Transcript מצגת ההרצאה

Sarit Kraus
Bar-Ilan University
[email protected]
Multi-issue Negotiation
 Employer and job
candidate.
 Objective: reach an
agreement over
hiring terms after
successful interview.
 Chat based
negotiations
3
Training People in Negotiations
4
The Training Methods
 Classical role playing with another human
counterpart:
simple
scheduling,
requires other people
 Agent role playing
accessible and available 24/7
modeling different counterparts
 no significant differences were
found between the different training
methods (148 human subjects)
5/19
Virtual Suspect to Train Investigators
7
Culture Sensitive Agents for
Culture related Studies
 The development of standardized agents to be used in
the collection of data for studies on culture and
negotiation
Bargaining
Automated Mediators for
Resolving Conflicts
Automated Speech Therapist
First place of the TedMed 2014 innovation competition for startups in medicine
11
Language Tele-Rehabilitation:
Monitoring and Intervention
‫טפוך‬
Patient:
Patient:‫תפוח‬
‫השניה‬
Cyndi:
‫האותיות גם‬
‫נכונות‬
‫והשלישית‬
‫זה פרי‬
Currently in use by patients.
12
Agent Supports Deliberation
Past
deliberations
accumulative
data
Agent
Update
Current
deliberation
Offer arguments
=Obtains information
13
Supporting Robots-Human Teams
15
Sustainability: Reducing Fuel
Consumption
Persuasion
Why not use only game theory?
 Game theory is a study of strategic decision making:
"the study of mathematical models of conflict and
cooperation between intelligent rational decisionmakers“
 Results from the social sciences suggest people do not
follow game theory strategies.
People Often Follow Suboptimal
Decision Strategies
 Irrationalities attributed to
 sensitivity to context
 lack of knowledge of own preferences
 the effects of complexity
 the interplay between emotion and cognition
 the problem of self control
Why not Only Behavioral Science
Models?
 There are several models that describe human
decision making
 Most models specify general criteria that are context
sensitive but usually do not provide specific
parameters or mathematical definitions
Why not Only Machine Learning?
 Machine learning builds models based on data
 It is difficult to collect human data
 Collecting data on specific user is very time consuming.
 Human data is noisy
 “Curse” of dimensionality
Methodology
Human
behavior
models
Human
specific
data
machine
learning
Human
Prediction
Model
Data
(from
specific
culture)
Game Theory
Optimization
methods
Take action
22
22
An Experimental Test-Bed
Interesting for people to play:
 analogous to task settings;
 vivid representation of strategy
space (not just a list of outcomes).
Possible for computers to play.
Can vary in complexity
 repeated vs. one-shot setting;
 availability of information;
 communication protocol;
 Negotiations and teamwork
25
25
CT Game
 100 point bonus for
getting to goal
 10 point bonus for each
chip left at end of game
 Agreement are not
enforceable
Collaborators: Gal, Haim,
Gelfand
26
26
An Influence DiagramTwo rounds interaction
Probability
of
acceptance
Probability
of transfer
Methodology
Human
behavior
models
Human
specific
data
machine
learning
Human
Prediction
Model
Data
(from
specific
culture)
Game Theory
Optimization
methods
Take action
Prediction of Acceptance and Reliability
Data Collection:
Human playing against adaptive agent (PURB)
Reliability Measure
People
(Lebanon)
People
(US)
Co-dep
Task
indep.
Task
dep.
Average
0.96
0.94
0.87
0.92
0.64
0.78
0.51
0.65
29
Personality, Adaptive Learning
(PAL) Agent In thisDatadatafrom
set the
“Nasty Agent”: Machine
Less reliable when Learning
fulfilling its
agreement
People
adapt their
behavior to
their
counterparts
.
