J. R. Anderson - modelling cognitive thought at CMU (AI)
Anderson's ACT-R system is a unitary theory of cognition. The theory has origins in the human associative memory (HAM) theory of human memory (Anderson & Bower, 1973). It also borrows ideas from Newell and Simon's symbolic framework (1973). The ACT-R theory started first as an ACT production system, presented by Anderson in 1976. The ACT production system proposed a distinction between procedural knowledge and declarative knowledge. In 1983, Anderson provided a fuller description of the ACT and developed a theory called ACT*. Integrated with a set of neurally plausible assumptions about how production might be acquired, the ACT* theory was evolved into the ACT-R (Atomic Components of Thought) theory (1993), in which an architecture of cognition is modeled to explain how the process of acquisition can be tuned to the statistical structure of the environment.
In ACT-R, the current goal acts as a filter to select relevant productions. There are two long-term memory stores: a declarative and a procedural memory (Anderson & Lebiere, 1998). The knowledge in the declarative memory, i.e., facts and goals, is represented in terms of chunks. At the symbolic level, chunks are structured as a semantic network. On the other hand, the knowledge in the procedural memory is represented as production rules in forms of condition-action pairs, in which the flow of control passes from one production to another when the actions of one production create the conditions needed for another production to take place. It is these production systems that provide the basis for a unitary theory of cognition. The selected production and the current goal will influence together the retrieval of information via their connections to declarative memory. Finally, the contents of retrieved nodes are used to update the current goal according to the production's action specification. Hierarchically organized goal structures are used to represent plans of action, and to control the course of cognitive processing.
Stages of Skill Acquisition
The ACT-R keeps with Fitt's three stages (1964) in the process of skill acquisition, which are the cognitive stage, associative stage and autonomous stage. The acquisition of a cognitive skill is a progressive process from cognitive stage to the autonomous stage, which, in terms of the ACT-R theory, is the transformation from declarative knowledge to procedural knowledge. The process starts with the interpretive application of declarative knowledge in the cognitive stage. Then it proceeds to compile declarative knowledge into production rules during the associative stage. Gradually the production, a set of condition-action rules, becomes increasingly fine-tuned. During the autonomous stage, the effort required by condition-action rules continually decrease.
At the beginning of the process of skill acquisition, new information enters in declarative form. In this stage, learners learn about a set of facts relevant to the skills, such as descriptions of the procedure. The knowledge of how to carry out a procedure is declarative, as step-by-step performance statements. At this point the learners generate actions through interpretations of the verbal statements, and carefully monitor the results of the actions when they carry out each step of the procedures. The processing in this stage is conscious, deliberate, slow and requires full attention.
The major development of this stage is knowledge compilation. The compilation process is aimed to produce successful procedures in order to speed up the execution of procedures, drop the verbal rehearsal and eliminate piecemeal application. During the associative stage, we have in the process of composition and proceduralization a means of converting declarative facts into production form. Composition is the process of organizing a series of actions together into a unified production. This produces considerable speedup by composing sequences of steps into one single action. Also, once the skill is proceduralized, the new integrated production no longer requires the domain specific declarative information to be retrieved into working memory. An important consequence of proceduralization is that it reduces the load on working memory, and thus achieves a great deal of efficiency.
After a skill has been compiled into a task-specific procedure, the learning process involves an improvement in the search for the right production. In this stage, the procedure becomes more and more automated and rapid. The process underlying this stage is tuning. Three learning mechanisms serve as the basis of tuning: generalization, discrimination, and strengthening.
The basic function of the generalization process is to extract from different productions what they have in common. The generalization process produces broader production rules in their range of applicability. It facilitates the transfer of knowledge in a novel situation. On the contrary, the discrimination process produces narrow production rules. The discrimination process restricts the ranges of application of productions to the appropriate circumstances. It helps identify specific conditions and multiple variants on the conditions controlling the same action. The discrimination process facilitates the development of powerful, domain specific productions. Moreover, the specificity of the condition statements can help resolve conflicts.
In this stage, learners are also getting better at selecting appropriate production in a particular context. The criterion of selection is degree of strength. Each production has a strength that reflects the frequency with which the production has been successfully applied.
Computational processing in ACT-R
The ACT-R is not only a theory that addresses knowledge acquisition. It has also developed an explanation to the question of how people select the appropriate knowledge in a particular context (Anderson, 1995). Using the rational analysis, the ACT-R theory claim (Anderson, 1995) that the mind determines what knowledge is available according to its odds of being used in a particular context. In fact, the mind implicitly performs a Bayesian inference to calculate these odds by keeping track of general usefulness and combining this with contextual appropriateness (Anderson, 1990). The basic equation is as follows:
Activation-level = Base-level + Contextual Priming
The main implication of this equation is that the accessibility of certain information is determined by both its past use and its relevance to the current goal.
In ACT-R 4.0 production compilation, the production rule is created immediately upon popping the dependency goal - hence at zero delay. So, the need is eliminated to maintain the declarative information.
There are chunks of type DEPENDENCY and represent a person's understanding of a particular step in a problem-solving episode.
A dependency is created when a person sets a goal t understand a bit of an example or instruction
When this dependency goal is popped, a production rule is induced form the dependency and added to the production system.
Four special slots of the dependency structure: goal slot (holding what the goal was like before the problem solving step); constraints slot (holding chunks that serve as the bridges from condition to action and becoming retrieval patterns in the compiled rule); modified slots (holding what changed goal look like); stack slot (indicating any changes to the goal stack in terms of pushes and pops) Cognitivism [http://www.personal.psu.edu/]