1. Chapter 8 Prediction Algorithms for Smart Environments
MavHome - U Tx Arlington

2. Smart Environments Design & Implementation requires breadth
Integration of disciplines
Machine learning
Human machine interfaces
Decision making
Wireless network
Mobile communication
Databases
Sensor nets
Pervasive computing

3. Benefits of Automation
Convenience
Turn off coffee, warm up car
Conservation
Manage heat & cooling, lawn watering
Do actual work
-- Order groceries, vacuum carpet
See "Bob's Day" -- pg

4. Role of Prediction Goals of Smart Environment: Maximize comfort
Minimize costs
Adapt to inhabitants
To Attain:
Use tools from artificial intelligence
Prediction, automatic decision making

5. Prediction Learn about devices by observation
At certain temperature, how long to warm house
Utilization of resources, also
To cool house, turn on ceiling fan and/or close blinds
Predict inhabitant's behavior
Hardware: video, power meters, motion detectors, load sensors, device controllers, vital sign monitors
Software: prediction algorithms

6. Prediction Outcomes Determine relevant features
Make maximum use of information
Minimal prediction errors
Minimal delays (quick predictions)
Prediction Decision Making Algorithm

7. Prediction Algorithms
* NOTE: Smart environment development is not all hardware *
Prediction Task: process of forming an hypothesis representing the future value of a target variable for a given data point.
Prediction Algorithm: learns a function that maps known information collected from past/current observations to future point in time.

8. Prediction Algorithms
Based on sequential ordering of events; input to algorithm (plus maybe timing)
Historical information + current state => prediction
Given event sequence {x1, x2, x3 …. Xj}, what is event xj+1?
Approaches
Pattern matching (sequence)
Markov Decision Process
Plan recognition

9. Sequence Matching - IPAM
Just one example algorithm
Collects sequential pairs; calculates probability of transitioning from one event to next
e.g. {a, b, c, b, c, b, a, a}
(a,b) (b,c) (c,b) (b,c) (c,b), (b,a) (a,a)
p (a,b) = 1/7
p (b,c) = 2/7
p (a,c) = 0, etc.
But, probabilities change over time and are kept in a table

10. IPAM (continued) When new event xj+1 is observed
p (xj, xj+1) increases by factor of (1-α)
For some constant α
All other p (xj, z) reduced by factor α
Weights recent events more heavily
Rank events by probability and prediction p (xj+1| xj)
Sequence matching algorithms: applied to UNIX command prediction

11. Markov Decision Process (MDP)
At each step, agent perceives environment state, selects action
Probability model + possible reward
Only last few stated used
Unlike pattern matching
Hidden Markov Models (HMM)
Observable vs. hidden states
Figure 8.3, pg. 179
Hidden: current task (in chair sleeping or reading) & health (feel good or tired)
Also probabilistic

12. Plan Recognition Prediction
Given a known goal, recognize possible plans to achieve
e.g. goal: cool house
Plan 1: turn on air conditioner
Plan 2: turn on air conditioner and ceiling fans
Based on Belief Network
DAG: nodes are RV, edge indicates influence
Figure 8.4, pg. 180
Includes evidence to support plan generation

13. Other Prediction Approaches
Decision Trees
Neural Nets
Bayesian Classifiers
Nearest Neighbor Algorithm
Support Vector Machines

14. MavHome - Smart Home - UTA
Designed as an intelligent agent
Goal: comfort of inhabitant & minimize cost of running home
Predict, reason, adapt
Figure 8.5, pg intelligent agent
Figure 8.6, pg architecture
4 layers
Decision, information, communication, physical

15. MavHome - 4 Layers Decision: selects actions for agent to execute
Information: collect information, generate inferences for decision making
Communication: routes information and requests between agents
Physical: contains hardware -- appliances, network, sensors, etc.
Bottom-up Process, pg. 182

16. Identifying Events
Need to identify repetitive tasks for potential automation
Need to predict next action
MavHome: prediction solely on previous interaction with devices plus current state
Prediction to decision –
algorithm that selects action to execute

17. Algorithms Repetition modeled as stationary, stochastic process
ED (Episode Discovery): identify sequences of regular & repeatable actions that could be used to predict - thru Data Mining
Active LeZi: uses sequence matching to predict next action
MDL (Minimum Description Length) Principle: pointer to database description of patterns (compression)

18. Experiment with Algorithms
30 days of data with noise
ED discovered 6 significant episodes (pg. 184)
Use of the knowledge?
Provides understanding of nature of home
Patterns will be used in decision making
Can improve prediction accuracy

19. Active LeZi Based on Ziv - Lempel compression (LZ78)
Good compression good prediction
LZ78 enhanced to improve prediction
Calculates probability of each action occurring in a sequence & predicts one with highest
Accuracy ≈ 48%
With random choice 2%
MavHome - Figures pg. 189+

20. Conclusion
Comfort: minimize number of manual interactions with environment
Overview of Prediction