ACT-R

from Wikipedia, the free encyclopedia

Adaptive Control of Thought-Rational (ACT-R; JR Anderson & C. Lebiere, 1998) is one of Adaptive Control of Thought resulting cognitive architecture .

The well-known cognitive psychologist John R. Anderson (born August 27, 1947 in Vancouver) played a key role in the development . The origin of cognitive architectures lies among other things. a. in cognitive psychology. They represent computer-aided models that are used to model human cognitive processes (memory, language, perception, problem solving, etc.), to evaluate psychological experiments and theories and to simulate them on the computer. Production systems are a fundamental component of the cognitive architectures. These consist of a declarative knowledge base, the facts, and a procedural memory, the rules.

There are now a number of cognitive architectures, including ACT-R (Adaptive Control of Thought-Rational), SOAR and 3CAPS , all of which are based on different objectives. All cognitive architectures are based on a number of empirical results, for example on processing time. Different emphases are placed on different aspects of human cognition .

Cognitive processes

Cognition is a general term for all forms of knowledge. Cognition includes both content and processes. The contents of cognition relate to what one knows - terms, facts, statements, rules and memory contents : "A dog is a mammal"; "Red light means to stop"; "I left home when I was 18". Cognitive processes refer to how you manipulate this mental content - so that you can interpret the world around you and find creative solutions to deal with life's dilemmas. Information from the environment is recorded and processed with the senses and then this information is coded into words, symbols or signs and analyzed, interpreted and sorted through experience. From this, conclusions and judgments are formed which trigger goal-oriented behavior.

Procedural and declarative memory

The ACT-R belongs to the group of production control systems . It differentiates between declarative and procedural memory. The declarative memory includes the so-called factual knowledge, while the procedural memory includes the knowledge of courses of action. The philosopher Gilbert Ryle also made a distinction in this context between the terms knowing that and knowing how . The latter includes knowledge of the way in which individual actions are carried out. This includes, for example, knowing how to ride a bike or how to dress. How individual actions are carried out is often difficult to convey verbally. In addition, a person is usually not able to perform an action perfectly if he has never done it before and only receives a verbal description of the course of action. Instructions are therefore usually not enough to acquire procedural knowledge. It has to be learned through practice. However, courses of action can be automated through training.

Declarative memory can also be divided into semantic and episodic . Semantic memory consists of factual knowledge, i.e. H. from everything learned and remembered that a person can communicate. It includes, for example, terms, statements or rules such as “A dog is a mammal” or “Red light means stopping”. This kind of knowledge is also known as world knowledge. However, one speaks of semantic memory because it contains the relationship between individual words and what they mean (semantics). In addition to semantic memory, there is also episodic memory. It includes the memories of everyday experiences and personal experiences. This includes, for example, the first day of school or remembering where you last parked your car.

Declarative knowledge is stored in the ACT-R in the form of so-called chunks . Procedural knowledge, on the other hand, is represented in the form of productions . These are rules that are used to process the chunks. In order to ride a bike, for example, a rule must exist that describes exactly how the cycle of action must be carried out. On the other hand, there must also be factual knowledge (chunks) to which these rules can be applied. For example, it must be known that a wheel is a means of transport that is generally made up of several individual parts. The production rule, on the other hand, specifies which of these parts must be operated in which way in order to move on the bike. The more often you have ridden a bike (part of the episodic memory), the more reliable this production rule is. Factual knowledge and production rules are therefore not independent of one another, but influence one another.

The role of working memory

Anderson (1993) suggests that working memory has an impact on memory and retrieval in information retrieval. He carried out a specific experiment: the test subjects were asked to memorize a sequence of digit combinations while they were solving a mathematical equation. You should then reproduce the sequence of digits. The two conditions (remembering the sequence of numbers and solving the mathematical equation) increase in their degree of difficulty. The size of the number combination to be remembered and the degree of difficulty of the mathematical equation are manipulated. As the difficulty of both conditions increases, there is a significant drop in performance in both tasks, as the results of the experiment show. The majority of errors are due to incorrect retrieval of information from memory. The ACT-R theory now assumes that there is a limited activation of resources. With the increasing degree of difficulty, there is very little activation potential left for the retrieval of the relevant information from the declarative memory. The incorrect retrieval of the information by the test persons could mean in the ACT-R theory that there is only a partial adjustment of the information in the memory.

