Greg Detre
Thursday, November 14, 2002
why haven�t people heard of it outside its field?
because it doesn�t work
it doesn�t
give you much more power than a programming language
what are the computational tools that people use in psychology???
it�s too high-level
is it even falsifiable???
Winograd � the program is the theory
push�s derivations from soar (he feels he�s taken the best ideas)
impasse �/span> mental critics
2 phases �
brainstorming + selecting
they went the opposite way from Marvin
went for a
very simple compact way of building all the more complex/diverse approaches
it gets hairy trying to do everything with one paradigm
e.g. difficulty of doing sequential stuff with the rule-based paradigm
Push: if you
can write something in Soar, you�re probably onto what the brain�s doing � it�s
so highly granular and super-parallel � it�s kind of like programming with NNs
problems???
too much abstraction in the perception interface
Push: you can�t have multiple/specialised representations of spatial things, for instance
you can�t even have maps
you�d have to specify rules about traversal
Hugo: but that�s misusing the abstraction
is the chunking learning mechanism powerful enough???
it summarises the result of solving an impasse
Hugo: it doesn�t have an abstraction mechanism
Barbara: you don�t get the association with what doesn�t work, you lose all the learning about what failed
it doesn�t learn anything strategic about problem-solving
it can�t generate its own goals
maybe if you give it just the goal of surviving, then it can generate all the other subgoals it needs to resolve the impasses
its own mind is closed to itself
it has no direct representations, it�s all procedural
they have quite a behaviourist approach, I guess???
hugo: what does it boil down to? a belief network, adaptation + generalisation tools, k-lines network
what is/how is it like a k-line network???
summarising the net effect of all �
Push thinks they�re more flexible
they were based on psychology experiments � trying to get Soar�s timing to mirror human timing
why were they interested in problems like:
������ S E N D
������ M O R E
�� M O N E Y
you aren�t going to see most of the interesting ideas of cognition at the level of neurons
you need to abstract up a few levels
refce: Psychology @ Harvard � Stickold (sleeping, dreaming)
main competitors:
ACT-R � John Anderson
it has all these parameters so that you can tune so that it matches human performance
Newell died, so interest in Soar has waned