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. The autho
rs feel that the issues addressed here apply more generally to AI appl
ication domains involving complex multidimensional search, in which co
ntrol knowledge is as important to successful problem solving as is do
main knowledge. Evolution is viewed largely from the context of the He
arsay-II (HSII) speech understanding system. The appeal of the blackbo
ard 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 co
mplicate the process of estimating the expected value of potential act
ions. Among the key themes in the evolution of blackboard control is t
he development of mechanisms that support more sophisticated goal-dire
cted reasoning. In the basic control mechanism of HSII, control decisi
ons could consider only the local and immediate effects of possible ac
tions. Thus, the value of potential actions in meeting the system goal
s could be evaluated in only a limited manner. The development of appr
opriate abstractions of the intermediate state of problem solving can
be used to evaluate the non-local effect of actions relative to the ov
erall problem-solving goals. In addition, blackboard systems went from
the implicit representation of goals in HSII to explicit representati
on of the goals that must be satisfied in order to meet the overall go
als of the system. This allowed the implementation of various styles o
f goal-directed reasoning (e.g., subgoaling and planning) that were no
t supported in the basic HSII control mechanism. Other architectural m
echanisms were concerned with efficiency issues. This article examines
a number of different blackboard control architectures that have evol
ved from the basic model of HSII: HASP/SIAP's event-based control, CRY
SALIS' hierarchical control, the DVMT's goal-directed architecture, th
e control blackboard architecture (BB1), model-based incremental plann
ing for the DVMT, and the RESUN interpretation framework. A longer ver
sion of this paper is available as a technical report. It also include
s analyses of the channelized, parameterized control loop version of t
he DVMT (Decker, Humphrey, & Lesser, 1989), ATOME's hybrid multistage
control (Laasri, & Maitre. 1989) and CASSANDRA's distributed control (
Craig, 1989).