From Imprinting to Adaptation: Building a History of Affective Interaction

1
From Imprinting to Adaptation Building a
History of Affective Interaction
Arnaud J. Blanchard BlanchardArnaud_at_Gmail.com Lola Cañamero L.Canamero_at_herts.ac.uk

• Adaptive System Research Group
• School of Computer Science
• University of Hertfordshire, UK

2
Introduction

• Imitation / synchronisation useful in development
• To learn
• To communicate
• To survive, satisfy needs
• Problems
• What and how to imitate
• When to imitate
• Who to imitate
• What makes a successful imitation

3
Approach

• Approach
• Bottom up as simple as possible
• Per-Ac no explicit representation of the world
• Biologically possible
• Developmental
• Consequences
• Perception and action at the same level
  (Gaussier, Prinz)
• The goal is not to do an action but to obtain a
  perception
• We will use Goal Perception

4
Who and when to imitate ?

• Imitation has to be done
• At the right moment
• With the right agents
• During the development
• The agent has to learn
• Where is better to be
• Who is better to imitate
• Affectivity between new born and care-taker

5
Affective bonds

• Emotional Contagion E. Hatfield (1994)
• Affective bonds increase imitation /
  synchronization
• Imitation / synchronization increase affective
  bonds
• Problems
• How to manage the affective bonds ?
• When to lead interaction ?
• To start these bonds we can observe the nature

6
Imprinting phenomena

• 1930s Lorenz
• Birds follow the first thing they saw
• Usually the mother but also humans or objects
• The birds try to keep the first perception
• This phenomena can be a cue to create affective
  bonds in robots

7
Implementation

• Goal Perception
• Average of the first perceptions
• We can express this as a stochastic learning
• The learning rate decreases with the time from
  hatching

learning rate
time from hatching
8
Experimentation

• We apply it to distance perception

9
Adaptation

• Observations
• Interesting as it reproduces natural behavior
• But not very useful for daily interaction
• Bateson (2000)
• Imprinting is not an instantaneous, irreversible
  process
• Notion of comfort
• Distance to ideal values (homeostatic control)
• Comfort between 0 and 1
• Decreases when the distance increases

10
Relativity of the comfort

• The learning rate
• Does not depend on the absolute value of comfort
• Depends only on the relative value of comfort
• Learning faster when the comfort is high
• How fast is faster ?

11
Times scales

• We use different time scales Desired
  Perceptions
• Long terms very slow changes
• Short terms very fast changes
• We use several time scales in between

learning rate
time from hatching
12
Acting

• Selecting a time scale
• High comfort using short time scale
• Low comfort using long time scale

13
New perceptions

• To learn, the robot has to accept new perceptions
• When the comfort is very high (no risk)
• When the comfort is very low (waiting for help)
• ?Varying openness to the world

activity
comfort
14
Final architecture
15
Experimentation

• Comfort management
• The comfort increases by touching a side sensor
• The comfort decreases with the time
• The robot tries to keep a high level of contact
• The frequency of beeps reflects the level of
  discomfort

comfort
lot of contacts
few contacts
16
Video
17
Results
18
Conclusion

• Affective bonds without symbolic representation
• Similarity with biology
• Robustness of the learning with the time
• Learning and acting simultaneous
• Continuity with reinforcement learning
• Capacity to decide when to lead the interaction

19
Perspectives

• Further works
• Notion of fear or avoidance
• Active exploration with notion of pleasure
• Taking account of different contexts
• Taking account of different modalities
• More analysis with biology
• Low level of imitation depending on affective
  bonds

20
Synchronization seems to make happy
21
I want make my robot happy
22
Thank you for your attention
Silvia Helena Cardoso, PhD and Renato M.E.
Sabbatini, PhD
Arnaud J. Blanchard e-mail BlanchardArnaud_at_Gmail.
com homepage arnaudblanchard.free.fr
Lola Cañamero e-mail L.Canamero_at_herts.ac.uk homep
age homepages.feis.herts.ac.uk/comqlc/