low AND Occupancy is High THEN Heating is High
  6. 6. Fuzzy Ad Hoc Data Driven System The ad hoc system is made up of four stages: Construction of fuzzy sets Basis of the process is to construct fuzzy sets and rules from numerical data. Fuzzy sets are created by segmenting each variable domain equally into fuzzy regions by 2N + 1. Each variable can contain differing quantities of regions. Min and max values of the variable are used to define the domain. m(t) VL L M H VH 1.0 0.0 Temperature (t) tmin tmax
  7. 7. Fuzzy Ad Hoc Data Driven System Construction of fuzzy rules Each instance of data is used to construct a rule. The maximum degree of membership of each feature is used to form the rule. This is carried out for each input and output feature. m(t) VL L M H VH 1.0 0.0 t1 t2 Temperature (t) tmin tmax m(o) VL L M H VH 1.0 0.0 o1 o2 Occupancy (o) omin omax m(h) VL L M H VH 1.0 0.0 Heating (h) h1 h2 hmin hmax (t1 , o1 , h1 ) ⇒ [t1 (0.65 in VL), o1 (0.7 in M); h1 (0.55 in L)] ⇒ Rule 1 (t2 , o2 , h2 ) ⇒ [t2 (0.8 in H), o2 (1.0 in M); h2 (0.8 in H)] ⇒ Rule 2
  8. 8. Fuzzy Ad Hoc Data Driven System Rule base reduction The initial production of rules is equal to the size of the dataset. Conflicts and additional complexity are produced as a result. Predominantly this is unmanageable and requires reduction. To reduce the rule base a degree of the maximum product (d) is assigned to each rule. Maximum product of membership This can be represented as: n m d= Pij (xi ) i =1 j=1 n is the number of partitions. m is the number of rules. P representing the partitions. x equals inputs. Each rule with equal antecedent conditions are compared. Those with the greatest degree remain within the rule base.
  9. 9. Fuzzy Ad Hoc Data Driven System Defuzzification Mapping of crisp input value to crisp output value. Many methods are available to achieve defuzzification. Within this implementation a centroid method was used. Centroid defuzzification This can be represented as: K i ¯i i =1 mo i y z= K i i =1 mo i mo i denotes the degree of output using a product t-norm. y i is the centre of the fuzzy region. ¯ K is number of fuzzy rules in the rule base.
  10. 10. Transfer Process Transfer Learning utilises a humanistic style of knowledge exchange. Crisp unlabelled data is taken from target task as an input. Transferable Fuzzy Inference System (FIS) provides the basis for the knowledge base and decision making. Previously constructed fuzzy rules and fuzzy sets are used as the basis of the model. Classification output is produced using previous related knowledge.
  11. 11. Context Adaptation To create an effective classification system, FIS needs to be adapted. Domains of the fuzzy sets are adapted in order to cope with the contextual changes. Assessment is made of the whole target data set to gain new minimum and maximum of each feature. Each feature fuzzy domain is separately adapted taking into account the minimum and maximum values. The domain of each feature is compressed, expanded or shifted relating to the new values. m(t) m(t) VL L M H VH VL L M H VH 1.0 1.0 0.0 Temperature (t) 0.0 Temperature (t) old new old new tmin tmax tmin tmin tmax tmax An example of temperature fuzzy sets with adaptive shift.
  12. 12. Application to an intelligent environment FuzzyTL was applied to Intel Berkeley Research Laboratory[5] sensor network. Wireless Sensor Network (WSN) consists of temperature, humidity, light and residual power data sensors. Spatial and temporal contextual changes were tested using differing sensors on different days. Sensors 7, 9, 12 , 24 , 34 , 42 and 51 where examined across seven days using time and light to predict temperature.
  13. 13. Application to an intelligent environment Baseline measurement of predictive values using comparison of the same sensor. Sensor to Sensor Classification and Predictive NRMSE of Temperature 1 Sensor to Sensor Classification and Predictive NRMSE of Temperature 0.8 NRMSE of Comparison of Temperature Values (˚C) 0.6 0.4 0.2 0 0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40 42 44 46 48 50 Sensor Number
  14. 14. Application to an intelligent environment 1 Comparison of Spatial Temporal Value and NRMSE of Temperature 0.8 NRMSE of Comparison of Temperature Values (˚C) 0.6 0.4 0.2 0 0 10 20 30 40 50 Spatial Temporal Combined Value (ST) Spatial Temporal Combined value is represented as: ST = (xi − yi )2 + (m + 1) where x and y are points in two dimensional space in metres, m is number of days separating the feature domains.
  15. 15. Conclusion and Future Work Conclusion Constructed a novel fuzzy transfer learning framework applicable to the issues within Intelligent Environments. Addresses the issue of modelling environments in contextually different situations. Used combined fuzzy and adaptive approach that absorbs uncertainty and dynamic nature. Utilised transfer of knowledge to enhance the learning of target tasks from source tasks. Applied methodology to a real world dataset to gain results. Future Work Greater work on impact of contextual information. Further investigation into relatedness of features. The establishing of limited / reduced information learning. Increased diversity of applications, currently applying framework to task recognition using eye gaze technology.
  16. 16. Bibliography Links Fuzzy Logic:An Introduction http://tinyurl.com/fuzzylogicvideo Bibliography [1] T. Bokareva, W. Hu, S. Kanhere, B. Ristic, N. Gordon, T. Bessell, M. Rutten, and S. Jha. Wireless sensor networks for battlefield surveillance. In Proceedings of Land Warfare Conference 2006. Citeseer. [2] J. Casillas, O. Cord´n, and F. Herrera. o Improving the wang and mendel’s fuzzy rule learning method by inducing cooperation among rules. In Proceedings of the 8th Information Processing and Management of Uncertainty in Knowledge-Based Systems Conference, pages 1682–1688, 2000. [3] Y.J. Jung, Y.K. Lee, D.G. Lee, K.H. Ryu, and S. Nittel. Air Pollution Monitoring System based on Geosensor Network. 3, 2008. [4] C.D. Kidd, R. Orr, G.D. Abowd, C.G. Atkeson, I.A. Essa, B. MacIntyre, E. Mynatt, T.E. Starner, W. Newstetter, et al. The aware home: A living laboratory for ubiquitous computing research. Lecture notes in computer science, pages 191–198, 1999. [5] S Madden. Intel lab data. http://db.csail.mit.edu/labdata/labdata.html, June 2004. Published on 2nd June 2004. [6] A. Wood, G. Virone, T. Doan, Q. Cao, L. Selavo, Y. Wu, L. Fang, Z. He, S. Lin, and J. Stankovic. ALARM-NET: Wireless sensor networks for assisted-living and residential monitoring. University of Virginia Computer Science Department Technical Report, 2006.

×