Task 3-cognitive modeling, production rules, ACT-R Flashcards
Retrieval request
production rule fired in procedural memory may require elements from declarative memory
Production compilation
new production rules can be created in procedural memory from chunks in declarative memory
Declarative memory
Declarative memory is collection of chunks – contain number of elements, usually between two and four → Chunks also have the “isa” slot – says what kind of chunk it is
• Chunks are not isolated but linked to each other through their values in a kind of network and they influence each other by Spreading Activation
• Each chunk has a level of activation (=kind of energy attached to it) and activation in one chunk tends to leak out and add to all the other chunks it is connected to through its values
• If a chunk has a lot of activation, it will be easy to find and to retrieve quickly from memory
• Conversely, it will be harder to find if it has only very low activation
• Highly active chunks therefore look a bit like STM
• Less active chunks look a bit like LTM
• Activation is not constant for a chunk and if nothing happens, it slowly loses activation → But if used, it gains activation and the more it is used the more it gains
procedural memory
• Procedural memory stores procedures in form of production rules – have a condition (i.e. IF) and action (i.e. THEN)
Rules tend to be written in cryptic form
• All specified production rules are available to be used depending on current state
• relation between these representations, the condition part of the rule (the IF part) can be found in the rows before the ‘==>’ character, and the action part (the THEN part) appears after it
The bits that begin ‘=something>’ are references to chunks, and each line that follows is a slot name followed by a value
When a word begins with the ‘= ’ character, it is a variable, that is, it can change each time the rule is used
Variables are very important to production rules, as they are what make rules sufficiently general to run with different problems.
• All specified production rules are available to be used depending on the current state
• For a production rule to be used, its condition part (IF) must match the current goal and use chunks that are already available in declarative memory
Gives control over timing
For example, the greater the activation of a chunk in declarative memory, the faster productions that use that chunk can be fired ACT-R uses this technique extensively to give reasonably accurate models of human response times
The more rule activated, the more likely…
The more rule activated, the more likely it is to be retrieved + used again in future
Operation of Actr
instructions for how to perform a task start out as chunks and then, as performance improves through practice, they are turned into rules in procedural memory, which do the same thing as the chunks but do it automatically
• To complete the circle, as the rules are used, they may themselves create new chunks in declarative memory
ACT-R criticized for
- ACT-R goal stack can store and perfectly recall arbitrary number of goals
- Humans cannot do that ACT-R criticized for its lack of psychological plausibility
Chunks
- Chunks are used to represent a list as set of groups (= can contain minimally 2 and maximally 6 items)
- Example: people mentally group telephone numbers
- Organizing list into groups increases number of items that can be remembered
- Chunk can hold a group of items or just single item
- Each chunk is represented using Slots and Values and a chunk associated with individual elements has slots for the group to which it belongs, its position within the group, the overall list to which it belongs and its content
- A chunk encoding a group has slots to indicate the list to which it belongs, the number of elements in the group and the position of the group within the list
- Chunks organize a seemingly flat list or elements into a Hierarchy, where the list is initially broken down into three groups and the groups are broken into individual elements
production rule 2
• condition part of the production rule may specify what must be the current goal, and may specify what chunks have to be made available in declarative memory
• action part of the production rule either transforms goals held in memory
or performs an action to the world outside the model, such as typing a retrieved word onto a computer screen
Goal tranformation
can be of three types: A goal can be modified, created (i.e. “pushed” on top of the goal stack) or removed (i.e. “popped” from the goal stack)
base level activation of a chunk
of a chunk depends on the number of times it has been rehearsed and amount of time since it was last rehearsed
Help account for primacy + recency effect:
Primacy effect is due to number of times the item has been rehearsed increasing base-level activation
Recency effect is due to small time lapse since last rehearsal of item increasing base-level activation
Association strength
Association strength – the strength of bond between an item and the required chunk and influences the flow of activation between chunks
• Strength of association between item + chunk depends on total number of associations that items has
fan effect
The greater the number of facts related to some concept that a subject has to memorize, the slower the subject will be to recall any one of them
• Activation is limited resource:
If an item is only associated with one chunk, then this chunk receives the full associative strength of the item and the full effect of any activation
If the item is associated with three chunks, then the association strength is split three ways and less activation will flow to each individual chunk
partial matching
Partial matching – possible to select a chunk that only partially matches the item sought in the condition part of the production rule
condition action pair
• Conduction-action pair:
Condition specifies what must be true for production rule to apply test for state of current goal and chunks in declarative memory
Action specifies set of things to do if production applies can change goal state
conflict resolution process
conflict resolution process (=chooses one production from among productions that match current goal)
Prototype Model
Learner estimates the central tendency from all the examples experienced from within each category during training
New stimuli are compared to the prototypes of different categories and the category with the most similar prototype is chosen
Exemplar Model
Memorize all examples that are experienced from each category
Category with the greatest total similarity is chosen
Cognitive modelling
is an area of computer science that deals with simulating human problem solving and mental task processes in a computerized model
The two hallmarks of cognitive models are that they are described in mathematical or Computer Languages and that they are derived from basic principles of cognition
cognitive model advantages
• By using mathematical language, they are guaranteed to produce Logically Valid Predictions
- Generalizability-The advantage of cognitive models over generic statistical models and empirical curve fitting is it Generalizability that can be used to derive new predictions
- The difference between cognitive models and neural models is that the latter describe actual neural substrates and neural interconnections that implement cognitive processes, and so they are better for inferring e.g. predictions for patterns of fMRI images
- Both types of models serve an important but somewhat different goal with respect to measures they try to predict but both should be bridged to relate the two types of models
What are the practical uses of cognitive models
- Clinical use: assess individual differences in cognitive processing between normal individuals and clinical patients
- Cognitive neuroscience: understand psychological functions of different brain regions
- Aging process and deterioration of cognitive functions with age
- Improve human-machine interactions
- Decision research predict outcomes
What are the steps involved in cognitive modelling?
- Take conceptual framework and reformulate its assumptions into more demanding mathematical/computer language description
need to formulate the prototype for each category as a vector of features and write down formulas describing how to compute the distance between the target stimulus vectors and the prototype vectors
first step uses the basic cognitive principles of the conceptual theory to construct the model. - Make additional detailed assumptions (ad hoc assumptions) to complete the model because conceptual theories are insufficient
- Estimate parameters from some of the observed data (e.g. weight parameters)
- Compare predictions of competing models with respect to their ability to explain empirical results design experimental conditions
- Start all over and reformulate theoretical framework and construct new models in light of feedback obtained from new experimental results
rule based models
• Rules – If-Then structures and similar to conditionals but have different representational + computational properties
• Rules are also called productions
• Rule-based models have psychological aims
• Rule-based systems:
Logical Theorist (Newell, Shaw, Simon) making proofs in formal logic
GPS (Newell & Simon) simulate human solutions to various problems
ACT system (Anderson)
SOAR (Newell)
