autoMinder
From: University of Michigan EECS News - Spring/Summer 2004

Autominder on the Move

The United Nations estimates that the percentage of people older than 60 will
double by 2050, while those older than 80 will quadruple. As the world's
population ages dramatically, Professor Martha Pollack is devising
artificial-intelligence systems intended to make the lives of these older
adults more manageable. 

Pollack sees that the elderly face a range of challenges - including
physical, social, emotional, and cognitive - and her work is designed to
assist those who have suffered from cognitive decline, especially decreased
memory. She is developing an intelligent cognitive orthotic system flexible
enough not only to remind people of the tasks they need to perform, but also
to schedule the reminders, cancel them, or change their nature, based on what
a person actually does and how he or she responds to the reminders issued.
The system is called Autominder. 

Autominder is an outgrowth of the Initiative on Personal Robotic Assistants
for the Elderly, a multi-university, multi-disciplinary research effort begun
in 1998, while Pollack was Director of the Intelligent Systems Program at the
University of Pittsburgh and a Professor in the Department of Computer
Science. Pollack has continued this effort since joining the EECS faculty in
September 2000. The project is funded by the National Science Foundation, in
partnership with Carnegie Mellon University and the University of Pittsburgh.
Pollack has received additional funding for this effort from the Intel
Corporation. 

Autominder has already been deployed on customized robots, or "Nursebots"
(the first was named Flo, the second one Pearl), which were designed and
built at Carnegie Mellon University, and have been sent on preliminary trials
with residents of the Longwood Retirement Community in Oakmont, Pennsylvania.
At nursing homes, getting seniors to and from their doctor's and other
appointments becomes a major time-intensive activity, especially for the
overworked and often understaffed personnel who would prefer to spend quality
time with these individuals. The Nursebots are able to provide reminders, and
help guide individuals from one location to another. 

From Robots to Handhelds
"We are continuing to pursue this work at Michigan," says Pollack, but we are
exploring other platforms than mobile robots. We are now experimenting with
handheld devices, which then communicate with sensors in the user's
environment. Pollack notes that while mere are "reminder systems" on the
market, they are "simply glorified alarm clocks that tell users to do this at
one time, and do that at another time. They're not flexible enough. With
Autominder, we can track and monitor the plans of an older adult with memory
impairment, and can do this in a flexible and adaptive way. Our goal is to
help elders stay living in their own homes longer."  

The technology is also useful for other people with memory or executive
function disorder. In fact, It is currently being tested with patients who
have suffered traumatic brain injuries. This work is being conducted in
association with Dr. Ned Kirsch, a faculty member at the UM Medical School,
Physical Medicine and Rehabilitation, and adjunct professor of Psychology. 

"Caring for the traumatic brain injury patient is very hard on the
caregiver," said Pollack, "so that any autonomy you can give the individual
in their own home is precious. The caregiver doesn't have to be there
constantly caring for the person, which is very good for the individual
because they are doing it `on their own' - with assistance from the reminder
system. We all need assistance in one way or another. A person who wears
glasses gets assistance - we don't say "they can't see on their own" - they
just need glasses. It's the same approach here."  

Upgrading through Reinforcement Learning The sophisticated Autominder
architecture is essentially a three-part interactive structure. The Plan
Manager stores the schedule of required tasks; much of the work here involves
constraint-based temporal reasoning. The Client Modeler maintains a record of
what the system believes has actually occurred. For example, explains
Pollack, "if you go to the refrigerator, can it be inferred that you got
breakfast? Well, it depends how long you stood in front of the refrigerator,
what time it is, when you usually eat breakfast, how long you stayed in the
kitchen. All these factors go into determining what you've done in the
kitchen."  

The third component is the Intelligent Reminder Generator. It compares the
Plan Manager and Client Modeler data and decides whether, when, and how
upcoming reminders should be presented. "Studies have shown that it's not
good to 'over-remind.' If you do, people become overly dependent on the
system. We want to allow people to initiate activities on their own whenever
possible," said Pollack. For example, assume your grandfather needs to take
out the trash no later than 8pm. Autominder would wait to see if he does this
on his own. But if his plans change, so that he is going to visit a neighbor
at 7pm, Autominder might determine that he needs a reminder to set out the
trash before he goes. 

Professor Pollack recently began collaborating with Professor Satinder Singh
Baveja to find ways to improve the crucial reasoning function of the
Intelligent Reminder Generator. Singh's main area of research is
Reinforcement Learning (RL) - a machine-learning approach to adapting and
thus improving automated decision-making through reinforcement techniques. 