Chunks in ACT-R

It is assumed that our memory divides larger amounts of information into subgroups (" chunks ") so that these can be represented "more easily" and thus take up less storage capacity in our memory .

A chunk is defined by its type and its slots . The type stands for the category to which the chunk belongs. Slots correspond to the attributes of a category. Each chunk has a unique name with which it can be referenced. The strength of a chunk affects how well and how quickly it can be recalled from memory. The more often certain chunks are used or the shorter it was last used, the stronger this chunk. These chunks are stored in declarative memory. Production rules that are used to process the chunks are stored in the procedural memory. In the ACT-R approach, chunks are coded in their own language. This has similarities with programming languages ​​such as Java or C ++. Information can be saved in this form and also verbalized again. As a prototype, chunks look like this:
(CHUNK type NAME SLOT1 SLOT2 ... SLOTN)

Productions and production systems

The building blocks of procedural knowledge are productions. The productions have a condition part and an action part. In the condition part, reference is made to the contents of working memory. If the constellation formulated in the condition part is in the working memory, the action part is carried out.

A system that controls the application of productions is called a production system. The data currently in the main memory can be, for example, the description of a problem state and the stored rules can be problem-solving operators. The rule application is controlled by an interpreter . First, output data are brought into the main memory via an input interface. After that, rules are applied to this data in the main memory until either no rule can be used any more or a rule is applied whose action is a command to stop processing. The current data is then output as a result. The transformation of the data through the application of rules takes place in the so-called match-select-apply cycle :

  • Pattern comparison (match ): a search is made for all production rules whose condition part is compatible with the data.
  • Selection (select): one of these rules is selected according to a predefined conflict resolution strategy.
  • Application (apply): the rule is applied to the data in memory.

A rule is used in each cycle that changes the data in the main memory.

The rules must be defined beforehand by the scientist, the modeler, according to a specific syntax (based on the LISP programming language ). Another part of the ACT-R cognitive architecture consists of the so-called "buffers". This is used to model the resource limitation in central sensory processing. It is a principle that only one “ chunk ” can ever be included in a buffer (for example in the “visual buffer”) .

In ACT-R, Anderson's now relatively undisputed theory of “Spreading Activation” is implemented, in which activation in declarative memory can spread over similar units of knowledge. This phenomenon explains the results in so-called "priming experiments" ex. 'Perception Study' / 'Reading Comprehension Study' by Higgins, Rholes & Jones (1977). (Priming is a pre-activation of memory content by means of appropriate stimuli that are associatively linked to a target content).

In contrast to other production control systems, which generally contrast with connectionist models ("networks"), ACT-R can fall back on a learning mechanism that is based on probabilities .

Activation processes in the ACT-R theory

The basic process according to the ACT-R theory according to Anderson is that a "production rule" is generated in relation to knowledge and according to this certain declarative knowledge is found (queried) in order to solve a specific problem can. The speed or the success of this process depends on how high the activation of the corresponding chunks is and how strong the activation of the "production rule" is. It is precisely this process that makes up the underlying fluidity of our behavior, our performance. Of course, against the background of constantly activated knowledge, a liquid is now difficult. First, in many situations it is unacceptable to show incorrect and slow behavior or performance. Second, the development of advanced competencies requires fluidity / harmony in the behavior or performance of the basic competencies. The ACT-R theory is a mathematically based theory of how strength and activation affect our behavior and ultimately our performance.

Thus, in a first process of parallel activation, chunks and productions are recognized or activated with the highest probability of being needed in the given context and these knowledge structures (chunks and productions) then in turn decide on further steps in which the activated knowledge is applied. This means that knowledge is made available and activated depending on the likelihood of being needed in a certain context.

