Transcript Effective Interaction Strategies for Adaptive Reminding

Efficient Interaction Strategies for Adaptive Reminding
university
Julie S. Weber & Martha E. Pollack
Computer Science and Engineering
michigan
artificial
intelligence
laboratory
University of Michigan, Ann Arbor, MI, U.S.A.
Adaptive Reminder Generation
• Features:
- Justifications
- Adaptive Reminder Granularity
- Adaptive Signaling
• Testbeds:
- Assistive Technology for People with Cognitive Impairment
(e.g., Autominder [2])
- Smart Office Assistants (e.g., CALO [1])
• Adaptation Techniques:
- Supervised Learning, with an active learning component
- Reinforcement Learning
Signaling
Intended Approach
• Different users may respond more readily to, or have different
preferences for, certain types of reminders
11:35am
Lunch meeting
in cafe at noon
Time for your walk
with Rhonda
Learning
Justifications
• A single, sparse reminder may not be enough to convince that
user to perform the task immediately.
• With a justification, the user is more likely to comply:
[5:30pm] Knight Rider is on at 6:00, so there
is just enough time for exercise beforehand.
[2:00pm] Reminder; your paper deadline is at
4:00, and your 3:00 meeting lasts 2 hours.
Creating a system that builds upon the initial efforts reported in [3]
and [4] requires:
• Adding additional features to the action space of an adaptive
reminding system, such that it learns to decide not only when to
issue a particular reminder, but also how the reminder should be
issued.
• Deciding when it is appropriate to perform active learning of
user preferences over the features of a reminder.
A new reminder plan is generated at the start of each day, based on the
user’s current schedule and to-do items. As new activities and tasks
are scheduled and depending on the user’s current task, this plan is
updated to reflect those new tasks requiring reminders.
Meeting with CEO in 10 minutes
OK
Techniques taken from the area of machine learning can effectively
enhance our solutions to the challenges described above: the system
can learn how best to interact with a particular user based on that
user's pattern of compliance with the reminders received.
When a user performs or fails to perform the task or activity
associated with a particular reminder, the system can update its policy
for issuing reminders accordingly.
Architecture
• Reinforcement learning to determine when certain reminders
should be issued [3].
• Supervised learning of scheduling preferences by way of active
learning [4].
Calendar
To Do List
Location Info
Reminder Granularity
References
[1] Mark, B., and Perrault, R. C. CALO:
Cognitive Assistant that Learns and
Organizes.
http://www.ai.sri.com/project/CALO
• Must track level of impairment
[8:30am] Time to prepare breakfast.
[8:30am] Remove eggs from refrigerator.
[8:31am] Prepare frying pan on stove.
of
...
[2] Pollack, M. E., Brown, L., Colbry, D.,
McCarthy, C. E. Orosz, C., Peintner, B.,
Ramakrishnan, S., and Tsamardinos, I.
Autominder: An Intelligent Cognitive
Orthotic System for People with Memory
Impairment. Robotics and Autonomous
Systems (44) 273-282, 2003.
[3] Rudary, M., Singh, S., and Pollack, M. E.
Adaptive Cognitive Orhthotics:
Combining Reinforcement Learning and
Constraint-Based Temporal Reasoning.
International Conference on Machine
Learning 2004.
[4] Weber, J. S., and Pollack, M. E. EntropyDriven Online Active Learning for
Interactive Calendar Management.
International Conference on Intelligent
User Interfaces 2007.
Task Info
Adaptive
Reminding
System
Reminder Plan
Reminders
Updates
Summary
There are a number of dimensions to intelligent reminding that must be
explored to create a personalized, adaptive system. Two learning
techniques that we propose to explore in this context are reinforcement
learning and supervised learning directed by an active learning
component.