Reading 5 - Cognitive Architectures: Research Issues and Challenges

Cognitive architecture can be thought of as the underlying structure of the human mind which specifies how information is stored and processed in different situations. For and intelligent system, it is more like a set of principles that should be followed at design time so the system can be as human-like as possible. This is the primary goal of Artificial Intelligence which aims to incorporate intelligence into machines even though they might not reach the level of organic intelligence the human mind possesses. Although this study gives numerous examples of robots and human-like AI systems, cognitive architecture can also be used to design systems on a smaller scale. Such as extremely specialized systems, for example, an intelligent vacuum cleaner. The aim is to draw inspiration from the human mind when developing a system so it can be as close to being as intelligent as possible. Multiple cognitive architectures have been proposed in research. This paper identifies several capabilities that should be present in a cognitive architecture.

Capabilities of Cognitive Architectures

These capabilities can be completely or partially present in cognitive architecture depending on how advanced the system needs to be. For instance, some systems like chatbots might need to think logically and plan what to say next but they do not need the ability to walk around and thus no motor skills will be necessary for the architecture developed for them.

The proposed capabilities are:

1. Recognition and Categorization - A cognitive architecture must be able to recognize patterns in new information (instructions or experience) and relate them to similar situations already stored in the memory. 
2. Decision Making and Choice - Being able to decide what decisions can be made in a given situation and then what decision should be made is one of the basic capabilities every cognitive architecture should support. When we play board games with a computer opponent, we see the computer making decisions regarding moves it should play. Human beings can easily beat most computers after playing with them for some time. I would say the decision making capability is there in board games but it is not very refined. It can be made extremely powerful though. For instance, AlphaGo is a computer program that can make perfect decisions and has beaten the world champion of the board game Go.
3. Perception and Situation Assessment - A cognitive architecture should have the capability to sense the environment around it much like human beings have five senses which they use to navigate around. This can be achieved through different sensors and cameras. Perception is a capability that we see even in the most basic commodities like vacuum cleaners.
4. Prediction and Monitoring - Authors give examples of pilots who are able to judge when there might be a leak in the fuel tank by noticing that fuel is decreasing unusually fast. However, I disagree with this capability being called prediction since a machine will have access to the fuel tank through some sort of digitization and can tell exactly when there is a leak. It can come under monitoring though.
5. Problem Solving and Planning - An intelligent system should be able to make plans using the information it gets from its sensors or information that is already stored in memory. Cozmo, a robot is capable of stopping and changing its direction when it detects it can fall from a surface.
6. Reasoning and Belief Maintenance - A cognitive architecture should support reasoning. Machine learning will be a good example through which a program gradually learns something and then makes it a standard for future operations but this model can be changed upon retraining.
7. Execution and Action - In some cases (like humanoid robots), motor skills are necessary for intelligent systems. The architecture should be able to store and process this information dynamically like human beings do and figure out when it wants to take action.
8. Interaction and Communication - Most intelligent systems require communication with human beings or other systems. A cognitive architecture should support some form of interaction and the ability to communicate. This capability involves inferencing and planning as well since the system will need to think about what to say and how to say it.
9. Remembering, Reflection, and Learning - A truly intelligent system can store events in memory, be able to recall them and understand new concepts. A number of sci-fi movies have been made on such advanced AI with the capabilities of remembering and learning.

Properties of Cognitive Architectures
We now discuss the properties it should have to be able to support those capabilities. We will discuss how cognitive architecture organizes knowledge in memory and how it accesses or improves this knowledge. There are multiple options to consider when dealing with the design of each property. The properties that cognitive architecture should possess are:

1. Representation of Knowledge - To be able to search through the knowledge already present in memory or for the acquisition of new knowledge, cognitive architecture should store it in a structured manner. Much like data structures that are used in programming to make CRUD (create, retrieve, update, delete) operations easier. This representation can be either 
• Declarative: Which gives the system the ability to recognize what it knows together with the ability to perform tasks
• Procedural: Through which system knows how to do a certain task but cannot recognize it if done by others.
2. Organization of Knowledge - The way in which knowledge is organized which can either be flat or hierarchical. 
• A hierarchical organization builds relationships among entities. For example, a car is a four-wheel vehicle and a bike is a two-wheel vehicle. The hierarchy will then be bile -> two-wheel -> vehicle etc. The "is a" relationships describe the hierarchy of the system. 
• A flat structure does not contain the representation of any relationship. In our example, car and truck will still be stored as vehicles but the distinction of two and four wheels might be removed.