The extent of activation can be derived using Bayes' theorem . The basic assumption here is that the extent of activation can be calculated from the basic activation and the context-related priming. In other words: activation = basic activation + contextual priming. If you compare this with Bayes' theorem, then:

Activation = posterior probability

Basic activation = a priori probability

contextual priming = likelihood quotient

Schooler (1993) showed that human memory seeks information based on Bayes' theory and then makes the required knowledge available as a function of posteriori probability.

Using the example of a chunk i (see chunks in ACT-R), the activation of which is described as a function of the various elements associated with it and its basic activation, the activation equation for the current activation of chunk i is :

...... basic activation of Chunk i

..... weighting of the contextual chunk j

..... strength of association between chunk i and j

Knowledge generation

New chunks can arise when productions are applied. In turn, productions are created by encoding chunks. In order to avoid circularity in theory, a second, independent source for the creation of chunks was defined in the ACT-R. That source is the environment. The visual perception system, which is responsible for encoding information, plays an important role here. It is assumed that what is visually perceived is not saved as a whole, but divided into individual objects. The specific characteristics of each object are also saved. However, not every perceived information is encoded and stored, only that which is in the center of interest. Attention processes are also essential when recognizing what has already been saved. Features of individual objects can only be recognized as the entire object if attention is directed to them. Productions are also made taking the environment into account. It is assumed that problem solving takes place by searching for existing solutions from other contexts and using them to solve one's own problem. Problem solving thus takes place through the formation of analogies.

ACT-R and visual attention

To apply the ACT-R model to visual stimuli, Anderson et al. (1997) generated a module for visual attention. It's about how ACT-R finds and extracts information from iconic memory (see illustration). The information or visual elements that are stored in the iconic memory consist of (visual) features and ACT-R can draw attention to certain contents of this intermediate memory. With the help of attention, a chunk is then formed from these features that the ACT-R can work with.

There are three types of information ACT-R can use to attract attention:

(a) certain places and directions,

(b) certain characteristics and

(c) Objects that have not yet received any attention.

history

ACT-R is the basis for many theories about models of human cognition. ACT-R has its roots in the early 1980s in the HAM model (Human Associative Memory), which is a model of memory and was postulated in 1973 by John R. Anderson and Gordon Bower. This model was later developed into the first ACT (Adaptive Control of Thought) model. Originally ACT was called “Active Control of Thought” and was a theory to explain cognitive performance of humans. On the one hand, it wants to explain how people structure their knowledge, and on the other hand, why people are able to show intelligent behavior. Anderson's theory also has its origins in research on artificial intelligence. According to Anderson, the name "homo sapiens" already says something about the fact that humans have their own "intelligence" that cannot be found in other species. To test the understanding of human intelligence, he developed, among other things, computers that can independently write recursive programs. This should provide the cognitive sciences and artificial intelligence research with a model that enables computer simulations of a person's mental performance. In the ACT model, procedural memory was included in the theory for the first time. This first theory was later developed into the ACT-R model of human cognition.

In the late 1980s, Anderson developed a mathematical model for cognition that he called rational analysis. The basic assumption of this theory is that cognition is optimally adaptable and that the precise assessments of cognitive functions reflect the statistical probabilities from the environment. This theory was later incorporated into the ACT-R theory for underlying calculations. Hence the “-R” in the name of the theory, which stands for “Rational”.

In 1998 John R. Anderson postulated a further development of the model, the ACT-R-5.0 model. Anderson did an in-depth understanding of the underlying processes in the human brain and used brain imaging methods to learn more about them. The need to localize the underlying processes led to a further development of the model, which was presented in 2005 as ACT-R 6.0. The coding language in this model has also been significantly improved. ACT-R 7.0 followed in 2015, but only brought about small changes. The long development of the ACT-R model led to numerous similar studies and projects. The most important of these are the PUPS production model and the ACT-RN model.

