Criticisms of symbolic

1
Criticisms of symbolic connectionist approaches

• Overly abstract
• modellers simplify environment, providing just
  the (intuitively) important features, do the hard
  work. Results in brittleness.
• Frame problem
• what needs to be included/updated in internal
  reps?
• Isolated systems
• modelling system components in isolation means
  they may not be ultimately connectable. Study of
  sub-problems.
• Inefficient, unnecessarily complex
• may not need much internal representation to
  function
• Realistic timing
• rich internal reps may not support realistic
  reaction times

2
Brooks approach to AI

• Basis for developing true intelligence
• mobility
• acute vision (sensory capabilities)
• ability to carry out survival related tasks
• Bottom-up
• Start small
• Incrementally build up capabilities of
  intelligent systems that interact with the real
  world.
• At each step make them complete/functional,
  ensuring validity
• Implemented in Mobots or Creatures
• autonomous mobile robots

3
Brooks (1991) p.139

• Representation has been a central issue in
  artificial intelligence work over the past 15
  years only because it has provided an interface
  between otherwise isolated modules and conference
  papers

4
Mobot requirements

• Cope with a dynamic unsimplified environment in a
  timely way
• Be robust against changes in environment
• Have a purpose for being do something
• Have multiple goals
• Capitalize on fortuitous circumstances

Mobot 1 Allen (a mobot that avoids hitting
things) 12 ultrasonic sonar sensors Explore
layer Looks at distant places and moves toward
them suppresses wander layer Wander layer
Generates random movement commands Avoid layer
Generates repulsive forces when detect obstacles
sums attractive and repulsive forces and
suppresses behavior
5
Subsumption architecture

• No distinction between peripheral and central
  systems
• avoid traditional functional decomposition
• Division into activity/skill layers that
• Each have a skill/goal (perception-action
  loops), avoiding need for higher level rep of
  goals
• Are largely encapsulated from each other
  (modular)
• Work in parallel, permitting fast reactions
• Suppress or inhibit other layers
• Big changes in environment may effect some
  layers, but probably not all (graceful
  degradation)

6
Activity/Skill Layers

• Contain finite state machines with
• fixed topology
• small number of states
• 1-2 registers
• 1-2 timers
• access to simple computational machines (e.g.,
  sum vectors)
• Run asynchronously send 1- or 24-bit messages
  to each other
• Change states in reaction to messages timers
• Incremental construction means lower layers never
  dependent on higher ones

7
Behavior of avoid layer

• If start in middle of room, doesnt move.
• If someone approaches, moves away.
• If other obstacles then encountered, halts.
• Exists without hitting or being hit.

8
FS machines of other layers

• Explore layer
• whenlook
• detects that robot not moving
• activates free-space finder
• inhibits wander layer
• send path to pathplan
• pathplan
• sends heading to avoid (so still avoids objects)
• integrate
• monitors actual path followed updates pathplan

• Wander layer
• wander
• output random heading every 10 seconds
• input to avoid layer where treated as attractive
  force, summed with repulsive ones
• output ignored when turn or forward is moving
  robot

9
Differences from earlier approaches

• Not connectionist
• no distributed representations, few connections
• Not production rule system
• no matching of database to preconditions of
  rules no variables to match to tokens in
  database no selection of rule based on entire
  rule base
• Unlike a blackboard system
• information is not globally shared nor
  anonymously output by layers
• Inspired by engineering rather than Heidegger

10
Potential limitations

• Max number of layers that function with
  subsumption architecture?
• Limits on complexity of behavior without central
  representations?
• Possible to learn with finite state machines?

11
Brooks (1991) p.139conclusions

• When we examine very simple level intelligence
  we find that explicit representations and models
  of the world simply get in the way. It turns out
  to be better to use the world as its own model.

12
Distributed cognition

• Model ant-like behavior in small robots
• So situated embodied too
• Mobile, have claws, communicate with infrared.

http//www.ai.mit.edu/projects/ants/
13
Software

• Use subsumption architecture for behavior.
• Move-forward, Move-to-light, Move-from-bumps
  (back away from something bumped into)
• Grouped behaviors hierarchically

14
Group behaviors II

• NOT-IT robots emit not it infrared signal.
• IT robot moves toward NOT-IT robots tries to
  bump them.
• When IT robot bumps into anything, it emits tag
  but only other robots respond.
• Tagged NOT-ITs emit I got tagged causing IT to
  become NOT-IT.

tag
15
Some applications

• Mars rover
• particularly locomotion over rocky terrain
  redundant legs
• Explosive ordanance disposal
• cooperative group behaviors

16
Social interaction

• Kismet from MIT AI lab
• Allow robot to influence its learning environment.

17
Main Points

• Complexity in behavior does not require
  complexity in the behaving agent.
• it may be in the environment.
• Cognition as intelligent behavior adaptive
  interactions between organism and environment.