The philosophy behind Subsumption Architecture is that the world should be its own model. According to Brooks, storing models of the world is dangerous in dynamic, unpredictable environments because representations might be incorrect or outdated. What is needed is the ability to react quickly to the present. Thus, sensor readings should be mapped quickly and directly to actuator commands in a decentralized fashion.
The behavior of insects has been used as a model for Subsumption Architecture.
Subsumption architecture defines layers of augmented Finite State Machines (FSMs). (FSMs are "augmented" with timers.) Sensors feed information into FSMs of all levels. The FSMs of the lowest level control actuator parameters. FSMs of higher levels may inhibit (attenuate the signal of one output wire) or suppress (attenuate the signal on all output wires) output values of the FSMs on the layers below them. In this way, a hierarchy of progressively refining behaviors may be established. A spinoff of Subsumption Architecture that allows limited global state is Behavior-Based Programming. Also, Subsumption Architecture has been incorporated in parts of other architectures like Gat's Atlantis.
Agents designed according to Subsumption Architecture are generally non-symbolic. They have no global representation, and are decentralized.
They have rapid-response for dealing with their dynamic and unpredictable environment. Their lack of symbolism and global representation adversely affects their taskability -- they are non-programmable, single-purpose devices.
Subsumption architecture agents are designed to perceive and act. They do not plan, let alone meta-reason. Their learning is on a small scale: the lack of global state means that each FSM can only try to optimize its own behavior. They have not demonstrated natural language ability, analogy usage, or naive physics, for these require symbols. They can navigate via reactive mechanisms. Getting such agents to cooperate is a topic of present research.
Brooks' Subsumption architecture was obviously designed for a dynamic environment. The architecture need not worry about an imperfect world model -- it has none.
Subsumption Agents are explicitly designed for real environments, where they face real-world sensor and actuator problems. Agent modules may attend to various aspects of their environment with different levels of urgency because higher-level modules may suppress and inhibit lower level ones from doing their default actions.
Brooks purposefully made his agents with limited computational power.
Because of its philosophy of using the world as its own model Subsumption Architectures are only applicable to Markov domains in which, at any time, the best next action is deducible from current sensor readings.
Subsumption Architecture agents are rational in that they are dedicated agents that try to achieve fixed goals.
They are not taskable at all. Taskable architectures need to hold state information to remember the current goals. State is not directly accessible in subsumption.
Scalability is a major concern for Subsumption Architects. Scaling them up seems so for to be an un-solved engineering problem.
They have Psychological Validity in the connectionist sense: symbolic mechanisms are not assumed given.
Other Properties
Back to the Title Page