General Production System

Production systems consist of two primary memory stores, a working memory that contains transient information and a production memory that contains rules that guide behaviour. Processing operates in two phrases. In the recognise phase, the system selects a rule whose conditions are met by the contents of the working memory. In the act phase the system applies the selected rule, modifying the working memory and/or generating external behaviour.

Match Memory buffer is required because, at any stage of processing, the conditions of multiple production rules may be satisfied by the contents of Working Memory.  In such cases, only one rule should be fired.

A production system's behaviour on any particular task is determined by the contents of Working Memory and Production Memory.Working Memory should initially contain a representation of the information on which the system is required to act.

We have seen that the behaviour of processes in Cogent follows from condition-action rules built into the model. These can be taken as an agent's intrinsic cognitive abilities.

They are distinct from the production rules held as data in Production memory. These can be taken as the particular cognitive skill being learned by the agent.

Here, we have to explicitly model refraction in order to avoid the same rule instantiation firing more than once.

Recall that ACT-R is also a general production system. Not surprisingly, the architectures of the two systems are similar. Working memory: similar to ACT-R Goal module.Production memory: combines ACT-R Central Production System and Declarative module.Match memory: Similar to ACT-R Retrieval buffer.Conflict resolution: Here, based on recency. ACT-R: based on subsymbolic activation levels.

ACT-R uses an activation based mechanism to ensure that only one production rule fires at a time. Production instances accumulate activation over time, and the instance that accumlates sufficient activation first is the one that fires. In soar, by contrast, multiple production instances may fire in parallel, but production firing does not directly change working memory. Instead it generates "preferences" for working memory changes, and these preferences are then evaluated, yielding a single most preferred working memory change.

Architecture

The Initial Contents of Buffers

Two 'levels' of production rules

Comparison to ACT-R