Agre and Chapman:
"Before and beneath any activity of plan following,
life is a continual improvisation, a matter of deciding what to do now based
on how the world is now. Our empirical and theoretical studies of activity
have led us to question the supposition that action derives from the execution of
plans and the corresponding framework of problem solving and reasoning with representations."
Pengi plays Pengo
"Pengo is played on a 2-d maze made of unit-sized
ice blocks. The player navigates a penguin around in this field with a joystick.
Bees chase the penguin and kill him if they get close enough. The penguin and bees
can modify the maze by kicking ice blocks to make them slide. If a block slides into
a bee or penguin, it dies."
Action based on Interactive Routines
Routines are not plans or procedures
* compute direction opposite from nearest bee
* run in that direction until you hit a wall
* kick through wall
Repeat until hit something
R1: when you are being chased, run away
R2: if you run into a wall, kick through it
The rules don't represent the iteration; the loop
emerges as a result of the interaction of the rules with the simulation.
Deictic Representations (see, e.g., Ballard)
Instead of representations like:
(at block-213 427 991)
(isa block-213 block)
(next-to block-213 bee-23)
Use situation-specific and agent-centered representations like:
This avoids variables altogether! It also
avoids the combinatorics involved in binding variables to different sets of individuals.
Other features of the Situated Activity approach
Simple Machinery--often based of finite automata. See Brooks on the subsumption architecture.
Visual Routines--perceptual routines analogous to the action routines described above. See work by Ballard.
Action Arbitration--required when multiple routines produce conflicting actions. See Maes on action arbitration via spreading activation networks.