3. Blackboard and Client/Server Systems

Transcription

1 3. Blackboard and Client/Server Systems 3.1 Introduction The blackboard model of problem solving arose from the Hearsay speech understanding systems [55,174,175] developed at Carnegie-Mellon University in early 70s as part of DARPA speech understanding project. The ideas of Hearsay then extended into what we now think of as the standard blackboard architecture in Hearsay-II (HSII). The HSII speech understanding system was first blackboard system whose architecture embodied all of the typical features blackboard systems. HSII brought key ideas of the blackboard model and these remain in today's blackboard systems. Numerous advances and enhancements have been made as a result of experience gained in using blackboard systems in widely varying application areas. The blackboard model support provides a great deal of flexibility in structuring the problem-solving process. Representational flexibility is an equally important in blackboard systems. The blackboard model does not place any prior restrictions on what information can be placed on the blackboard. It separates a system into: 1> a central, global data containing input data, partial solutions, and other data in various problem solving states, typically called as the blackboard, and 2> a number of independent computational modules, often called as knowledge sources (KSs). The knowledge sources perform computations based on the contents of the blackboard in order to change the blackboard. One of the good features of the blackboard model of problem solving is that the problem domain is divided into hierarchy unique to that domain. The key ideas behind the blackboard model are that the problem solving should be both incremental and opportunistic [176]. Incremental problem solving means that complete solutions are constructed piece by piece and at different level of abstraction [176]. The opportunistic problem solving means that the system chooses the actions to take the next that will allow it to make the best progress towards meeting its goal given the current situation i.e. given the available data and intermediate state of problem solving [176]. The client/server computing model delivers the benefits of network computing model alongwith the shared data access and high performance characteristics of host-based 50

2 computing. In client/server architecture the application processing is divided (not necessarily evenly) between client and server. A simple definition of client/server computing is that the server software accepts requests for data from the client software and returns results to the client [53]. In client/server architecture, the clients are separate and independent of the server software and hardware. The client applications handle interactions with user, local machine and devices while the server (database server) handles database access and control functions. This chapter briefly describes the blackboard model of problem solving and client/server computing model, and, how these two models can be integrated to develop a new kind of blackboard framework. 3.2 The Basic Blackboard Model Figure 3.1 shows the basic blackboard model [57] with two major components: 1> the blackboard, a global data structure and 2> knowledge sources, the number of computational modules containing the knowledge useful in context of application for which system is designed. KS, KS 2 Blackboard i i i KS n Figure 3.1: The basic blackboard organization [57] The blackboard serves as the medium for all communication amongst the knowledge sources. It contains the data and hypotheses (solutions, partial solutions) of the problem at hand. Usually, the blackboard is organized into different levels, representing the hierarchy of the knowledge domain, with each level representing information about the domain at different level of granularity. The hypotheses on one level are typically 51

3 linked to the hypotheses on other level. In Hearsay-II, the blackboard levels correspond to a part-of hierarchy of components of speech utterances, such as phonemes, words, word sequences, and phrases. The elements of each level are composed of elements from the level below. Sometimes, the levels are themselves structured in terms of set of dimensions or areas. This makes it possible to provide efficient associative retrieval of hypotheses based on area. A KS examines a number of elements from one level and suggests a solution element at the next higher level as shown in figure E5 3" '2~5" Figure 3.2: Organization of the blackboard The internal presentation and problem solving mechanism of a KS is hidden from direct view. KSs communicate through the blackboard only. The KS examines the state of the blackboard and once the necessary information is found, it proceeds without any assistance from other KSs. A KS is triggered for execution based on the state of the blackboard or as a result of changes in the blackboard. Rather than having each KS scan the blackboard, each KS informs the blackboard system about what kinds of events in which it is interested. Normally, the blackboard changes are described as events. The KS is executed once it is enabled, in turn it may change the state of the blackboard. The KS has two components: the KS preconditions and the KS body. The KS preconditions 52

