Challenges for Artificial Cognitive Systems

From EUCog Wiki

Jump to: navigation, search

Scientific Atelier at the International Summer School on Embodied Language Games and Construction Grammar, Cortona 29 August-4 September 2009

Contents

Link to event page

http://www.alear.eu/events/cortona2009/Ateliers.html

"Challenges" Events

Organisers

Antoni Gomila, Vincent C. Müller, Robert Porzel, Yulia Sandamirskaya

Summary of Workshop Results

Aims of the atelier

  • Develop the theoretical requirements for a set of challenges (A)
  • Develop a framework for the formulation of a specific set of challenges (B)

Theoretical requirements for a set of challenges

  • providing a long-term vision and fruitful orientation for present work
  • theory and strategy neutral, not fashion dependent, open to new approaches and strategies
  • not domain specific, not oriented towards toy problems or scenarios
  • flexible
  • systematic

-- The main challenge in defining challenges is the development of a metrics

Our result

A metric for what is more or less of a challenge for a cognitive system:

  • system that successfully pursues goals; flexible (adaptive) and autonomous
  • defined in terms of behavior, not internal structure of underlying theory
  • not defined in terms of extant natural cognitive systems (allowing alien AI, not identifying AI and cognitive science)
  • not specifying where a 'system' must be individuated (a system may have sub-systems: birds & flock of birds)
  • cognition allows of degrees, perhaps specific levels
  • degrees can be measured by the predictability of behavior, given a particular input or environment (how many parameters to you have to freeze to predict the system 100%)
  • Note: the higher the cognitive level, the lower the controllability and security of a system

A line of challenges

... for example:

  1. input-output reflex system
  2. digital input leads to determinate results (finite state automaton)
  3. analog input is accepted and leads to flexible results (control structure)
  4. re-shape itself (work offline, anticipation, planning, social learning, flexible use of resources [e.g. time], real cooperation)
  5. shape the re-shaping of itself, i.e. shape my own preferences (freedom of will)

For each domain of cognitive systems (perception, motor control, spacial orientation, social interaction, ...) specific challenges should be defined on these levels.

Benchmarks

For each of the challenge levels, esp. the domain-specific ones, benchmarks should be defined:

  • benchmarks are abilities
  • meeting a benchmark is to reach a level x for a domain

For example:

  • 'clean up the house',
  • 'walk the dog',
  • 'make a cup of coffee'
  • 'be the porter',
  • 'replace this worker/soldier/...'
  • 'do what this animal can do' (single cell, worm, insect, reptile, bird, mammal)
  • 'do what humans cannot do' [space, underwater, underground, ...]
  • 'RoboCup', 'RoboCup@home'
  • ... note that these are traditionally massively underestimated!
  • ... note also, that societal impact & ethical relevance is underestimated (we need sociology/philosophy/psychology)

List of files

Related Areas

Retrieved from "http://www.eucognition.org/eucog-wiki/Challenges_for_Artificial_Cognitive_Systems"
Personal tools