task 5 Act-R
Act-R
--> has serial processing ( searching problem space)
---> and parallel processing ( spreded actication in chuncs and looking fo rrule thats best for the constraints / fits best the chunks activated )
Rule based
rule = if then statement
contradictory rules can coexist with each other :3!!
--> so one rule not absolute / universal
--> e.g. if x = student then overworked (as a default rule)
--> e.g. if x = student and x= took easy courses THEN x= not overworked (as additional rule :3 !!)
--> these seem contradictory but they can easily coexist so IF then rules are not ultimately one way/ absolute but rather flexible
search of problem space (rules) = main method of computation / problem solving in rule based!!
(in logic base it would be deduction form logic rules)
heuristics
--> for complex problems searching the whole problem / rule space = impossible
------> so they rely on heuristics
analogy to human brain:
--> many rules in long term memory BUT
--> only few accessible at a given time cause they are more accessible in short term memory thus through heuristics these get choosen more #
conscious = serial mostly but also depends on parallel processing / many rules applied at once
unconsicous = PARALLEL PROCESSING especially if chuncks activate other chunks like manyo of them and rulesbecome active :3!!
Parallel search / processing = unconscious
Problem solving
searching problem space for possible solutions to determine substeps necessary to solve the problem
--> can happen forward, backward or bidirectional :3 !!
deciding between alternative substeps
--> cant really be rule based / not ideal so needs other supplementary process :3 !!
--> utility approach (expected-value calculation) the node with most utility is chosen
or deliberative coherence determination
Explanation is based on rule based reasoning
--> only possiblel if chain of rules exist that allow to generate explanation youre looking for
learning = acquisition, modification, application of rules !!
--> e.g. chomsky's language device which as he proposed operates based on innate rules like syntax etc ..
------> children overgeneralizing "-ed" if they want to say something in past tense = abductive learning cause they reverse the rule e.g. i want to say in past tense that someone was bringing me something but i only know if pasttense i shoul dput “-ed” hence if wanna say something past tense i need to put “-ed” on it !!
Cognitive modelling
2 modelling types
prallell distributed networks
--> as in neural network models (connectionism / synapses etc)
--> trying to apply how neurons might work in brain to computers
Rule based models
--> like ACT-R and SOAR
--> less regard to neuron like behavior in brain
ACT-R
--> by Anderson
3 components
--> interaction between those leads to problem solving, learning, planning and understanding
Declerative memory
--> things that we know and were conscious of at a given moment
--> consists of chunks with 2 to 6 elements per chunk
--> Hippocampus
Procedual memory
--> things that we know how to do at any given time
-----> e.g. like tying your shoe laces :3 !!
--> basal ganglia / PFC
Goal module (most up to date model)
--> represents what ACT-R currently wants to do so the actual goal :3 !! thsi is used as reference point every step along the way and possible nodes in problem spaced get compared to this :3 !!
Connected by:
--> retrival request: production rule in procedual memory needs retrival from declerative memory
--> production complication: new production rules can be created in procedural memory from ‘chunks’ in declarative memory. This is especially important in ACT-R models of learning, where new rules are formed as learners become more skilled at solving problems.
.
ACT-R also has a range of very simple perceptual and motor components, that allow it to model visual stimuli, auditory stimuli, and actions. These components are not directly linked to production memory, but, like declarative and goal modules, are connected through small shared buffers. Only the buffer contents affect production.
if part = condition
then part = action
product = solution
Chunks
chunks are connectect if elements of one chunk also present in another chunk
activation
--> = the more elements in a chunk match the stimulus or rule or somethign the stronger it activates
--> the stronger active a chunk the more salient in becomes !!
strong active chunks are more salient = same as short term memory (salient at forefront of attention chunk / rules)
--> important for heuristics
weak active chunks = less salient = same as long term memory
activation = NOT constant !!
