Subsumption/Symbolic Dichotomy

Traditionally artificial intelligence has taken a symbolic approach to intelligence. The theory is that a symbol system is necessary to produce intelligent behavior. The subsumption architecture was created as an alternate method for implementing robotic control. Though agents that maintain a symbolic world model can engage in cognitive activities such as planning and problems solving they have had a notoriously difficult time operating in a dynamic , real world environment. This is due to the impossibility of maintaining a consistent internal state that accurately reflects the state of the environment. In particular, imperfect sensory information and actuator control hamper the efforts to accurately maintain the world model.

The subsumption architecture addresses these issues by maintaining minimal internal state and using the world as its own best model. This allows for highly reactive, sensor based behavior with very little computation. Subsumptive agents tend to have "hard wired" bahaviors and thus are not highly flexible in nature. Furthermore, since their behavior is so closely connected to sensory input, they are not generally capable of high level cognitive acts such as planning and learning . Though Brooks maintains that the subsumption architecture can potentially form the basis for a bottom up development of such intelligent functionality, only symbolic agents have thus far been able to demonstrate such capabilities.

Back to List of Dichotomies