Greg Detre
Wednesday, April 16, 2003
synthetic
char � autonomous agent with some stage-presence
emotional intelligibility + appeal too
Reynolds
(1987) boids � behavioural animation
Tu + Terzopoulous
(1994) � undersea
Blumberg +
Gaylean (1995) � Silas, Blumberg PhD project
directability, multiple levels of control (e.g. at a motor vs emotional
level)
Syn chars
group
draw in particular from ethology
reference:
Clutton-Brock, Origins of the Dog, 1995 � first domesticated 14000 years ago
quotes from
the Homer passage when Odysseus is recognised by Argos(sp???)
Dobie the
coyote
�do the right thing at the right time� � maximising reward
recognise imporant new contexts
synthesise new actions (via luring)
trainer-friendly
credit assignment
percepts =
discrete classifiers
action
tuples
trigger (when?)
object
action
do until (for how long?)
wasn�t there a fifth though???
temporal
windowing � animals only really pay attention to two windows of time
the cue window
consequences window
from
Dobie�s perspective, causes consist of things that I have done
the verbal
cue is a cause in the sense that it�s a releaser for the behaviour
this is
operant conditioning � where the cause is self-actioned
I don�t understand the distinction from classical conditioning
credit goes
to tuple with:
same action
active trigger percept
most novel + specific trigger
convex hull
� cross-product of the decomposed pose space into separate limb poses
one of the
things about clicker training is that you keep your hands off the dog �
otherwise you stress the dog, or you distract it, or it�s not intiating the
movements itself and learns slower
e.g. if you push the dog�s butt down to teach it to sit
Deb: considering whether you could add the code to Ripley, if you allow
the user to use a mouse to drag the robot arm around manually
hard to build a 3D interface for manipulating� the dog in software, vs a robot
in the
future: more complex
more than one action
more subtle/ambiguous feedback
what and who to learn from?
no scripted
hand movements
humans use
operant conditioning a lot
we often use the verbal (say) equivalent of clicker training
and there�s even some languages where a �click� sound is used as a sign
of disapproval
Killeen �
Arizona state university in behavioural modelling
timing, animal modelling, addicted to meta-thought, addicted to science
major researcher in behaviourist circles
he chaired
the conference of Society of Quantitative Behaviourists
Garcia �
noticed that rats can often have an aversive reaction to food that makes them
ill, hours later � it took him ages to get it published
Chomsky �
Skinnerian behaviourism can�t capture long-term dependencies and recursion,
because Markovian models are stateless (once you�ve got this far, you can
ignore what came before)
description
vs explanation
well, if you can describe it in terms of the lower level, haven�t you
explained it