This paper examines the issues that arise in the control of blackboard systems for applications with large and complicated search spaces by analyzing the evolution of blackboard control architectures. We feel that the issues addressed here apply more generally to AI application domains involving complex multi-dimensional search, in which control knowledge is as important to successful problem solving as domain knowledge. Evolution is viewed largely from the context of the Hearsay-II (HSII) speech understanding system. The appeal of the blackboard model is that it provides great flexibility in structuring problem solving. On the other hand, many of the features that are responsible for this flexibility make effective control difficult because they complicate the process of estimating the expected value of potential actions. Among the key themes in the evolution of blackboard control is the development of mechanisms that support more sophisticated goal-directed reasoning. In the basic co...

chematically in Figure 1. It comprises a set of data structures (labeled boxes) in a global memory and a three-step execution cycle: (a) an executor executes the next action, making changes to memory and producing associated events; (b) an agenda manager triggers known actions whose conditions are satisfied by those events and puts contextspecific instances on an agenda of possible actions; and (c) a scheduler rates possible actions against the current control plan and chooses the one with the highest rating as the next action to be executed. Control plans play a central role in the architecture. As shown in Figure 1, the control plan is a data structure containing any number of component plans, each with its own hierarchical and temporal organization. Typical control plans are abstract and do not specify sequences of particular actions. Instead, each "step" in a plan specifies: (a) a class of intended actions in terms o

for Correspondence An agent should adopt different control modes in different situations. Depending on the predictability of its environment and the constraint imposed by its goals, the agent should modulate its sensitivity to run-time events and its commitment to specific actions. We propose an opportunistic control model that supports this flexibility. 3 Abstract Given its multiple goals, limited resources, and dynamic environment, an intelligent agent must decide which of many possible actions to execute at each point in time. Planning and reactive models embody two different modes of control. By contrast, we characterize a two-dimensional space of control modes, each of which maximizes the quality of run-time behavior in the corresponding region of a two-dimensional space of control situations. The situation space is defined by dimensions representing the predictability of the agent's task environment and the constraint imposed by its goals. The space of control modes is defined ...

We are developing a generic control architecture suitable for use as a single intelligent agent or as multiple cooperating agents. The generic architecture combines a task-oriented domain controller with a meta-controller that schedules activities within the domain controller. The domain controller provides functions for model-based situation assessment and planning, and inter-controller communication. Typically, these functions are performed by modules taken from a repository of reusable software. In tasks that are simple, deterministic or time-stressed, the modules may be compiled into or replaced by conventional control algorithms. In complex, distributed, cooperative, non-deterministic or unstressed situations, these modules will usually exploit knowledge-based reasoning and deliberative control. To improve the controller development process, we are combining many of the best ideas from software engineering and knowledge engineering in a software environment. This environment incl...