Cognitive Architectures and General Intelligent Systems

The following is a report I wrote for my Artificial Intelligence course in September 2006. It discusses Pat Langley’s article entitled “Cognitive Architectures and General Intelligent Systems.”

In his article, Langley claims that artificial intelligence is a fragmented field that has lost sight of its original goal: to create general intelligent systems. Such a system would resemble humans in their ability to reason and make intelligent decisions, and be applicable in various environments. But because of the difficulty in designing such a system, much of the artificial intelligence field has become focused on much more specific tasks, such as computational linguistics or computer vision. Langley asserts that more attention should be refocused on the development of general intelligent systems.

There are three primary paradigms that attempt to achieve general intelligent systems: multi-agent systems, blackboard systems, and cognitive architecture. Multi-user agent systems resemble object-orientated programming; each agent has a role and defined interfaces for communicating with other agents, much like objects in a programming language like Java. The next paradigm, the blackboard system, is similar to multi-user systems in that modular agents are used. However, they communicate indirectly via a shared resource, which makes the agents less directly dependent upon each other. Because it allows for greater flexibility than a multi-user agent system, the blackboard system is thought to be more similar to cognitive thought.

Cognitive architecture is the third paradigm. This approach emphasizes interdependence – the modules constrain each other. This high degree of interdependence is considered to best model cognitive thought. Langley uses the cognitive architecture Icarus as an illustration of a Cognitive architecture. Icarus represents knowledge as propositions. For example, the proposition “(current-street me A)” could represent “I am currently on street A” within a system. This analytical representation allows Icarus to match conditions with existing knowledge and supplement its knowledge with the result of the condition.

The Icarus architecture stores long-term and short-term memories for an agent. Long-term memory consists of conceptual memory and skill memory. Conceptual memories are logical predicates that are used to make deductions. Skill memories are a set of steps that are taken to achieve a goal. The short-term memory consists of belief memories, goal/interaction memory, and two buffers: perceptual buffer and motor buffer. The perceptual buffer is refreshed each cycle with information about the environment, such as the existence and properties of objects. The motor buffer consists of all the actions the agent will perform that cycle. Belief memory is built by applying what is known in the perceptual buffer to the conceptual memory to produce inferences. When the conditions of a concept are contained in the belief memory, the result of the concept can be added into the belief memory. The goal/intention memory attempts to use skill memories combined with the information accumulated in belief memory to execute skills that reach a desired effect. When the conditions of a skill are satisfied by the belief memory, the actions specified by the skill may be executed.  Much of Langley’s article is devoted to the explanation of this interdependent architecture.

Two important aspects of the Icarus architecture are its problem solving and skill learning abilities. Icarus can achieve goals for which it does not already have a skill by using other skills and concepts. Icarus will attempt to use backward chaining to execute other goals (concepts or skills) until it is able to achieve its original goal. This is a remarkably human approach: when confronted with a problem with an unknown solution, the system applies what it already knows and is capable of doing to the problem in the hopes of discovering a series of steps that lead to the solution. Additionally, whenever Icarus reaches a goal by problem solving, a new skill is saved that contains the subgoals used to achieve the goal. Since these new skills are available for future use, these skills can be utilized by later attempts at problem solving. It remembers the solution to a previously unknown problem so that if the system encounters the problem again, it can draw upon its experience to quickly arrive at the solution. Incremental increases in skill memory occur when these later attempts successfully use a previously learned skill and incorporate it into the new memory.

Langley advocates for advances in the development of general intelligent systems, and describes cognitive architecture as a potential avenue for achieving such a system. He familiarizes the reader with cognitive architecture by quickly explaining the structure and mechanisms of the Icarus architecture. He finishes by acknowledging other promising architectures and by identifying issues, like episodic memory, which need more research. The resulting article raises awareness about the desire for general intelligent systems and the potential of cognitive architectures.