4 are like conjunctive conditions of an if-then rule, they test the state of the blackboard and determine when the KS can be executed. The KS body contains the computational logic to be performed by the KS when preconditions are satisfied. Each KS can have its own knowledge representation and problem solving method or reasoning strategy. In literature [56], the term KS activations or instances (KSIs) is used for active processes while KSs is used for static repositories. A KSI is combination of the KS knowledge and a specific triggering context. The basic execution cycle consists of steps as shown in figure [57]: 1. Determine which KSs are enabled 2. Choose which of the enabled KSs is to be executed. 3. Execute the KS. 4. Repeat above steps until all KSs are executed or problem is solved Determine which ^ KSs are enabled -\ Execute KS Choose a KS for execution Figure 3.3: A basic control loop The issue in step-2 is how and which of the enabled KSs is to be selected next for execution. This is the problem when the application can execute only one KS at a time. This problem is known as the control problem. This may not be problem in parallel computing environment where KSs can operate in parallel. To address the control problem, a third component control is included into the basic blackboard model as shown in figure

5 KS, KS 2 i i i KS n ^ ^ ^ ^ C 0 N T R 0 T. Blackboard Figure 3.4: Basic blackboard model with control The control component acts as a gatekeeper, determining which KSs can talk to the blackboard at what times. It makes run-time decisions, by looking at the current state of the problem solving, to determine the next course of actions to be taken, in order to make good progress towards the solution of the problem. The method used to solve the control problem is known as control strategy and its implementation as a control mechanism. Blackboard systems differ greatly in their individual control strategies [176]. In some of the blackboard systems, the control strategies are quite complex and a lot of factors are considered before making decision to select next KS for execution [176]. Two major control mechanisms are used: Data-directed control and goaldirected control. Data-directed control makes the best use of available data, it places all possible KSIs into agenda and in each cycle the KSIs are rated and the most rated KSI is chosen for execution. In goal-directed control, the system goals are set, and, the role and ultimate importance of KSIs in satisfying these goals are considered. Effective control typically requires the integration of both goal directed and data based control factors [176]. The control component is often called as scheduler because it schedules the activations of KSs. Some architecture such as HASP/SIAP [177,178] use event-based control based on occurrence of predefined events. The blackboard changes are described in terms of set of blackboard event types. The predefined event types served as the KS preconditions. Since the blackboard events are directly mapped to the sequence of KSs, there is never an uncertainty about how best to respond to each event. HASP/SIAP supports another three categories of events in addition to the set of blackboard event types: clock, 54

6 expectation and problem. Clock events consist of a time and a list of KSs to be executed at that time. Expectation events are represented as the blackboard events that are expected at some point in future. Problem events are meant for the problems encountered by KSs. Figure 3.5 shows the basic control loop in HASP. In the blackboard community, three levels of details have been distinguished [57]. The general model of blackboard problem solving is the blackboard model as described above. The practical extension of blackboard model is blackboard framework, which specifies the construction details of the model. These details define the components of the model, how the components react, and, how they are controlled. The framework should specify enough information for a practical information to be built. An application, which is implemented using a particular framework, is called as blackboard application. The terms blackboard architecture and blackboard shells are essentially synonymous with blackboard framework but with stronger connotations in that their purpose is to serve as a template for the creation of actual systems [57]. KS Execute KS Rules events event andks STRATEGY MODULE: Select event category event category T EVENT CATEGORY MANAGER: Select event type identify appropriate KSs Figure 3.5: The basic control loop for HSAP [57] 55

7 3.3 Parallelism in Blackboard Systems Parallelism can be introduced in the blackboard model of problem solving at various levels [179,180]: 1> blackboard interaction machinery, 2> implementation of each KS, 3> multiple KS instances running in parallel, and, 4> the blackboard control component in parallel with domain KSIs. Corkill [181] analyzed three alternatives for parallelizing a blackboard system at KS execution level (see figure 3.6): 1. a distributed blackboard approach with multiple KS instantiations (KSI) and multiple blackboards, 2. a blackboard server approach with multiple KSI queues and a single blackboard located at one processor, and, 3. a shared memory approach, with all processors sharing a single blackboard and a single KSI execution queue. Q P "* Q p Q P Q, 1 i P Q p > Q p BB BB BB BB Distributed Blackboard Approach Blackboard Server Approach Q "" \ / 1 v ( p p I p KSI Queue Processor BB A Blackboard Shared Memory Blackboard Approach Figure 3.6: Three design alternatives for KS-level parallelism [186] 56

