Understanding Cognitive Architectures

Cognitive Architectures (CA) can be defined as a move from distinct boundaries of the human-technology relationship to Artificial Intelligence (AI) that is to mimic, as closely as possible, the cognitive faculties of an intelligent human being. The term “architecture” is to be understood as a framework that supports different types of intelligent behavior wherein the challenge lies in designing a system that supports behavior(s) that are as close to human cognitive processes as per the full spectrum of predefined action, learning, unlearning, choosing between options, foresight in decision making and so on.

Soar

(I found a great article that was also very consoling in the way that the author had named it: “A Gentle Introduction to Soar”🤞https://web.eecs.umich.edu/~soar/sitemaker/docs/misc/GentleIntroduction-2006.pdf

I just figured it’s by John Laird himself! So he understands the mess we’re in too..great!

The following excerpt from the article was particularly helpful in putting things in perspective:

To make sense of our everyday experience and function in the world, we need to use our
knowledge about objects and actions in pursuit of our current goals. You know many things
about a cup but you use only some of that knowledge when you want a cup of coffee, and some
other part of it when you want a place to put pens and pencils. Moreover, once you have applied
some of that knowledge, a particular cup will be full of coffee or full of pens. Your ability to
reason about other cups as pencil holders or coffee mugs will not have changed but there will be
something in your current situation that will have taken on a single role — your general
knowledge will have been tied to the particular instance of the cup in front of you. This
dichotomy between general knowledge and specific applications of that knowledge is captured in
Soar by the existence of two different types of memory. Figure 6 shows the basic structure of
Soar’s memories. (Lehman, Laird, Rosenbloom, 2006)

This is an architecture that uses procedural long-term knowledge in the form of “production rules” which organizes “operators” based on the associated “problem spaces”. We must understand each of these terms separately:

Procedural representation: There are three types of Long-term-memory (LTM): Procedural; Semantic and Episodic. Procedural memory does the following: 1. Generally the encoding of knowledge to accomplish tasks. 2. Production rules are a means for representing this knowledge. 3. It cannot recognize activities done by others. LTM must be “searched” to find the relevant material, it is not readily available.

Working Memory: In Soar, WM is represented as a set of the features and values that
make up the current state (and substates), which might include representations of the current
goal, problem space, and operator. (Lehman, Laird, Rosenbloom, 2006)

Production Rules: These refer to the specific coding rules that guide the tool to perform certain tasks. These maybe deductive (i.e. choosing from a set of pre-designed/pre-experienced solutions) or they may be abductive i.e. which observes a problem and then tries to find a new, previously unknown solution for it.

Operators: operators cause things to happen, either in the world or in the imagination (i.e. of the architecture itself).

Properties of cognitive architectures

There are some internal properties of architectures that produce the requisite capabilities that perform various functions. These include the architecture’s 1. representation of knowledge (procedural, semantic and episodic) 2. organization of that knowledge (flat or hierarchical; fine-grained or coarse-grained structures) 3. utilization of knowledge (heuristic vs. retrieval solutions vs. plans from long-term memory) and lastly the 4. acquisition and refinement of knowledge.

Though the semantics may be complex and confusing, the general idea is that architectures need to be flexible and “able” to different frames and not limited in ways that inhibit growth and refinement of designed procedures-a quality similar to humans of learning and unlearning concepts as and when required while maintaining some concrete basis for doing so.

Evaluating Cognitive Architectures

Moving towards a “unified theory of cognition” that combine the various capabilities (mentioned earlier) to create and support synergies in AI-much like actual human-beings that can perform on a spectrum of problems from very basic ones and very abstract and complex ones all the same.

The evaluation of CAs is done on the following criteria:

1. Generality, versatility, and taskability: the ability of an architecture to be able to serve diverse functions, with the least amount of effort in ways that are responsive to external commands (not just long-term memory).
2. Rationality and optimality: that architectures not only are able to match the right task to the correct problem but also use the most “optimal”/most efficient way/solution in doing so.
3. Efficiency and scalability: This metric evaluates whether a system can satisfy hard constraints of time and space and may be based on recognize-act cycle or actual complete-task scenarios. Scalability refers to the level of difficulty that a system is able to handle.
4. Reactivity and Persistence: Reactivity refers to the speed with which it reacts to unexpected situations and persistence refers to the system continuing to pursue its goal despite changes in the environment such that it does not lose track of the main objective.
5. Improvability: no explanation required, however improvement may take the form of the ability to perform new tasks, improved speed, learning from the environment etc.
6. Autonomy and extended operation: such that the system does not fail to perform in a new situation and likewise does not slow down due to an accumulation of long-term memory.

Thus, a general understanding of the capabilities and evaluation metrics provides us with a good vantage point to consider what an ideal architecture should encompass and what areas need more focus than others. Generally speaking, there needs to be more focus on categorization and understanding (than executing hard-wired tasks); on episodic memory and reflective processes; the need to encode knowledge in a variety of formalisms apart from only logic; real embodiment and how physical embodiment impacts thinking; the use of emotions and lastly the re-suability of knowledge.