1 / 34

CHAPTER 12

CHAPTER 12. Lessons From an Adaptive Home University of Colorado. Scenarios. Page 273 What is wrong? Smart home vs. Smart inhabitants Forgetfulness Technologically challenged Failure to use Failure of system Get out. Smart Home Failures. Why? People satisfied with traditional home

alair
Télécharger la présentation

CHAPTER 12

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. CHAPTER 12 Lessons From an Adaptive Home University of Colorado

  2. Scenarios • Page 273 • What is wrong? • Smart home vs. Smart inhabitants • Forgetfulness • Technologically challenged • Failure to use • Failure of system • Get out

  3. Smart Home Failures Why? • People satisfied with traditional home • Obstacle to learning new interfaces is high • Only if benefit outweighs effort, cost

  4. Failure Example U.C. San Diego - Lab • Dozen light switches - "random" • Replaced with "map" of switches on wall • Too many, too complex, look up or down • People just turned on all • Interface too difficult Solution • Invisible (non-existent) interface

  5. Adaptive House • 8-year project • No interface beyond normal ones • "Intelligence" -- ability to predict • Systems implemented • Lights, air temperature, water temperature • Home programs itself by • Monitoring environment • Sensing actions • Observe occupancy & behavior • Learning to predict future states

  6. Predictions • If incorrect prediction (lights, temperature), inhabitant manually adjusts, system notes "error" & learns • Scenario -- page 277

  7. Comfort Systems • Why? Correction is trivial • Prediction on statistical behavior • Many variables • Time of day, day of week -- limited • Occupancy patterns, weather, recent departures & returns • ACHE -- Adaptive Control of Home Environments • Boulder, CO. • 75 sensors (lights, temperature, door & windows, sound level, motion, weather) • Several projects -- focus on lighting • Figure 12.6, pg. 288

  8. Lighting Control • 22 independently controlled banks of lights: • Each has 16 intensity settings • Some rooms have several banks • Trained by: • Actions taken by inhabitants • Energy consumption • "Cost" of errors

  9. Why is Lighting Hard? • Not just ON/OFF • Mood, intensity • Motion sensors not sufficient • Sit still for long time, lights go off • Roll over in bed, lights go on • Pets? • Lag time between sensing/action (700ns), due to outdated X10 protocol use • Lag is inconvenient

  10. Why is Lighting Hard? • Range of time scales • Change quickly • Remain same for hours • Two constraints-conflict • Preferences vs. Energy Conservation

  11. Optimal Control Framework • Not "supervised learning“ • Won’t adequately consider cost Energy Cost: actual cost based on setting Discomfort Cost: incurred if inhabitant changes setting Expected Average Cost: considers both factors Optimal Control Policy: map from state to decision that minimizes expected average cost

  12. Reinforcement Learning • A stochastic form of dynamic programming • Q learning (specifically) • Incremental algorithm that determines expected cost of action in a given state • Requires all states be visited to determine cost • Converges to optimal policy to minimize cost • Formula, page 279-280

  13. Temporal Credit Assignment (TCA)-- A Poor Choice -- • Start of Time Interval - a fixed value (e.g. 200 ns) • Observe current state • Make control decision • Observe (calculate) cost • May be attributed to current or any previous decision • Figure 12.3, page 280 • TCA Problem • Determine which decisions are responsible for the cost & assign credit • Shorter the time interval, more complex problem becomes

  14. TCA (continued) • Fixed-time-interval method not practical • For lighting, use event-based approach • TCA problem is eliminated • That is, intervals begin/end when an event occurs, so can be very long intervals, or short • As opposed to 18,000/hour • Using 200ns

  15. The Lighting Problem -- Lighting Zones -- Assumption: Light level in one zone will not affect inhabitant preference in other zones • Not totally true • Greatly reduces problem (1622 possibilities) ACHE House • 8 lighting control zones • largest zone -- 7 banks • smallest -- 1 bank

  16. Defined Events (pg. 281) • Zone entry; zone exit • Significant change in outdoor light level • Change in inhabitant activities • Manual change of light after 2 minutes of satisfaction • (in great room) movement from one region to another • Anticipation (prediction) of a zone entity