8 The early work by Fennell and Lesser [182] studied the effect of parallelism on Hearsay II speech understanding system using shared memory blackboard approach. One major contribution of that project is a detailed study of blackboard locking mechanism that ensures integrity. Decker et al [177] investigated the effects of parallelism on blackboard system scheduling heuristics and described a parallel blackboard system that allows multiple KSIs to execute in parallel using shared memory blackboard approach. The advanced Architecture Project at Stanford University generated two prototype blackboard architectures, POLIGON and CAGE, that exploit parallelism in rather different ways. The POLIGON system [183] was designed for distributed memory, multiprocessor hardware, with large number of processor/memory units enjoying high bandwidth communication. It was intended to run with a high degree of granularity. The knowledge sources are executed, as data becomes available instead of being polled by a central module. The CAGE system [184] - essentially a concurrent version of the AGE blackboard [185] was targeted on a shared memory system with a smaller number of processors. Programming language constructs were introduced to support parallelism at the discretion of the application programmer. Nii et al [186] reported that the performance gains are sensitive to the ways in which applications are formulated and programmed. 3.4 Blackboard Model Strengths and Weaknesses Strengths The problem domain is divided into separate and independent KSs, allowing systems to be developed in modular way. In conventional programming language, the execution flow is explicitly defined, and, the modules are invoked directly by their names and parameters. All modules have to be defined at the system design and implementation stages. In case of the blackboard systems, a KS never knows the identities or implementations of other KSs. The KSs can easily be modified, added or removed from the system. This provides greater flexibility in designing and structuring the problem. KSs need not be called by their names; they are self- 57

9 enabling, opportunistic and volunteer themselves for execution as soon as situation warrants. The table 3.1 shows the differences between conventional and blackboard systems.blackboard model of problem solving supports multiple strategies of reasoning and problem solving methodologies. Weaknesses Since the blackboard is shared global data structure, a change in representation of data may require changes to a number of KSs. This problem can be lessened, by using a separate region for each KS, thereby limiting the modifications required in other KSs [57]. Since there is no data hiding, any constraints on the contents of the data objects on the blackboard, such as integrity or consistency constraints, must be carefully maintained by every KS that manipulates that data [57]. In general, since KSs are completely independent, the reuse of KS components is less. Table 3.1: Comparing conventional and blackboard systems Conventional Systems Data structures are predefined Execution sequence is predefined All processing elements needs to be present at the time of system integration and implementation Blackboard Systems Flexibility in knowledge representation Depends on the current-state of the blackboard Flexibility in adding or removing the processing elements at any time 3.5 Client/Server Computing Model A client/server system has three distinct components: 1> database server 2> client, and, 3> network as shown in Figure 3.7. The client/server architecture distributes the processing between client and server as shown in figure 3.8 [54]. 58

10 Request for data Database Server Figure 3.7: Client/Server Architecture [187] Client Application Logic 4 ft Presentation Logic SQL requests Results Server Application Logic 4- DBMS * Figure 3.8: Distribution of processing in client/server model [54] The database server manages resources and database of information among the multiple clients efficiently. The tasks of the database server include [54]: Managing a single database of information among many concurrent users Controlling database access and other security requirements Protecting database information with backup and recovery features Centrally enforcing global data entry rules across all client applications The users interact with the system through client application. Significant advances in microprocessor technology have made today's client machines more powerful in terms 59

11 of processing power and storage thereby performing intensive computations on their own. These are smart and thick clients and use the database as an information source that is accessed only when necessary. Increased client functionality can potentially improve the performance and scalability of the system by lowering the query processing times and off-loading the server, thereby allowing larger number of clients. The tasks that the client performs include [53]: Managing presentation logic using Graphical User Interface (GUI) objects for user interface. Performing application logic. Validating the data Requesting and receiving the information from database server. The exchange of information between clients and server through the network is done through communication software at both the ends. To improve the performance of the client/server system it is important to make the proper division of work between client and server and the exchange of data should be optimized to make proper utilization network bandwidth. Modern relational database servers are more powerful and effective to develop client/server applications. Following are the some points that can be considered while utilizing the capabilities of the relational database servers [188,189]. The database servers have capacity to process huge volume of data. They are capable of processing many records at a time. Instead of using the SQL Insert, Update and Delete statements for a single record, these statements can be used to affect multiple records. This means, instead of processing a single record, the group of records is processed at server side utilizing server's capacity. Thus, it also reduces the network traffic and frees the memory required for client process. The database servers facilitate building of application logic at server side using stored procedures. The stored procedures are compiled SQL statements, which are stored as objects and as part of the relational database. Two important advantages of using stored procedures are 1> they enable SQL code to run in compiled mode instead 60