Lebanon
people
specific
almost
always kept
culture
the agreements 
PAL never kept
agreements
Optimization
methods
Human
Prediction
Model
Take action
30
Pal vs Humans
The model’s used by PAL had a very
low accuracy prediction of people
actually played against PAL
250
200
150
PAL
Human
100
50
0
U.S
Lebanon
Israel
31
Multi-issue Negotiation: The
Negotiation Scenario
 Employer and job
candidate
 Objective: reach
an agreement over
hiring terms after
successful
interview
32
Chat-Based
Negotiation
33
NegoChat’s Offers
 Find a “good” offer to use as anchoring
 Predict which offers will be accepted using on past
negotiation sessions and create clusters of possible
offers per time period
 Aspiration Adaptation Theory (AAT)
 proposes issues sequentially based on aspiration
scale learned from data
 retreat from previous values
Collaborators: Avi Rosenfeld, Zukerman,
Dagan, Gelfand
[email protected]
34
34
600
1200
500
1000
400
800
300
600
200
400
100
200
0
0
KBAgent
Time to Reach Agreement
Agent Score
NegoChat vs KBagent in Israel
Agent Score
NegoChat
Time (sec)
35
NegoChat in Egypt
 Collect new data of human vs human (negotiations
much longer)
 Build classifier
 Extract aspiration list
 Got many complaints on the agent.
 Challenges in running experiments.
 Results: slower than in Israel; lower score; almost the
same happiness and fairness; females scored much
lower.
36
One operator – Multiple robots
 Search And Rescue (SAR)
 Warehouse operation
 Automatic air-craft towing
 Fire-Fighting
 Military applications
 Etc..
Semi-Autonomous Robots
Noisy signals
Controls the robots
Agent
Controls the robots
Agent design
Data on
human
behavior
30 human
Operators in
simulation
Data on robots
performance
Machine
Learning
Human model
Optimization
Provide
Advice
150 hours of
simulations (no
human operator).
Robot model
Evaluation: Three Environments
12 subjects
16 subjects
16 subjects
Objects found per condition
Simulated office
Physical office
Simulated
warehouse yard
Computer Games for Learning
 Relatively low impact
 Possible Reason: lack involvement of the teachers
 Better results with a human teachers in the loop.
Capabilities for Agents in Learning
 Planning
 Monitoring
 Intervention
 Encouragement
45
Human-agent Working Together
 Planning: human forms the basic learning plan;
agent adapts it to the student’s progress.
 Monitoring & Intervention: Agent performs M&I
and asks the human for help when it is not sure.
 Encouragement: point up to the human
that a student needs encouragements.
46
Team of Agents and Teacher
 Many automated agents provide service to many
students.
 One human teacher supervises many students.
 Special agent
supports the
teacher.
47
Reducing Energy Consumption
by Advising People on how to set the
Climate Control System
GM Chevrolet Volt 2011
Presenting Advice to User
Presenting Advice to User
~80% of drivers explicitly accepted.
Methodology
 Collecting data on 38 subjects for an hour session in
the car)
 Predicting drivers reactions to offers
 Modeling the car and environment
 Integrating into an MDP
 Solving the MDP
51
Evaluation
 45 drivers - 15 per condition, 3 rounds.
The lower the better.
Why Did MDP Outperform the
SAP?
Agent
Avg. go eco %
Avg. save %
Avg. consumption
MACS
0.835
23.1
0.174
SAP
agent
0.641
33.7
0.237
 SAP was aggressive.
 Some subjects stopped clicking on the advice.
This does not look
as a real person!!
Remove the
Avatar!!
AniMed evaluation
Agents interacting proficiently
with people is important
Human
Data
behavior
(from
Challenges:
models
How to integrate machinespecific
culture)
learning
and
behavioral
[email protected]
Human
machine
models?
How to use in
learning
specific
data
agent’s strategy?
Challenges:
Game Theory
Optimization
methods
Human
Experimenting with people
Prediction
is very difficult !!!
Model
Working with people from
other disciplines is
challenging.
Take action
56