Elephants Dont Play Chess

1
Elephants Dont Play Chess

• By Rodney A. Brooks
• Presented by Yan Ha

2
Purpose of the paper

• 2 approaches of AI
• explore second approach which emphasizes ongoing
  physical interaction with the environment as the
  primary source of constraints (physical
  grounding)
• examples and future work

3
Classical AI

• classical AI (symbol system hypothesis) is flawed
• bases its decomposition of intelligence into
  functional information processing modules
• none of the modules themselves generate the
  behavior of the total system
• improvement in the competence of the system
  proceeds by improving the individual function
  modules

4
Nouvelle AI

• base on physical grounding hypothesis
• bases its decomposition of intelligence into
  individual behavior generating modules, whose
  coexistence and co-operation let more complex
  behaviors emerge
• improvement in the competence of system proceeds
  by adding new modules

5
Symbol System Hypothesis

• states that intelligence operates on a system of
  symbols
• perception and motor interfaces are sets of
  symbols which the central intelligence system
  operates
• symbols represented entities in the world (ex
  objects, emotions, molecules)

6
Inadequacy of Symbol Systems

• symbol systems assume a knowable objective truth
• there is a limit on the complexity that modal
  logics can be built for the symbolic system
• frame problem impossible to assume anything not
  explicitly stated
• rely on emergent properties

7
Physical Grounding Hypothesis

• states that to build an intelligent system, its
  representations need be grounded in the physical
  world
• no need for traditional symbolic representations
• the world is its own best model

8
Physical Grounding System

• connect system to the world via set of sensors
  and actuators
• no typed input and output
• built from the bottom up
• system has to express its goals and desires as
  physical action, and extract its knowledge from
  physical sensors
• forms of low-level interfaces have consequences
  that ripple through entire system

9
Subsumption Architecture

• built on a computational substrate that is
  organized into a series of incremental layers,
  each connecting perception to action
• substrate is network of finite state machines
  augmented with timing elements
• subsumption compiler???

10
Old Subsumption Language???

• each AFSM has a set of registers and timers
  connected to a conventional FSM which control a
  combinational network fed by the registers
• registers can be written by attaching input wires
  to them and sending messages from other
  machines-get replaced
• arrival of a message can trigger a change of
  state in the interior FSM

11
New Subsumption Language

• groups multiple processes (AFSM) into behaviors
• message passing/suppression/inhibition between
  processes within a behavior, or between behaviors
• behaviors act as abstraction barriers-one
  behavior cannot reach inside another

12
Physically Grounded Systems

• seemingly goal-directed behavior emerges from the
  interactions of simpler non-goal-directed
  bahaviors

13
Allen

• sonar range sensors and odometry, offboard lisp
  machine
• 3 layers
• first layer-avoid both static and dynamic
  obstacles
• sit in middle of the room until approached, move
  away and avoid collision
• sonar return represented repulsive force
• vector sum tells robot where to move

14
Allen

• additional reflex halted robot whenever there was
  something right in front of it and it was moving
  forward
• Second layer-randomly wander about every 10 secs
• Third layer made robot look for distant places
  and try to head towards them
• suppress the direction desired by the wander
  layer

15
Tom and Jerry

• 2 identical robots that demonstrate how little
  computation is necessary to support subsumption
  architecture
• 3 1-bit infrared proximity sensors, 3-layer
  system
• first layer-use vector sum of repulsive forces
  from obstacles for obstacle avoidance
• second layer-wander about
• top layer-detect moving objects and create a
  follow behavior
• wander behavior was suppressed when chasing
  objects

16
Tom and Jerry

• demonstrate the notion of independent behaviors
  combing without knowing about each other (chasing
  obstacles but staying back a little)
• subsumption architecture can be compiled to gate
  level

17
Squirt

• smallest robot (50 grams)
• 8-bit computer, on board power supply, 3 sensors
  and a propulsion system
• acts as a bug, hiding in dark corners and
  venturing out in the direction of noises, after
  noises are gone, looking for a new place to hide
  (near to previous noises)

18
Squirt

• high level behavior emerges from a set of simple
  interactions with the world
• its lowest level monitors a light sensor and
  causes it to move in spiral pattern searching for
  darkness, then stops
• second level monitors 2 microphones and measures
  time of arrival of sound at each
• waits for pattern of sharp noise followed by
  silence, ventures out to the direction,
  suppressing the desire to stay in dark

19
Other examples

• Herbert
• Genghis
• Toto
• Seymour

20
Future Work

• how to combine many behavior generating modules
  to be productive
• how to handle multiple sources of perceptual
  information when theres a need for fusion
• how to automate building of interaction
  interfaces between behavior modules so that
  larger systems can be built