Please enable JavaScript.
Coggle requires JavaScript to display documents.
Yang: Learning from a Learning Thermostat :check: (Goals (Study the…
Yang: Learning from a
Learning Thermostat :check:
Goals
Study the experience of living with an advanced thermostat
Better understand the challenges of deploying intelligent systems in the home
Study the UX of living with a smart appliance in the wild, particularly one with machine learning
Understand real-life, long-term experience with technology in the home
Inform the design of intelligent home systems broadly (not just heating)
Method
Summary
Households that had a Nest installed
Interview - 13 ppl
Interview + Diary study - 10 ppl
Sample
23 participants, 19 households
Male / tech-savvy / affluent
Interviews
Questions
Compare usage of Nest and reg. thermostat
Overall experience with Nest
Overall understanding of Nest
Phone
1-3 interviews
Diary Study
Gathered details
Individual situations
Decision-making processes
Changes in perception and understanding over time
Reported
Daily routines
Changes made to the thermostat
Reactions
Weekly screenshots of schedule and energy history
Mechanics
Done by primary
Done online
3 weeks
Recruitment
Social networks
Public forums
Findings
Well-received overall :star:
Easy to use
Nest lighting up as they passed by
Phone control
Intuitive GUI
Increased awareness of energy use
People found Energy History useful
Nest perceived:
Less useful than expected
Less intuitive than expected
People frustrated by:
Failure to understand their intent
Nest can't distinguish routine and temp behavior
Most problems came from ppl who actively managed the temp
Some felt it was
over-eager
to adapt the schedule
Others felt it was
arrogant
, doing its own thing
System's behavior opaque
How much data is necessary for the Nest to make a sch?
How many changes will confuse the Nest?
How long does it take for the Nest to learn a new pattern?
How does the Nest react to multiple people adjusting it?
Does the learning happen all at the beginning or continually?
How does Auto-Away work with the Auto-Schedule?
Doing the schedule manually seems easier...
Is Auto-Away on while I'm at home, just working on my computer?
Will my corrections confuse the Auto-Away feature when I'm really not home?
If I delete inappropriate changes, will the Nest stop learning my behavior?
Energy savings unclear
Auto-Sch and Auto-Away shortcomings hampered savings
Users used the Nest to pursue comfort first (cranking up heat while still in bed)
Auto-Sch's learned schedule not necessarily more energy efficient
Schedule based on ppl's adjustments, so prioritizes comfort over efficiency
Auto-Away bugs led to wasted energy
$$$ Most savings came from users' engagement with monitoring
Workarounds
Modifications to the schedule were easy and doable
Limited # of adjustments after realizing Nest limitations
Actively monitored schedule and Energy History
Recommendations
Exception Flagging
Way for users to tell the smart system
when a situation is unusual
Incidental Intelligibility
Helping users understand how a system works, in bite-size pieces (incidental) not requiring separate attn
Constrained Engagement
Recognize that engagement with smart home tech is not a user's priority. Make interaction fun and easy because the more the user is engaged, the more $ they will save.
Critical Context
"Smart home" vision
28
"a home which seeks to adapt to its inhabitants and respond to their information and comfort needs"
30
becoming a reality! Digitally-enhanced home appliances coming out
Promise to
reduce manual work
operate efficiently on behalf of users
require little to no intervention
provide new types of information
Research
9
in lab
10
in lab
Design Principles
For machine learning, manageability means intelligibility and control (1, Edwards 5)
Offer advanced functionality without becoming unmanageable (Edwards, 5)
Gaps between mental models and system cause: inefficiency, confusion, dissatisfaction, and quitting (27)
Research on designing INTELLIGIBLE ui
1
12
24
Managing Energy Consumption
Research
7
eco-feedback
8
prototype
22
prototype
23
info didn't always result in behavior change
18
usability is a barrier for programmable thermostats
Commercial
30
THE NEST
First mass-market thermostat with machine learning
Promises
Promote comfort
Energy savings
Convenience
More enjoyable interaction
Features
Auto-Schedule
Generates a personalized heating / cooling schedule
Uses user controls to get to know preferences
Takes about a week to general schedule
Afterwards, adapts schedule according to temp adjustments
wall-mounted
additional smartphone & web interfaces
Auto-Away
Embedded motion sensor on wall unit
If 2 hrs with no movement, goes into this mode
Adjusts temp to user-defined "away" level
History
Released 2011
Cost $249
(reg. therm $40)
Study uses v1.0
22% of energy used @ home (6)
Related research
Existing home tech
16
20
Automation prototypes
2
14
26
:no_entry: not mainstream users
:no_entry: not adaptive tech
Smart home demos
Illustrate feasibility
10
9
3
Show lived experience
Mozer, 15
Deployed adaptive tech in his home for months
Concl: adaptive home systems need to educate occupants in case of failures
Limitations
Sample of affluent, tech-savvy males
Diary study focused on recent Nest users
Analysis
Data processing
Interviews
Transcribed
Audio-recorded
Compared
Diary entries
Nest schedule
Energy history
to find explanations for changes
Interview / diary data
Coded
Higher Aims
judgments about benefits compared to previous thermostats
changes to household routines and thermal control patterns
perceived improvements to their home's energy efficiency
Lower Aims
problems / successes with learning / sensing features
mental models of the Nest
perception of usefulness / desirability of the features
Main aim: gain insights into how to deploy intelligent systems into the home