12 of.terete form, makmg their execute faster, and, 2> they provide the layer of strae on that can hide the detahs of the datahase design from chent aprons. Putt n g the Ilcat 10 n,o glc at server makes the elient code sunpler and database design can be modtfied without changing the client code. Stored procedures can also be writtl to process large number of records at server side. The client can use batch updating which helps to improve performance by locally caching changes to data, then writing them all to the server in a single update. The database triggers can be used to enforce complex business rule at server side. 3.6 On integration of Client/Server Computing and Blackboard Model The very basic similarity between client/server (C/S) systems and blackboard system is sharing of the data among many problem-solving modules. The database in C/S system resembles blackboard while client applications as knowledge sources. Table 3.2 shows the similarities between C/S and blackboard systems. However, C/S systems differ many in ways from blackboard systems as shown in table 3.3. Table 3.2 Basic Similarities in client/server and blackboard systems Client/Server Systems Client Applications Relational Database Blackboard Systems Knowledge Sources Blackboard Clients database share the information stored in Knowledge sources share and communicate through the blackboard In C/S systems, the functionality of blackboard systems such as controlling and scheduling activities at run-time, flexibility in adding or removing the knowledge sources can be incorporated. This can help to make the best use of functionality of both the systems. For example, treating a database as a blackboard would help to use existing database server as a shared blackboard server. However, this type of blackboard framework would limit what kind of information that can be represented on the blackboard unlike in conventional blackboard where there is a great degree of flexibility 61

13 in representing the information. The blackboard frameworks resulting from integration of C/S computing model and blackboard model can be flexible and interactive and suggest a new kind of class of frameworks to develop blackboard systems. Table 3.3 Differences between C/S and blackboard systems Client/Server Systems The relational databases have built-in mechanism to handle concurrency, integrity, security etc. The relational database allows storage of data in the form of tables. Difficult to store and manage complex data structures The flow of problem solving is event-based or user-driven Contents of relational database are nonvolatile and may be stored using data files Database transactions are used to insert, delete or update the contents of relational database Blackboard Systems Needs to be implemented for parallel or distributed access Blackboard provides greater flexibility in knowledge representation and knowledge can be organized in hierarchies. Complex data structures can be represented Flow of problem solving is based on current state of the blackboard Since the blackboard elements reside in memory, its contents are volatile Since, the contents of blackboard reside in memory, it is easy to add, delete or update the contents The Relational Blackboard Framework (RBF) proposed in this dissertation adds the functionality of blackboard model into client/server architecture. Figure 3.9 shows the general concept behind the RBF. The RBF uses a relational database as a blackboard {relational blackboard) and C/S computing model as a blackboard model {relational blackboard model). The relational database server is used as blackboard server {relational blackboard server) that manages the relational blackboard. The ESs, ANNs and other software are independent knowledge sources share the data and communicate through the relational blackboard as shown in figure 3.8. Information in conventional blackboard is stored in levels or regions and data in one level may be linked to next level. In relational blackboard, data is stored in tables and can be associated using views and linked through database triggers. View is virtual table in relational data model [190] in which data from base tables are combined [187]. Views 62

14 enable to work in a just one table instead of several base tables. The database recordsets and their types play a very important role in the exchange of information in C/S environment. Instead of opening a recordset and fetching a row at a time, client-side batch recodrsets are used to reduce network traffic and increase overall performance. Batch recordset and batch updating help to improve performance by locally caching changes to data and then writing them all to the database server in a single update. In RBF, the set of rows (batch) is a major unit of exchange of the information among the KSs. When KSs need to share the data they can open recordsets having common database portion. The KSs can also share the parameters through relational blackboard. r 1 1 ES r 1 1 ANN H 1 Spreadsheet r- 1 Other Software / A A A \ f KSI Queue > >\ Control and Scheduling Logic Relational Datab >ase (Relational Black board) Relational Database Server (Relational Blackboard Server) Figure 3.9: Relational Blackboard Framework The control in RBF is event based like in HASP/SIAP blackboard architecture [176] and database triggers are used to implement it. The scheduling logic is implemented using database triggers. In most of the blackboard systems, the blackboard changes are treated as events [176]. In relational blackboard, these changes are execution of data manipulation operations such as insert, delete or update. A database trigger can be used to perform an action whenever some data manipulation operation takes place. In C/S architecture, the KSs are GUI applications. A KS may respond to many GUI events like mouse click, menu-selection etc. and perform corresponding functions. An 63

