In General
Designer's goals
The designers of an architecture usually have explicit goals, and reasons why they want to build a particular architecture (other than the fact that it is fun!)
Limitations
The design of each cognitive architecture usually has certain limitations that make it difficult to achieve certain properties, capabilities, or make it difficult for to build agents for certain kinds of environments.
Distinguishing features
Each architecture usually has one or more notable or distinguishing features that set it apart from the rest.
Agent properties vs. Architecture properties
In the analysis of cognitive architectures, sometimes it is difficult to separate the analysis of the architecture from the analysis of agents implemented in that architecture. Some architectures have been used to implement many agents (like Soar), and some are really only part of a single agent (like Homer). Others, like MAX, leave a large part of the functionality to be implemented in the agent.
Sound bites
Sometimes the authors of a particular architecture like to repeat certain phrases that summarize some aspect of their view on cognitive architectures...
Important aspects of building cognitive architectures
This is similar to "Sound bites", but perhaps there is no catchy phrase that goes along with it.
Psychological plausibility
Architectures are designed with different relationships to human cognition. Some are simply inspired by human cognition in some way (as almost anything in AI is), some try to impose some of the constraints of human cognition on their architecture, in the hope that it leads in the right direction, some are designed as cognitive models, and others are somewhere in between. If the authors said something like "humans can do X and our architecture can do X", that counts as no. If a parallel was drawn between mechanisms, or cognitive modelling attempted, that counts as "yes".
- Yes:
- No:
Subsumption,
Atlantis,
Theo,
PRODIGY,
MAX,
AIS,
Homer,
Soar,
RALPH,
ERE,
Real vs. simulated
Different architectures are designed to function in different environments, like office environments, sea environments, etc. There is an orthogonal dimension on which environments can be viewed, and that is whether or not there is an actual robot functioning in the real world, or a simulated agent functioning in a simulated world. Two ends of the spectrum of positions taken on this issue are:
- There is complexity involved with real robots operating in real environments that can not be duplicated by simulation, which tends to simplify things (real-time constraints can be relaxed, noise can be eliminated, and the issues of sensors and actuators become much simpler). Cognition cannot be addressed independently of real perception and action. If you develop an agent under simulation, when you try and make it work in the real world you will have to go back and redesign everything.
- Simulation provides a powerful tool for testing the cognitive capabilities of an architecture, without the cost of building extra hardware, and in reasonable amounts of time. Simulated environments can be made sufficiently complex as to require complex behavior. Not all agents need to ultimately be "ported" to the real world; some of them may be developed to operate in the internet environment, for example.
"Robot machismo" and "simulations are doomed to succeed" are some of the sound bites overheard here at Michigan.