Though RL has not often been applied to the design of human-computer
interaction systems such as Autominder, Singh says in this case it's a
"natural fit" because Pollack's system is "a great application and a
wonderful opportunity for a reinforcement learning researcher to make an
impact with something that helps people." Together they have been building
and testing the behaviors of a machine-simulated "person," customizing and
adapting it to short term changes in activity schedules as well as long term
effects such as increased memory deterioration. Professor Pollack says this
will ultimately lead to a new and improved intelligent reminder module with
"a different computational mechanism inside of it."  

In addition, Singh explains that he wants the system "to adapt automatically
- which is how an individual usually behaves--and to learn what an individual
user's preferences are. Some people may be very forgetful about medicine, but
may remember to eat. Others may have the reverse situation. Even depending on
the time of day, they may remember to do some things, and not others. We
would like to build a system that will learn overtime, through repeated
exposure, to actually modify its strategy to be most effective for that
person."  

Pollack adds, "What we'd like to do Ä what absolutely needs to be done, but
is still in the lab - is to embed the Autominder system in a ubiquitous
computing environment in the home, with motion detectors and contact sensors
on the refrigerator and doors; pressure sensors in the bed; and flow sensors
on the faucets. Then we will get it all hooked up wirelessly, which is fairly
normal these days." Pollack envisions transmitting and receiving wireless
signals in hearing aids, cell phones, television remotes, and other devices
in the home or on one's person. Singh reinforced this goal, stating that the
home sensors will "detect what it is that the user has already done, will
have a sense of what they should be doing the remainder of the day, and in a
context sensitive way, will remind them of what they should be doing. So if
they are supposed to drink water 4 times a day, and it is the afternoon, you
might want to remind them then, so they don't have to drink late in the
evening which may cause them to need to get up in the middle of the night."  

The Pull of AI for Student Researchers
When Joe Taylor was graduating from Notre Dame, he wanted to work in the area
of Al; coming to UM and working with Prof. Pollack seemed to click. Joe
enabled Autominder to be run on any system connected to the internet, while
allowing for remote modules to be run on a variety of systems that would be
more useful to patients. "It's good that the patient doesn't have to carry
around an entire computer," said Joe. "We also hope to branch out into other
types of interfaces, like a hearing aid - any other interface that will make
it easier on the patient. While Al can't do everything, Joe feels that "the
brain can really use help. And that's what we're trying to do for the elderly
with cognitive impairment. We want them to be able to function on their own,
and allow them to stay in their own homes."  

Ph.D. candidate Bart Peintner was similarly drawn to the field of artificial
intelligence Ä and to Professor Pollack's work in particular. "All of my
background was in programming, but I took one of Martha's classes in which
she explained her work, and I became interested and asked if I could do a
summer project." Peintner has worked on Autominder ever since - and his
Higher-Order Markovian Reasoner (HOMR) software is the key ingredient that
enables Autominder not only to gather, but also to reformulate information
obtained from sensors in such a way that it can act on that information and
change its messages accordingly. "If the sensor data indicates they are
already cooking breakfast," Peintner says, "we don't want to remind them to
cook breakfast."  

Peintner admits that while "a computer system can never be as complex as the
human mind, for simple tasks it can be faster."  

Boomers: New Age for Old Age
While most researchers have found that seniors are more likely to accept
high-tech gadgetry if it is packaged in a familiar form - such a telephone,
or even a robot - aging Baby Boomers will likely be more willing and able to
accept nearly any form of technological assistance. The Baby Boom generation
- whose youngest members turn 65 in 2011 - is not only familiar with
technology; it is already highly dependent on it. Cell phones and personal
digital assistants are ubiquitous - and so, Pollack expects, within 10 to 15
years, will be the practical, affordable, off-the-shelf memory aid products
she and others are currently working to develop. 

Prof. Pollack and graduate student Joe Taylor demonstrated the handheld,
artificial-intelligence device at the largest aging services technology
demonstration of its kind March 16, 2004, at the Dirksen Senate Office
Building in Washington, DC, organized by the Center for Aging Services
Technologies (CAST) (see http://www.agingtech.org for information about CAST). 

Prof. Pollack testified at a hearing of the U.S. Senate Committee on Aging
April 27. (see http://www.agingtech.org/announcement.aspx?id~3)

Read - and see - more about it! Click on http://www.ecs.umich.edu/~pollackm
and click on "In the News". 

Captions:

1. Professor Martha Pollack