Applications of the ACT-R model

The ACT-R is now used in many different areas of activity. Common software based on ACT-R are "ACT-R 6" and "Cogtool". This software can be used to create prototypical user interfaces (UI) with the aim of predicting the behavior of users. This is done on the basis of the ACT-R model. CogTool simulates the cognitive, perceptual and motor behavior of people when dealing with the generated prototype in the course of processing certain specified tasks.

A modified model of the ACT-R (ACT-R / PM) plays a role in predicting how much time is needed to complete a task, particularly in human-computer interaction (HCI). Furthermore, the ACT-R model is used to describe how people can learn or manage cognitive tasks. It also has a prescriptive character, i. H. it helps to design tutorial programs by modeling learning processes. These “tutorials” assist users in using computer programs or other cognitive skills.

For example, the ACT-R model is currently used successfully in schools, where it can help students learn math problems.

The ACT-R model is successfully used not only in mathematics, but also in language research to decipher aspects of the mother tongue in terms of understanding and production.

When researching neuronal mechanisms in the brain, the ACT-R model was used to predict patterns of brain activation. It was found that four modules of the ACT-R are associated with four brain regions that are active during complex task management. On the one hand, activity was found in lateral inferior prefrontal regions, for example, which can be equated with the retrieval of information from declarative memory. Furthermore, the activity in the caudate nucleus is reflected in executive acts of procedural memory.

Web links

Individual evidence

  1. ↑ Anderson, JR, Bothell, D., Byrne, MD, Douglass, S., Lebiere, C., & Qin, Y. (2004). An integrated theory of the mind. Psychological Review, 111, 1036-1060.
  2. ↑ a b Gerrig, R., & Zimbardo, P. (2008). Psychology (18th ed.). Munich: Pearson studies.
  3. ↑ cognition . Lexicon online. Retrieved April 8, 2014.
  4. ↑ Oberauer, K., Mayr, U., & Kluwe, ER (2006). Memory and knowledge. In H. Spada (Ed.), Textbook General Psychology (3rd ed., Pp. 115-195). Bern: Verlag Hans Huber.
  5. ↑ Anderson, Reder, Lebiere (1996): Working Memory: Activation Limitations on Retrieval. In: Cognitive Psychology (30). 221-256.
  6. ^ Sven Brüssow, Daniel Holt: Introduction to Cognitive Modeling with ACT-R (PDF) psychologie.uni-heidelberg.de. October 24, 2007. Retrieved April 8, 2014.
  7. ↑ a b Wolfgang Schoppek: Mental arithmetic from the perspective of the ACT - R theory (PDF) uni-saarland.de. Retrieved April 8, 2014.
  8. ↑ Müssler, J. & Prinz, W. (2002). General psychology . Heidelberg: spectrum. 715-733.
  9. ↑ Anderson, JR, & Schunn, CD (2000): Implications of the ACT-R Learning Theory: No Magic Bullets. In R. Glaser (Ed.), Advances in instructional psychology (5th ed., Pp. 1-34). Mahwah, NJ: Erlbaum.
  10. ^ Glaser, R. (Ed.) (2005). Advances in instructional psychology (5th ed.) . Mahwah, NJ: Erlbaum.
  11. ↑ a b J. R. Anderson: A Simple Theory of Complex Cognition , 1996 In : American Psychologist , 51, 355-365.
  12. ↑ Anderson, JR, Matessa, M., & Lebiere, C. (1997). ACT-R: A Theory of Higher Level Cognition and its Relation to Visual Attention. Human-Computer Interaction, 12, 439-462.
  13. ↑ ACT-R 6 Official Website
  14. ↑ Cogtool Official Website
  15. ↑ CogTool User Guide - Version 1.2 - May 23, 2012 Accessed April 10, 2014
  16. ↑ Anderson, JR, Fincham, JM, Qin, Y., & Stocco, A. (2008). A Central circuit of the mind. Trends in Cognitive Science. 12 (4), 136-143.