At roughly the same time as BBAI was emerging, so were reactive plans (16,20). Reactive plans are powerful plan representations that provide for robust execution. A single plan will work under many different contingencies given a sufficiently amenable context. An agent can store a number of such plans in a library, then use context-based preconditions to select one plan that should meet its current goals in the current environment. The alternative -- constructing a plan on demand -- is a form of search and consequently costly (12). In keeping with the goal of all reactive intelligence, reactive plans provide a way to avoid search during real-time execution.
A hybrid behavior-based system takes advantage of behaviors to give a planning system very powerful primitives. This in turn allows the plan to be relatively high-level and simple, a benefit to conventional planners as well as reactive plans (28). Most three-layer hybrid architectures (as described above) have a bottom layer of behaviors, which serve as primitives to a second layer of reactive plans. They may then optionally has a third `deliberative' (searching) layer either to create new plans or to choose between existing ones.
Consider this description of the ontology underlying three-layered architectures:
The three-layer architecture arises from the empirical observation that effective algorithms for controlling mobile robots tend to fall into three distinct categories:
Gat (19, p. 209)
In this description, behaviors are the simple stateless algorithms and reactive plans serve as state- or context-keeping devices for ordering the activity of the behaviors. In Gat's own architecture, ATLANTIS, (18) the second, reactive-plan layer dominates the agent: it monitors the agent's goals and selects its actions. If the second layer becomes stuck or uncertain, it can reduce the agent's activity while consulting a third-level planner, while still monitoring the environment for indications of newer, more urgent goals.