Technology to Support Individuals with Cognitive Impairment

Martha E. PollackComputer Science &

EngineeringUniversity of Michigan

Autominder

• Model, update, and maintain the client’s plan– Including complex temporal and causal constraints

• Monitor the client’s performance– Updating the plan as execution proceeds

• Reason about what reminders to issue, and when– To most effectively ensure compliance, without sacrificing

client independence

Client Client ModelerModeler

Plan Plan ManagerManager

IntelligentIntelligentReminderReminderGeneratorGenerator

ClientPlan

Activity Info

Inferred Activity

Sensor Data

Reminders

Client Model Info

Activity Info

Preferences

Plan Updates

ClientModel

Autominder ArchitectureWhat should the client do?

Technologies: Automated Planning, Constraint-Based Temporal Reasoning

What is the client doing?

Technologies: Dynamic Bayesian InferenceIs a reminder needed?

Technologies: Iterative Refinement Planning, Reinforcement Learning

Autominder Example

Req/Opt Activity Allowed Expected Observed

R toilet use 10:45-11:05

R lunch 12:00-12:45

O TV 14:00-14:30

10:55

R toilet use13:55-14:15

REMIND 12:25

REMIND 13:55

12:28

Robot Platform

• Nomadic Technologies Scout II

w/custom-designed head

– Multiple sensors: lasers, sonars, microphone, touchscreen, camera vision, wireless ethernet

– Effectors: motion, speakers, display screen, facial expression

“Pearl”[courtesy Carnegie MellonUniv. Robotics Institute]

“Ubicomp” Platform

• Handheld or wearable device– Currently: HP iPaq

• Deployed in a “smart” environment with multiple sensors (ubiquitous computing environment)

The Plan Manager

• Maintains up-to-date record of client’s planned activities and their execution status– Eating– Hydrating – Toileting– Medicine-taking – Exercise – Social activities – Doctors’ appointments– etc.

How Does it Work?

• Models constraints on future actions– Lunch takes between 25 and 35 minutes – Take meds within one hour of finishing lunch – Watch the news at either 6pm or at 11pm

• Performs efficient constraint processing when key events occur:– New planned activity added.– Existing activity modified or deleted.– Planned activity performed.– Critical time bounds passed.

Small Example

ClientPlan

1. New Activity2. Mod/Deletion3. Activity Execution4. Passed Time Bound

PLAN MANAGER

:0 MS – LE :60“Take meds within 1 hour of lunch”

LE = 12:15“Lunch ended at 12:15”-----------------------------12:15 MS 13:15“Take meds by 1:15”

Temporal Reasoning in AI

An important task & exciting research topic, otherwise we would not be here

• Temporal Logic• Temporal Networks

– Qualitative relations: • Before, after, during, etc.

• interval algebra, point algebra

– Quantitative/metric relations: • 10 min before, during 15 min, etc.

• Simple TP (STP), Temporal CSP (TCSP), Disjunctive TP (DTP)

Temporal Network: example

Tom has class at 8:00 a.m. Today, he gets up between 7:30 and 7:40 a.m. He prepares his breakfast (10-15 min). After breakfast (5-10 min), he goes to school by car (20-30 min). Will he be on time for class?

Simple Temporal Network (STP)

• Variable: Time point for an event

• Domain: A set of real numbers (time instants)• Constraint: An edge between time points ([5, 10] 5Pb-Pa10)

• Algorithm: Floyd-Warshall, polynomial time

Example

A

B

C

[10 15]

[5 10]

?

Example

A

B

C

15

10

-10

-5

Example

A

B

C

15

10

-10

-510

Example

A

B

C

15

10

-10

-510 0

Example

A

B

C

[10 15]

[5 10]

[0 10]

Other Temporal Problems

Temporal CSP: Each edge is a disjunction of intervals

STP TCSP

Disjunctive Temporal Problem: Each constraint is a disjunction of edges

STP TCSP DTP

Search to solve the TCSP/DTP

• TCSP [Dechter] and DTP [Stergiou & Koubarakis] are NP-hard• They are solved with backtrack search• Every node in the search tree is an STP to be solved• An exponential number of STPs to be solved

CM: Client Modeler

Given what can be observed• Sensor input: client moved to kitchen • Clock time: at 7:23 a.m.• Client plan: breakfast should be eaten between 7 and 8• Model of previous actions: client has not yet eaten breakfast• Learned patterns: 82% of the time, client starts breakfast between 7:10 and 7:25• Reminder information: we issued a reminder at 7:21

Infers what has been done• Client Activity: probability that client has begun breakfast

How Does it Work?

• Models probabilistic relations among observations and actions

• Performs Bayesian update, extended to handle temporal relations• Asks for confirmation when needed!

started

breakfast

breakfastreminder issued

went tokitchen

reminder kitchen start-breakfast Y Y .95 Y N .10 N Y .8 N N .03

Intelligent Reminders

• Decides whether and when to issue reminders• Given a client’s plan and its execution status:

– Easy to generate reminders• Remind at earliest possible time of each action

– Harder to “remind well”• Maximize likelihood of appropriate performance of

ADLs and other key activities• Facilitate efficient performance• Avoid annoying client• Avoid making client overly reliant

How Does it Work?

LB D

TV

Midnight

8:00 16:0012:00

12:00

LB D

TV

Midnight

8:00 16:0012:00

12:00

LB D

TV

Midnight

8:30 16:0012:00

12:00

8:30 12:32• Initially: schedule reminders for earliest possible time• Apply “rewrite rules” to improve remders:

• Used preferred times for reminders• Combine “near” reminders that are compatible

• e.g.: “drink water” and “take pills”• Reschedule reminders for conflicting activities

Current Status of Autominder

• V.0 (Autominder + Pearl) field-tested for client acceptability on Pearl at Longwood Elderly Care Facility in Oakmont, PA, summer, 2001

• V.1 of Autominder implemented – Java, Lisp on Wintel machines

• Data collection with three Oakmont residents completed summer 2002; with Ann Arbor TBI patient summer 2003

Key Challenges for Cognitive Orthotics

• Technological– Advanced AI Techniques

– HCI

– Sensor Networks for Inference of Daily Activities

– Mechanisms to Ensure Privacy and Security

• Policy– Mechanisms to Ensure Privacy

– Reimbursement Policies