From Cognitive Architecture Wiki
This document is provided as an introduction to new readers and users of the Cognitive Architectures wiki. The primary purpose of the document is to provide a quick and gentle introduction to a number of different cognitive architectures. Of interest to more experienced readers will be the contextual links used to connect related ideas within and between architectures. We use the term cognitive architecture more generally than is often used the literature of Cognitive Science and Artificial Intelligence. In particular, many of the architectures introduced in the document makes no claim about any particular neuroscientific or psychological validity. Thus, the architectures are more generally agent architectures than cognitively-plausible ones.
This document was developed from a number of similar documents that were written as part of the course requirements for EECS 547, Cognitive Architectures, at the University of Michigan (Department of Electrical Engineering and Computer Science). The authors of both the original documents and this document are graduate and undergraduate students at the University of Michigan.
For those new to both Cognitive Architectures and Artificial Intelligence
For readers unfamiliar with cognitive architectures and artificial intelligence (AI) in general, a good place to begin examining the document is the theory section. This area of the document presents a number of different theories and ideas from AI and cognitive science independent of the cognitive architectures themselves. Although the information there is not organized in a coherent argument, the information does present some motivating factors for the study of cognitive architectures, especially as an experimental endeavor.
Since the reader may not be overly interested in the contextual links which are used to relate different facets of architectures, architectures may be read linearly using the "NEXT" labels at the bottom of architecture nodes. In general, there will be a non-contextual link at the bottom of each node in a document which leads to a definition of that term. For instance, if you were reading the frame-like representations node in the Theo architecture, there is a link there with the label "Go to a discussion of this property for multiple architectures" which would take the reader to a node discussing the concept of frames, including both a definition of frames and their relationship to cognitive architectures. Thus defintions of terms specific to cognitive architectures (and AI) may be readily accessed and when a definition is not needed, reading may continue uninterrupted by the intrusion of the definition.
For those familiar with concepts of Artificial Intelligence but not familiar with Cognitive Architectures
In general, the cognitive architectures introduced in this document are described in terms of typical artificial intelligence concepts; e.g., representations, methods for control, and capabilities such as planning, learning, and robotic action. For a reader already familiar with such topics, one way to quickly learn about cognitive architectures is to enter the document from the descriptions of the common properties or capabilities which correspond to these familiar topics. These descriptions include not only a definition of the term as it is used in the document (which will probably be of only marginal interest to the reader) but also entry points into architectures demonstrating that particular feature. For instance, the property frame-like representations includes a definition of frames and points out that three architectures -- Theo, MAX and Homer -- use frames as a basic mechanism for representing knowledge. By reading these three nodes from the original frame node, the differences and similarities in the use of frames in these different architectures may be examined. Additionally, other facets of an architecture may interest the reader, bringing she/he into an architecture from a particular point of interest. For example, in the discussion of frames in Theo, there is a contextual link to Theo's impasse-drive control, demonstrating how the use of frames leads to this control paradigm.
For those already familiar with Cognitive Architectures
Of primary interest to the reader already familiar with cognitive architectures will be contextual links between nodes. These links demonstrate the relationship between the motivation for a particular architecture, the architecture itself, properties of a particular architecture, its resulting capabilities, the environments in which it does (and does not) function and general issues relating to all architectures. For instance, the discussion of frames for the Theo architecture includes contextual links relating frames to Theo's utilization of meta-knowledge and an impasse-driven control strategy. The former then suggests that meta-knowledge is used to realize meta-reasoning and self-reflection in Theo while the latter demonstrates that frames and the impasses defined in relation to frame slots allow focused behavior in Theo. Thus, the use of contextual links makes the relationship between different facets of the architecture explicit. We believe that the framework we use for developing these relationships and their explicit identification constitutes a contribution to the field of agent and cognitive architectures.
Navigating the Document
One of the reasons that such vastly different user popualtions can utilize this document is that is is built using a consciously consistent organization. This node attempts to make that organization explicit in order to make navigation through the document as easy as possible. It is divided into two parts.
