Sarit Kraus Bar-Ilan University
29
Transcript of Sarit Kraus Bar-Ilan University
Computerized Agents that Interact with People
2
Automated
Speech Therapist
Persuading people
to save energy
Culture sensitive
negotiation agent
Automated
mediators
Agent supports
Argumentation
Supporting
Teams of
Robots &
Operator
Virtual suspect for
training investigators
Training people in
negotiations
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
3 Kahneman Selten
Why not Only Equilibrium Agents? Nash equilibrium: stable strategies; no agent has an
incentive to deviate
Results from the social sciences suggest people do not follow equilibrium strategies:
Equilibrium based agents played against people failed.
People rarely design agents to follow equilibrium strategies.
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
5
Why not Only Machine Learning?
Machine learning builds models based on data
It is difficult to collect human data
Collecting data on a specific user is very time consuming.
Human data is noisy
“Curse” of dimensionality
6
Methodology
7
Human Prediction
Model
Take action
Machine Learning
Game Theory Optimization
methods
Data
Human behavior models
Human specific data
Ariel Rosenfeld, Amos Azaria, Sarit Kraus, Claudia V. Goldman, Omer Tsimhoni
CCS reduces ~10% of the car’s power efficiency!
Reduced ecological footprint.
Extending travel distance of EV.
Economically efficient.
Why bother?
• Driver’s and system’s goals are
partially conflicting.
Partially Conflicting Interests
Let’s minimize energy
consumption... I’m Hot!
Ariel Rosenfeld et al. AAMAS 2015 @ Istanbul,
Turkey. May 2015
Challenges in Repeated Advice Provision in CCS in Real Cars Repeated interaction
Drivers’ preferences.
Long-term effect of advice.
Changing environment.
Estimating expected energy consumption.
Climate Control System (In GM Chevrolet Volt 2011)
Advice
Controls
Effects
Agent
Effects Effects
Goal: minimize the accumulative energy consumption.
Driver/Environment models We recruited 38 subjects. (not that easy!)
Each subject spent 30 min. in the car,
simulating 3 different trips.
Subjects were presented with different advice.
ML algorithm for extracting probabilities:
Drivers likelihood to accept an advice
Car’s condition likelihood to change.
Presenting Advice to User
Presenting Advice to User
~80% of drivers explicitly accepted.
78% accuracy (post-hoc).
Influential Features: Current internal temperature. Change from current setting (Reference point). % of accepted advice (Trust). Saving percentage (Expectation bias).
Not influential:
External temperature. Average temperatures\fan. Accepted deltas.
Prediction of drivers reactions
MACS – MDP agent
Uses the predictions for the transition function.
State of the art – SAP (Azaria et al. 2012)
Considers the Social Utility of advice.
The weight provides a trade-off between short and long term gain.
Agents
driveragent UwUw )1(
Evaluation
45 drivers - 15 per condition, 3 rounds.
The lower the better.
Why Did MDP Outperform the SAP?
SAP was aggressive.
Some subjects stopped clicking on the advice.
Agent Avg. go eco % Avg. save % Avg. consumption
MACS 0.835 23.1 0.174
SAP agent
0.641 33.7 0.237
Ariel Rosenfeld , Noa Agmon,
Oleg Maksimov, Amos Azaria,
Sarit Kraus
One operator – Multiple robots
Search And Rescue (SAR)
Warehouse operation
Automatic air-craft towing
Fire-Fighting
Military applications
Etc..
Semi-Autonomous Robots
Controls the robots
Noisy signals
Agent
Controls the robots
Agent design
Provide Advice
Machine Learning
Optimization
Data on robots performance
Data on human behavior
Robot model Human model
150 hours of
simulations (no
human operator).
30 human
Operators in
simulation
Evaluation: Three Environments
16 subjects
16 subjects
12 subjects
Objects found per condition
Simulated office Physical office Simulated
warehouse yard
29
Agents interacting proficiently with people is important
Human Prediction
Model
Take action
machine learning
Game Theory Optimization
methods
Human behavior models
Data (from specific culture)
Human specific data
Challenging: Experimenting with people is very difficult !!! Working with people from other disciplines is challenging.
Challenging: How to integrate machine learning and behavioral models? How to use in agent’s strategy?
