Skip links
Main content

Agent architecture

zondag 17 mei 2009 22:05

Here is a sketch of the raw idea of the agent I am working on. I want to implement a reference architecture.


The agent aims to be an implementation of Bernard J. Baars' Global Workspace Theory of mind. In this theory the mind is like a theater. In this theater the stage corresponds to Working Memory and the audience to the unconscious processes. The audience may scream for attention, and may be allowed on stage (similar to Jackson's Pandemonium theory). This means that most brain activity is unconscious and only when it is done, or encounters a problem, becomes conscious.

The most important problem this theory solves is that of coordinating the massive parallelism of the mind with the serial interaction with the world. Also, the theory ensures that time appears like a single arrow and events appear one after the other.

Raw input is first processed by a language module into a "Speech Act", to lend a term from John Austin. The speach act is not passed to the WM directly, because the WM would not know its importance. It is first sent to the Dialog Manager, or Inner Dialog, to process it (unconsciously) and only when further action is required it is send to the Working Memory.

The "inner dialog" is not part of Baars' theory per se. It is a thought I picked up from reading about natural language processing. In this area, such a component coordinates language use. It guards the structure of the dialog the agent has with its conversation partner.

Working Memory (an active process, despite its name) selects one of the attention seeking events as the one to be in the theater's spotlight. This event is processed and enters Episodic Memory (thus ensuring the sequential nature of time). Working Memory is not very intelligent. It listens to all events, but may store just a small amount of them (say 7). The rest will just be disregarded. Based on some priority function it picks an event to pass through. Only this event will be remembered.

Unconscious modules may communicate with eachother directly. This will not be consciously experienced and can not be communicated to the outside world.


my agent

« Terug