1RADAR – Scheduling Task © 2003 Carnegie Mellon University RADAR – Scheduling Task May 20, 2003 Manuela Veloso, Stephen Smith, Jaime Carbonell, Brett Browning,

1RADAR – Scheduling Task © 2003 Carnegie Mellon University

RADAR – Scheduling Task

May 20, 2003

Manuela Veloso, Stephen Smith,

Jaime Carbonell, Brett Browning,

(Jay Modi, Eugene Fink)


The Challenge

• Main Functions -- Calendar Management» Respond to meeting requests (extracted from ongoing email stream)

» Initiate meetings requests and establish meetings

» Continuously acquire user preferences and negotiation profiles

• Why not yet available » Requires capture and use of complex, ill-structured user preferences

» Continuous scheduling

» Management of rich multi-threaded information exchange under conflicting constraints and preferences

• Why now» Explore collaborative, user + EPCA, scheduling

» Build upon integration of many leading technologies, I.e., information extraction, constraint satisfaction, iterative scheduling

» Log, analyze, learn profiles to incrementally improve scheduling


Calendar Scheduling is Complicated

• Meeting constraints may be hard to satisfy, requiring counter proposals, or relaxing, or negotiation

• Pre-emption of a meeting can cause a ripple effect

• Users do not put all commitments in their calendars

• It may be necessary to secure additional resources (e.g., room, projection facilities)

• Preferences and interaction protocols will vary according to context and participants involved

• There may be several meeting requests in various stages of commitment at any given time


Diversity and Complexity

• Can we meet tomorrow at 10am?

• Can we meet with Pat some time this week?

• The admissions committee needs to meet every week until the end of February.

• The interested teaching AI faculty need to meet to schedule the courses for the Fall.

• We should arrange an AI retreat, as the one we did a few years ago.

Templates


The Approach

Email Stream

User

KnowledgeBase

LearningProcesses

ExtractorMessage Stream

Email Stream

Preferencesand Profiles

EditorCalendarDisplay

SchedulerManager

Need for “Sliding Autonomy”


Scheduler: Responding to a Request

Request: Template, T» When: Thursday 15th» Duration: 1 hour» Who: Visiting Researcher

(Priority: “medium”)» Where: 1502E NSH

Response, R:» 4:00 - 6:00

Infeasible

Commited Pending

11:00 - 12:30

2:00 - 3:00

4:00 -

Policy preference:Avoid lunch hour

Pending reservation but lower priority

… but would 1/2 hour be sufficient?

Generate Options

Evaluate Options

…

Preference Order:4:00 - 6:002:00 - 3:00

11:00- 12:00

Threshold


Manager: Multi-Thread Processing

Manuela’s Calendar

Manuela

Raj

Student

Steve

10am?

Time

Meeting request for blocked time

12pm

2pm

4pm

Student, Steve10am

Confirmed Pending




Manuela

Raj

Student

Steve

10am?

Resch. 12pm?

Time

Conflict: try rescheduling

12pm

2pm

4pm

Student, Steve10am

Confirmed Pending

Raj

Student, Steve



Steve’sCalendar

Manuela

Raj

Student

Steve

10am?

Resch. 12pm?

Time

Student12pm

2pm

4pm

Student, Manuela10am

Confirmed Pending

Conflict: try rescheduling

Student, Manuela



Student’s Calendar

Manuela

Raj

Student

Steve

10am?

Resch. 12pm?

12pmokay

Time

12pm

2pm

4pm

Manuela, Steve10am

Confirmed Pending

No conflict Manuela, Steve




Manuela

Raj

Student

Steve

10am?

Resch. 12pm?

12pmokay

4pm?

Time

Another meeting 12pm

2pm

4pm

Student, Steve10am

Confirmed Pending

Raj

Brett

Student, Steve



Manuela

Raj

Student

Steve

10am?

Resch. 12pm?

12pmokay

4pm?

2-4pm better

Time

Rescheduling difficult: suggest an alternative Steve’s

Calendar

Student 12pm

2pm

4pm

Student, Manuela10am

Confirmed Pending

Student, Manuela

Student, Manuela




Manuela

Raj

Student

Steve

10am?

Resch. 12pm?

12pmokay

4pm?

2-4pm better 2pm?

Time

Choose best alternative 12pm

2pm

4pm

Student, Steve10am

Confirmed Pending

Raj

Brett

Student, Steve


Student’s Calendar


Manuela

Raj

Student

Steve

10am?

Resch. 12pm?

12pmokay

4pm?

2-4pm better 2pm?

2pmokay

Time

Pending 12pm

2pm

4pm

Manuela, Steve10am

Confirmed Pending

Manuela, Steve




Manuela

Raj

Student

Steve

10am?

Resch. 12pm?

12pmokay

4pm?

2-4pm better 2pm?

2pmokay

2pmconfirmed

10amconfirmed

Time

12pm

Student, Steve2pm

4pm

Raj10am

Confirmed Pending

Brett

Confirmed


Main Tasks

• RESPONDING to a request for availability » Multi-thread conflicting:

– request, availability, response, reschedule

• INITIATING organizing a meeting» Request meeting» Collect replies» Merge and solve scheduling» Until solution is found

• LEARNING» Priorities, contexts, profiles


Manager and Scheduler

Manager

EmailExtractor

EmailGenerator

Scheduler

Knowledge Base

• Preferences• Profiles

R*

T T … T

T

Pending

T …

…History

mtg i

T


Manager and Scheduler

Manager

EmailExtractor

EmailGenerator

Scheduler

Knowledge Base

T

Pending

T …

mtg i+1 ……

History

mtg i

T

R*

T T … T

T’


The Science

• Algorithms» Dynamic, incremental constraint-based reasoning

» Priority-, preference-driven minimum disruption optimization

• Main open questions» How to effectively computer assist a user in calendar management?

» How to represent and exploit an ill-structured set of calendar scheduling preferences and profiles?

» How to learn these preferences and profiles from episodic logging?

• Novel ideas for open questions» Collaborative meeting scheduling based on context and history

– Acquired preferences in different contexts

– Acquired beliefs of scheduling preferences of others

– Determination of profiles for management

» Use of learned profiles to overcome user burden managing calendar

» Direct, closed loop integration with user email stream


Learning

• Accumulation of episodes

• Control learning – State/action models

• Probabilistic dependencies

• Statistical strategy selection

• Multiagent learning


The Impact

• Scientific advances» Continuous mixed-initiative scheduling

» Multi-threaded process management and logging

» Learning of interaction preferences and profiles

» Seamless integration of scheduler, manager, learner

• Performance» Full implementation – RADAR improves user’s activity

• GEMs – Generalized modules for similar activity management – extend to space task


The Plan

• Next Steps» Collecting data from the team

» Templates as a stub for email extractor

» Representation of scheduling preferences and profiles

» Assemble architecture:

– Scheduler, manager, knowledge base, user, learner

» Scheduling engine

» Logging of scheduling process

• Long Run» Learning over collected data

» Development of protocols and algorithms for distributed resolution of scheduling conflicts

» Multiagent collaboration and sharing among EPCAs

1RADAR – Scheduling Task © 2003 Carnegie Mellon University RADAR – Scheduling Task May 20, 2003 Manuela Veloso, Stephen Smith, Jaime Carbonell, Brett Browning,

Documents

scheduling slide

scheduling build

steve slide

pm student

iterative scheduling

confirmedpending slide

autonomy slide

given time slide