Transactional Blackboards

The blackboard architecture is a popular structuring framework for expert systems. W i t h this structure, an expert system is bui l t as a collection of knowledge sources which are scheduled by a controller and communicate through a shared data region, called a blackboard. The performance of such a system may be signif icant ly enhanced by the concurrent execution of the knowledge sources. However, introduct ion of concurrent execution into blackboard systems requires extension of the architecture w i th new mechanisms for scheduling knowledge source activities, synchronizing knowledge source interactions, and accessing shared data. This paper describes our design for transactionbased facilities support ing parallel execution of knowledge sources in a blackboard system. I . I N T R O D U C T I O N The blackboard architecture is an important structural framework for expert systems. In this architecture, an expert system consists of a shared data region (called the blackboard), a set of knowledge sources, and a control mechanism. The blackboard is a data base which is shared by the knowledge sources as their communication medium. Containing rules and hypotheses which express the domain expertise of the system, the knowledge sources respond to each other through observed changes in the blackboard. The cont ro l mechanism schedules execution of the knowledge sources according to information from its goal queues and the blackboard. Several expert systems have been bui l t according to the blackboard architecture. Examples include a speech-understanding system (Erman et al, 1980), a sonar interpretat ion system (Ni l and Feigenbaum, 1978), a vehicular t racking system (Lesser and Cork i l l , 1978), and a protein crystallography interpretat ion system (Terry, 1983). Al though these systems are founded on the blackboard architecture, they vary signif icantly w i th in the framework, demonstrating the u t i l i t y and f lexibi l i ty of the paradigm. Experience suggests that this architecture is part icular ly suitable for systems representing mult ip le areas of expertise and for systems solving problems wi th complex information interdependences. Multiprocessor computing environments should be capable of increasing the scope and u t i l i t y of expert systems and successfully addressing problems beyond the reach of most uniprocessors, such as real t ime speech recognition or robot control. The domain and control knowledge of an expert system may be distr ibuted onto several processors. The interactions of modular knowledge sources may simulate their modeled events, w i th both communication paths and t iming of interactions. Thus multiprocessor configurations have the potential to support the construction and execution of expert systems wi th new and useful properties. Multiprocessor computers are often di f f icul t to use. While the processors can execute in parallel, the exchange of data, code, and results among these processors can often make the overall system slow. Therefore, a balance must be reached among the costs of loading code, accessing data, and communicating requests and responses. Two extreme approaches have received most at tent ion by researchers. At one extreme are systems in which processing nodes frequently exchange small sets of data and do small computations w i th each data set (e.g., Dennis, 1980). At the other extreme are systems that place a large, autonomous program on each processing node. In these systems, the nodes exchange data infrequently and spend most of their t ime performing " local" computations (e.g., Lesser, 1978). The work described in this paper focuses on support ing systems closer to the latter extreme. We present mechanisms for constructing expert systems as collections of knowledge sources communicating through a shared data medium. These are systems in which knowledge sources executing on different processors perform moderate to large computations between communications. The integr i ty of data that is accessed asynchronously by several clients must be maintained. Providing transactional access to shared data bases is a common solution to this problem. A sequence of operations on one or more data elements, beginning w i th a start-tran3action request and ending w i th either a J. Ensor and J. Gabbe 341 commitor an abort-transaction request, a transaction is a uni t of activity w i th three properties: atomicity, consistency preservation, and permanence. Atomici ty means that , in net effect and even when failures occur, either all operations in the unit happen (the transact ion commits) or none of them happens (it aborts). Consistency preservation means that a transaction moves data from one consistent state to another. Permanence means that the effect of a committed transact ion persists, surviving any noncat as trophic failures, unt i l the next transaction involving that data is commi t ted. We extend the blackboard architecture to support systems executing in multiprocessor environments by providing transactional access to the blackboard. Our extensions are novel in their ease of use and in the richness of structure that they support. Two mechanisms are provided for safe access to the blackboard data. Knowledge sources can communicate by accessing shared data in separate transactions. Furthermore, several knowledge sources can participate in a common transaction if they need to see a common, consistent view of shared data. I I . S Y S T E M S T R U C T U R E Figure 1 illustrates a system that we designed to understand the use of the blackboard. We term the control and knowledge sources agents because they are both modular units of activity. The agents are distr ibuted on various processors and may execute concurrently. Knowledge source activities on each node are controlled by the control sources on that node. (The collection of control sources is the controller mentioned in the blackboard architecture description.) In our present implementation, the distr ibution of agents is subject to restrictions. The init ial distr ibut ion is specified by the system designer, and we provide no mechanism to support agent migration among processors. Al though the blackboard resides on a single machine, it could be distr ibuted without changing its interface.