Hierarchical Organization of Nodes
This diagram represents the overall organization of the document. Each of the topics available at the first level of the figure are accessible from the Title Page (the zeroth level) while nodes underneath each of the first level nodes reflects nodes that are accessible from these nodes. Only the nodes from the Guide are expanded to the second level in the diagram. The heavy lines show connections between nodes that are enumerated via index files, normally consisting of lists. For example, the node Guide in the figure consists of a list of twelve links introducing each of the individual architectures introduced in the document. The dotted lines represent contextual links. These are used only to suggest the type of links that exist, not any particular explicit links. In particular, in the diagram, there are links connecting specific architectures to Theories, Properties, Capabilities, etc., links connecting Properties to Capabilities and Capabilities to Environments, and links between architectures (Soar to Theo in the diagram). Not shown in the figure are contextual links within an architecture (See Architecture Organization for these details). Individual links are uni-directional although, for the most part, within the document there will be a link at a destination node leading back to the original (source) node.
The actual number of links in the document will be much more dense than that represented in the diagram. These in the diagram are used to demonstrate that the hierarchical organization does not enforce a particular reading pattern since following its links can lead to far-distant nodes in the hierarchy. The chief benefit of the overall organization is to aid users in navigation. This becomes particularly important when users make use of contextual links and do find themselves in new portions of the document. The Current Location markers at the bottom of each node reflect this hierarchical organization and inform users where the current node resides in the hierarchy. For instance, the marker below indicates that this node is a third-level node, its parent node "Navigation of the Cognitive Architectures Document" and that node's parent "Reader's Guide". The Title Page, as the single, zeroth-level node, is implicit in the hierarchy even though it is not explicit in the Current Location marker.
Organization of Individual Architectures
An effort has been made to organize the nodes under each architecture consistently. The information is organized analogously to the hierarchical organization of the entire document. The primary difference is that the index node for each architecture contains links to each of the nodes in the description of the architecture (rather than nodes for properties, capabilities, etc). This gives the user a quick, concise enumeration of all the properties, capabilities, and environments for the architecture. When applicable, individual components of the architecture are also listed under the architecture description section of the index node as well. The only nodes not accessible directly from the index node are the individual issue nodes. These links may be accessed through the Issues link on the index node. (Since the issues are generally consistent throughout all the architectures, they are not included as part of the index node in the interest of conciseness.)
From the index node, users already familiar with architectures or users with some knowledge of artificial intelligence may proceed directly to topics on interest. On the other hand, users with little knowledge of the architectures or AI may wish to begin with the first node, Philosophy and Methodological Assumptions, and then follow the NEXT links at the bottom of each node to navigate through the document. These links take the user, in order, down through the items enumerated in the index node.
In addition to the NEXT links, each node in the architecture includes a non-contextual link which will return the user to the architecture's index node. Additionally, terms which are used in several architectures are not generally defined for each architecture; instead the term is described as it relates to the particular architecture. However, another non-contextual link, usually entitled Go to a discussion of this property for multiple architectures will take the user to a definition of the term (see Hierarchical Organization for understanding where these definition nodes are located in terms of the overall document). Both of these links are set off from the context of the node by two horizontal lines (as below).
Finally, even though there are not explicit nodes enumerating properties only, capabilities only, etc. for individual architectures, the Current Location markers at the bottom of each node do indicate which of the higher-level categories the current node falls under. This is done so that users following contextual links can quickly identify the category of a new node, making navigation simpler and allowing concentration on content rather than traversing the document.
Sources and Document Maintenance
As mentioned in the introduction to the Reader's Guide, this project developed from projects organized during the Winter 94 offering of EECS 547, Cognitive Architectures, at the University of Michigan. Most of the readings for that course were based on papers generally written prior to 1993. The information in this document then reflects our understanding of the architectures based on these readings (which correspond to the references for this document). In many cases, work has continued on the architectures described in this document and it should be remembered that the descriptions reflect the state of an individual architecture around 1992. An attempt has been made to identify WWW sites which describe on-going architecture work. These may be found in the reference section of the individual architectures. Although this document may be updated in the future to reflect both new architectures and increased/modified coverage of the currently described architectures, for the short term, this document is going to remain fairly fixed.
It is certainly possible that some of the information in the document may reflect an incomplete understanding of a particular architecture or concept. In general, the references themselves, rather than this document, should be the basis for additional exploration of the architectures. However, if you identify an error, have suggestions or ideas concerning future versions of the document, or general comments concerning the current document, please let us know by mailing us at: firstname.lastname@example.org.
Return the Title Page