peter sawyer caise
TRANSCRIPT
Modeling Pesonalized Adaptive Systems
Alistair Sutcliffe and Pete Sawyer University of Manchester and Lancaster University
Motivation
• Traditionally: – Systems developed for an idealized/homogenized user or
categories of users • Possibly with some tailoring
• Increasingly: – Systems need to be tailored to individual users and their
contexts and to adapt as users’ needs evolve – Recent trends towards self-adaptive systems reflect this
• Our aim: – A model-based framework for personal and contextual
modeling
Personal Contextual Knowledge People have …
…. information describing them
… and information held by them Some of this information changes over time and according to context.
Is a journalist
Values privacy
General stakeholder requirements
Spacial change
Business and domain evolution, expert users, ..
Culture and localisation, interaction language, …
User characteristics requirements
Individual user skill and ability, …
Physical context, social context, …
Personal goals
Attain individual goals
Location, social context
A two-layer framework
User Characteristics
• The information describing someone • Includes their skills and knowledge but also cognitive,
physical and perceptual abilities. – e.g. Truck driver has a commercial driving license (skill) but
also myopia (perceptual ability)
• May change with time – Myopia may get worse, spacial orientation skills may decline
as driver ages
• May change with context – Driving around home town vs. driving around an unfamiliar
city, reactions and spacial awareness when fresh vs. when tired
User Characteristics (contd.)
• Where useful? – Depends on application (of course), e.g.:
• Socio-technical systems involving physical action, or action in virtual environments that might lack physical cues
• Users of assistive technologies in which each user will need care tailored to their condition(s)
– Systems tailored to skills and abilities, but also adapt as these change over time e.g.:
• Skills improve • User’s condition progresses
Personal Goals
• The information held by someone, but also their attitudes and preferences
• Values are key components of personal goal attainment.
Personal Goals (contd.)
Relations with others
“Big five”
Personal Goals (cont.d)
• The information held by someone, but also their attitudes and preferences
• Values are key components of personal goal attainment.
• Values tend to be less time-variant than personal characteristics
• Values are a diverse set of properties but their usefulness lies in being able to understand their impact on personal goal attainment, e.g.: – concerns about privacy is something social network systems
have to deal with
The Framework: broad principles
• User characteristics used as a checklist to identify potential obstacles to system goal satisfaction
• In the personal goals layer, values may represent “weak obstacles”; probabilities that user behaviour will inhibit attainment of system goals
Case Study
• An Ambient Assisted Living System • Two users
– Mary – Mary’s carer
• Mary has: – Limited mobility
• Able to live at home with assistance – Mild Cognitive Impairment
• May forget things
AAL KAOS Goal Model 1
Achieve[Release Dose]
Achieve[ MedicineTaken]
Maintain[Is Healthy]
Achieve[Correct MedicineDose]
Dispenser
Mary
Top level system goal
An expectation
Human agent
System agent
A requirement
Personal Characteristics
• Mary’s key characteristic is her MCI – She might forget to take her medicine
• Leads to the obstacle Forgets to Take Medicine
• Mary’s MCI will probably get worse over time
Achieve[Release Dose]
Achieve[ MedicineTaken]
Maintain[Is Healthy]
Achieve[Correct MedicineDose]
Dispenser
Mary
Forgets to take medicine
Underdose
Overdose
Top-level obstacle
AAL KAOS Goal Model 2
Threatens
AAL KAOS Goal Model 3
Achieve[Release Dose]
Achieve[ MedicineTaken]
Maintain[Is Healthy]
Achieve[Correct MedicineDose]
Dispenser
Mary
Forgets to take medicine
Underdose
Overdose
Achieve[Prompt ToTakeMedicine]
Achieve[Remind MedicineUntaken]
Maintain[Monitor DispenserTray]
AAL
AAL
Achieve[Detect UntakenMedicine]
Mitigates
Mary wants to maintain her health
Domain assumption
So far …
• We’ve used one of Mary’s personal characteristics to guide obstacle analysis for the system goal model
• Now we will analyse Mary’s personal goals to investigate whether there are any further (“weak”) obstacles that arise from Mary’s values and that need to be mitigated.
• We use Mary’s values to help understand softgoals; system qualities
• We start by eliciting these system qualities and modeling them as softgoals
AAL KAOS Goal Model 4
Achieve[Release Dose]
Achieve[ MedicineTaken]
Forgets to take medicine
Underdose
Maintain[Is Healthy]
Achieve[Correct MedicineDose]
Overdose
Dispenser
Achieve[Prompt ToTakeMedicine]
Achieve[Remind MedicineUntaken]
Maintain[Monitor DispenserTray]
Mary
AAL
AAL
Achieve[Detect UntakenMedicine]
Mary wants to maintain her health
+
-
Avoid intervention
Minimize intrusion
Effect of Mary’s values on Minimize Intrusion quality
Minimize intrusion
frustration resentment
emotional resonse
cooperation openness
+ +
=
+ =
Mary is open and cooperative. This potentially gives hare a neutral attitude to intrusive interventions by the AAL
Effect of Mary’s values on Minimize Intrusion quality
Minimize intrusion
frustration resentment
emotional resonse
cooperation openness
+ +
=
+ =
However, Mary suffers feelings of frustration at her condition and may feel resentment that interventions are needed. This strengthens her intent to avoid interventions but may lead her to ignore or attempt to subvert reminders
This should lead us to mitigate the risks – perhaps by careful design of the reminders – e.g. to make them empathic
Uncertainty 1
• Mary’s values may be estimated in a number of ways: – Online tests – Estimates from her carer – Estimates from domain experts
• Clearly there will be significant uncertainty about the nature and strength of her values
Uncertainty 2
Minimize intrusion
frustration resentment
emotional response
cooperation openness
+ +
=
+ =
• Is this the right set of values for Mary? • Do they have the same or different relative weights? • Are the propagated values (the ‘+’, etc.) the right ones?
• These are probably best understood probabilistically • Baysian reasoning may be of help here.
Time and Context
• So far we have not considered time or context • But these might be relevant, e.g.:
– Mary’s condition might progress and reminders might have to become more frequent
– There may be contexts in which Mary more commonly forgets to take her medicine
• Monitoring (A key component of an adaptive system) could be used to discover these.
• The framework should help us identify what to monitor
Next Steps
• We are applying the framework to the EPSRC SAMS project1
– Software Architecture for Mental health Self management
• SAMS aims to encourage self-referral for people with the early signs of dementia
• SAMS will apply text and data mining techniques to look for signs of Mild Cognitive Impairment by passive monitoring of peoples’ interaction with their computer 1EPSRC grant EP/K015796/1
Conclusions
• We are interested in how to systematize the design of systems that have a focus on the individual user
• We propose a two-layer framework based on well-developed models of personal characteristics and personal goals
• We use these to help us identify strong and weak obstacles to attainment of system and user goals, using goal modeling.
• Still to do: – Validate the sets of characteristics and values – Develop means of reasoning about their effects on goal
attainment