  17. Processing of an Event When event detected • Control decision made • Interval: time between 2 events Interval (Window) • Starts with each new event • So intervals are unitary, independent • Cost entirely determined by last event, within current interval (window)

  18. Excellent Example of Research Eliminating Long-Term Decision Consequences 3 Domain Properties (allow elimination) • Effect of decision completely undone by subsequent decision • Only true of settings are absolute not relative • Inhabitant activities are unaffected by ACHE decisions • Change light, not activity • Current sellings are irrelevant; do not need to consider for decision * Current state does not depend on earlier decisions -- TCA problem vanishes

  19. ACHE ArchitectureFig. 12.4 p. 283 • Sensors continually "sense", but nothing happens unless an event is detected • Light switches represents manual intervention • Q learning controller is trained by cost evaluator, makes decisions

  20. Q-Learning Controller • Each zone & device in a zone is treated as independent • 2 lookup tables: room empty vs. room occupied • Map state + decision = expected cost

  21. Occupied Table States • Natural light level • Number of zones changed by inhabitant in last minute • Number of zones changed by inhabitants in last 5 minutes • If great room: location in room

  22. Empty Table States • Number of entries to zone under consideration in last minute • Number of entries to zone under consideration in last 5 minutes • Relative power consumption in current state -- leave device as is or turn off

  23. Occupancy Model • Determines zones currently occupied • Motion sensed  occupied • Remains occupied until motion in adjacent zone & no motion for k seconds • Value for k • Depends on number of occupants • Conservative: 600 sec for multiple, 10 sec if single occupant • If 2 zones occupied for 10s, sets to multiple until home is empty

  24. Anticipator • Motion detectors -- sluggish (700 ms) • Inhabitants across room before light • Anticipator = a neural network • Predicts zone to be occupied within 2 seconds • Note: Use better hardware? Still must deal with noise, undependable hardware

  25. Anticipator Input Data • Average of binary motion detector in 1-, 3-, & 6-sec window (36 inputs) • Instantaneous 2-s average of binary door (20 input) • Instantaneous 1-s, 3-s average of sound level (33 input) • Current zone occupancy & durations (16 input) • Time of day (2 input: 24 hour clock) Averages provide recent history

  26. Anticipator Output • Probability zone will become occupied in next 2 seconds, given currently unoccupied • If occupied, anticipator ignored • Anticipator runs every 250 ms Neural Net • Single hidden layer • 107 inputs, 50 hidden units • 8 output units, direct I/O connections

  27. Anticipator Training • Provided by occupancy model • Inductive: • Anticipator predicts • If error, new data added • Sufficient data (200) retrain • Errors -- 2 types • Miss -- fail to predict occupancy • False Alarm -- predict but no occupancy Figure 12.5 -- pg. 287; Example

  28. ACHE Parameters Exploration Probability • 95% - take highest probability action • 5% - take other choice • e.g. next lower energy cost - gradually lowers settings until manual adjustment (increase cost) Learning Rate • For Q-table, rate at 0.3 provides balance between adaptability & stability

  29. ACHE Costs Energy: actual $0.72 kw/hour Discomfort: manual adjustment • $0.01 per device adjusted • multiple devices in zone Errors: (as above) • $0.01 per device Initialization: no inhabitant preferences

  30. Results Figure 12.7 (page 290) Training Scenario • Controller selects lowest intensity (least cost) • Inhabitant changes (discomfort cost) • After 3-4 trials, learn correct level • tends to be lowest acceptable level Occupant's Evaluation • X10 controllers slow - frustrating • Overall benefits outweigh inconveniences

  31. Future?ExtensionsNotto Pursue Home Entertainment System • Depends on undetectable state of mind • Usually explicit intention • lights & temperature are implicit • Benefit vs. cost  manual control Provide Explicit Information • e.g. calendars, preferences • This is not the minimal interface

  32. Activity Classification • Could help controller with predictions, decisions • E.G.: if "cooking dinner", certain lighting settings even if move among zones to answer phone, start washer

  33. Training the Inhabitants ACHE developer found himself choosing to go home, for example, due to the house expectations. Feedback • Tell user how much it will cost to change thermostat

  34. Conclusion Do you want a Smart House?

More Related