Download Chapter 2 - Gert Kootstra

Chapter 2
Alper Aydemir
CVAP, KTH
Origins of Artificial Intelligence
• Classical approach: “the brain” is the one and
only holy grail of AI
• 1956 Darthmouth meeting
– Good old Fashioned AI
• Expert systems were thought of replacing
human intelligence “soon”
• Symbol processing, set of rules, when we have
enough rules  success!
Origins of Artificial Intelligence
• The world state can be described as a
configuration of clear, unambiguous set of
symbols.
– This is how computers work, everything should be as
precise as possible: not in nature!
• The more complex the model is the more care to
maintain it
• GOFAI has failed to live up to its promises
• Let alone human-like, even insect-like intelligence
is not achieved yet
Origins of Artificial Intelligence
• Common sense: “surely a hallmark of
intelligence” (same as chess?)
• Very large set of “rules” is crucial to teach
robots common sense
• Speech is also another area where GOFAI
failed
– Google Voice
• Asking the right question
Robots with bodies
• Brooks: “The world itself is its own best
model”
– Studying insects
• Simply act and react, no need to model the
world to its atoms
– Decision theoretic planning  super expensive
• Truly autonomous robots need to learn the
environment on their own
– Kuiper’s garbled pixels
Robots with bodies
• Interaction with the real world is never clean
but messy and ill defined.
– Modeling everything is simply impossible
• Effect of embodiment idea on AI
– Research on animal intelligence
– More collaboration with biology, neuroscience
Neuroscience
• 80’s: The rise of Artificial Neural Networks
– Simulation/abstraction neurons and their connections
– Some success in computer vision (classification,
pattern recognition) and language acquisition
• Connectionist approach
– “I am my connectome”, S. Seung
• However this was mostly without embodiment
• Recently, taken interest in embodiment,
neuroscience gained popularity again
– Computational Neuroscience, Neuro-Informatics etc.
Multidisciplinary AI
• Computer Science
• Linguistics / Computational linguistics
• Philosophy
– A. Sloman
• Biology, robotics, biomechanics
• Embodied AI called robotics, biomimetics,
adaptive locomotion et cetera.
• Various conferences sprung up
Biorobotics
• Build robots which mimics certain organisms
(typically simple and non-human)
• Example: Sahabot mimicking Tunisian desert
ants
• Snapshot model by Sussex Univ.
– Short-range navigation based on horizon
appearance
– Long-range navigation based on light polarization
– Empirically proven model
Biorobotics
• No need for a detailed map, course,
completely imprecise navigation and
localization
– SLAM!
• Snakebots by Hirose Lab. – Univ. of Tokyo
• Auke Ijspeert’s salamander – EPFL
• ...
Developmental Robotics
• Brooks falsely claimed “we have reached insectlike intelligence” and moved to a “sexier” topic:
how does human learning work?
• Humanoid robotics became popular in Japan
– However mostly concentrated in mechanics
• With Brook’s “Cog” project a new interest in
human-like intelligence flamed up
• HRP humanoid project
• Ishiguro’s baby robot, RoboCub
Ubiquitous computing
• Moving away from interacting with computers
with only mouse/keyboard
• “Scatter” sensors everywhere
• More recently researchers became interested
in not only sensing but changing the world as
well
• Various ways of interacting with computers,
“wearable computers” and so gained
popularity
Multi-agent systems
• Important insight: complex behavior can
emerge from very simple rules in the
individual level
• Cellular automata: the Game of life,
movement of bird flocks
• Swarm intelligence, self-organization
– Book: “Order out of Chaos” by Prigogine
Self organization
• Emergence of a pattern from the local
interactions of many individuals
• Ant trails
– Marked by pheromones
– P(followingTrail(X)) ~ trail scent
– Shorter paths = more ants prefers  shorthest food
source gets consumed first!
• Modular robotics, self-configuration and selfassembly
• Murata, Tokyo Inst. of Technology
Multi agent systems
• Rather than emergence of pattern aimed at
solving a particular task
• RoboCup: a team of robots playing soccer
– Various leagues: small, humanoid, middle
– 100.000 spectators in Fukuoka match!
Evolutionary Robotics
• Trying to understand and simulate natural
evolution
– Genetic algorithms for designing electronic circuits
– Only “the brain” evolves so far in previous work,
though there are some few examples of the body
evolving.
Summary
• The journey of AI has changed significantly
from GOFAI to embodied systems
• Embodied intelligence is now the artificial
intelligence (or has become)
• By building synthetic systems (robots) we can
learn a lot about the nature of intelligence
• Also they allow testing of concrete ideas
rather than just thought experiments
– “Put your money where your mouth is!”