15 instance of a KS can be in running mode waiting for an event to happen unlike in conventional blackboard systems where a KSI is executed only once i.e. load, execute and exit, when it is scheduled for execution. Therefore the control mechanism required in such a framework is different from control in conventional blackboard systems. The events are mapped not only to KSs, but also to the events to which they respond while scheduling the activities. For example, table 3.5 shows what are the events to which SIN KS responds and what corresponding functions it performs. Sunners et al [191] have described an interactive framework that uses GUI based knowledge sources. They have used object-oriented approach to implement the blackboard framework. In RBF, there are two types of events to which a KS can respond: user-event and control-event. Userevents are GUI events that occur when the user interacts with the KS. Control-events are generated by the control component to execute a KS or some function/s inside the KS. For example, in case of SIN KS, the user may select menu option that will execute the function TrainQ, or, the control component may send the event TRAIN NEXT to execute the same function. Hypotheses can be mapped to the events within KS instead of KS itself. Table 3.6 shows the hypothesis Continue Sin Training is mapped to the event TRAIN NEXT within SINKS. Whenever the hypothesis Continue Sin Training is posted, the control component would schedule the execution of event TRAIN_NEXT of SIN KS. An event may be considered as precondition to execute some function/s within the KSI. Therefore there may be many preconditions for a single KSI. Table 3.5: A KS (SIN) example Event TRAIN NEXT SAVE NETWORK TERMINATE Function Train() SaveNetwork() Exit() Table 3.6: Hypothesis mapping Hypothesis KS Continue Sin Training SIN Event TRAIN_NEXT 64

16 The relational blackboard stores the pending KSIs and events (KSI Queue) in a table (control table) as shown in figure 3.8. This table is shared among all running KSs. When a KSI is idle, it itself looks into the control table for any pending events to be executed. The event may terminate the KSI itself. Whenever a hypothesis is posted, the database trigger activates the particular knowledge source or event within a KS. The scheduling of KSs can be done in such a way that no more than one KSI can simultaneously update the same data, and, no redundant execution of KSs. Due to GUI KSs, the RBF is highly interactive and user-friendly. 3.7 Summary The blackboard model offers a powerful problem-solving integration architecture suitable when there are many diverse and special knowledge representations are needed. It provides the development of a system in a modular way where each knowledge source can be developed and tested separately. It facilitates the dynamic control of problem-solving activities in an incremental manner so that the progress can be made towards the solution of the problem. The blackboard frameworks differ in way how they control the activities dynamically. Due to modularity and independent of knowledge sources, the blackboard systems can be implemented in parallel and distributed processing environment. In C/S architecture, the client applications are separate and independent of the database server. Each client application is a separate and many instances of client application can be running at a time. The database servers are powerful and can handle many client requests at a time. The performance of C/S systems depends on the division of work between client and server, and, proper utilization network bandwidth. Although the blackboard model of problem solving and C/S computing model are different paradigms of problem solving, they have basic similarity like modularity and access to common database of information. These models can be integrated to develop a new class of framework that can use features of both the models. The RBF described 65

17 in this chapter essentially integrates the functionality of blackboard model into C/S architecture. Each client application is treated as a knowledge source while the relational database as the blackboard. A knowledge source in this framework can be an ES, an ANN, a spreadsheet package or any other software that has database connectivity. The database triggers play an important role in implementing the control component. The resulting framework has been very modular, extensible, interactive and user-friendly. The RBF can be used to develop the hybrid applications that need blackboard model of problem solving in a distributed C/S environment. The advantage of the RBF is that it integrates the intelligent systems as well as the database system to these intelligent systems. 66