--> if nothing happens activation slowly decreases until chuk in LTM and not acesible until reactivated :p
---> but if activated again = increases proportional
-------> explains priming effect (if chunk was active and then cues hortly later its more likely to activate again)
activation also applies here the more activated the more likely to be used ,
--> the less activation the less likely its gonna be used even if applicable
learning in 3 stages:
understanding
production compilation
--> new production rules can be created in procedural memory from ‘chunks’ in declarative memory.
practice
Learning and the power law !! add form baes notes
downside of rule based:
—>rule based system less not as representational as logic based systems
upside:
—> they have higher computational power
—> rule are not abosulte / mutually exclusive as logic rules are.
——> in other words contradictory rules can coexist with each other :3 !! (the more applicable rule is applied on per case basis!)
--> rules can represent goal states !!
---> all thsi enhances computational power AAAND psychological plausibility
Plausibility
Psychologically + computational
--> rule based systems out of all models has the best psychological plausibility :3 !! (so good cognitive model )
--> many examples of how rule based system can describe human behavior !
Power law
--> learning at first fast then slower
--> SOAR can explain cause first chunks can be build pretty quickly, but then at some point they become so specialized, that they are rarely applied, simply cause the situations happen rarely and performance of learning seemingly slows down a bit :3 !!
--> also cause if less used then chunks = less active and thus less salient and less accessible :3 !!
Neurologically
--> rule based systems like ACT-R not only cognitive model But also neurological model :3 !!
Conditioning
--> rules activated if they hear a tone (activation of chunks which spread activation to other chunks etc and response elicited)
Tick tack go or any other game lol
Human language !! (anderson)
IF the goal is to communicate a meaning structure of the form (relation, agent, object), THEN set as subgoals
- to describe agent
- to describe relation
- to describe object.
basic relation of rules on neurons
--> if input strong enough then fire
ACT-R Brain correlates
--> production rules (just rules cause production is same as rules) = applied / used by basal ganglia !
--> stored in PFC buffer :3 !!
KNOW ALL THESE THINGS from course manual
how does ACT-R enable to :
acquire knowledge
—> inductive generalisation (from other rules)
—> chunking (soar) /compisition (ACT-R) ruleis formed by other rules / generalizing the rule
—> specialisation (rule is adapted to special circumstance
—> abductive learning = rule is reversed to find possible explanation e.g. why a person is sad (getting bad grades = sad ; person sad = mabye got bad grades :P lol)
—> slow incremental learning= the better a rule works / the more succesful it is the higher the likelyhood its getting used again infuture , cause more active thus saliance in declertive memory increases !! (high activity = high saliance = short term / declerative memory ) (low activity = low salience = long term autobiografical memory)
solve problems
—> search problem space forward backward bidirectional and create subgoals based on which rule will get them to their goal fastest
find a solution more quickly than under other circumstances
—> chunks in declarative memory more active for that situation or instance or maybe was pre-activated / priming !!
why people make mistakes :3 !
—> in complex problems HEURISTICS !!
—> HOW does ACT R work for reals AAAAND how does sniff ACT work in example / ereader below !!
Current goal
--> goal that is the focus of interest at the moment
Goal stack
--> where other goals are stored that we are currently not working on but may in the future
Push and Pop
Push = from current goal to stack
Pop = from goal stack to current goal :3 !!
History
HAM 1973
--> declerative memory = collection of chunks/declerative memory elemts WHICH contain between 2 and 4 elements
--> inductive generalisation = rule is created through summary of many experiences/ examples
--> chunking (SOAR)/composition(ACT old one) = rule is formed by other rules (or many rules/options chunked together into a more general one!) generalizing the rule !!
--> specialization, in which an existing rule is modified to deal with a specific situation
--> abductive learning, in which a rule is run backwards to provide a possible explanation of what happened.
--> slow incremental learning = the more successful a rule operates the more like to be used again in future :3 !!
symbolic vs subsymboic
Symbolic
--> something that has actual representation in the real world
Subsymbolic
--> something that has no real physical representation in the real world :3 !!
important for priming effects cause of chunk activation and spread activation !!
Connectionism has the best neurological plausability though !!
if part = condition
then part = action
product = solution