Neural Computing and Production Systems*

The application of neural computing to the problem of matching in production systems is addressed. The computation time required by this problem can be significantly reduced by using the massive parallelism and pattern recognition capabilities available through neural computing. A new neural computing model, called here the ProNet, is introduced and explained in detail. The ProNet is applied to the match phase of the production system interpreter in an attempt to yield a reduction in time and space requirements by matching a!: of the productions to all of the working memory elements simultaneously. 1.0 INTRODUCTION The production system, a special type of expert system, will probably continue to be used to assist both humans and computers in specific tasks for future applications of artificial intelligence to Intelligent Control. If this type of system is to be used efficiently in real-time applications, the speed at which the production system operates must be considered. Present production system schemes are becoming faster, but still can be improved. A new alternative approach to present production system schemes uses neural computing to increase the speed of the production system. This is accomplished by performing the match phase of the developed production system using special hardware. Production systems are expert systems which use rules, called productions (rules, production rules), to represent knowledge and which use a particular interpreter to perform the actions of the production system. The form for the production addressed in this paper is IF a1 and a2 and ... and aj THEN bl and b2 and ... and bk where the ai are the antecedents and the bi are the consequences of the particular production. Customarily, the conjunction of the antecedents is referred to as the left-hand side (LHS), and the conjunction of the consequences is referred to as the right-hand side (RHS). The working memory (WM) contains the data which is compared to the productions. The individual elements of the WM are referred to as the working memory elements (WMEs). The production interpreter performs the comparison of the WM to the productions. It is commonly assumed that the interpreter should have a three phase cycle: (i) Match. Compare the LHS of all of the productions to the WMEs. If the LHS is satisfied, include the production in the conffict set, the set of satisfied productions for the present WM state. (ii) Select. Choose one production from the conflict set to execute. (iii) Act. Execute the production in accordance with the RHS of the chosen production. Of the three phases, the match phase traditionally consumes the most time of the production interpreter. Using conventional approaches, a production interpreter can spend more thsn 90% of its time in the match phase of the production cycle [I]. The Rete Match Algorithm, introduced in [1,2], avoids the brute force approach ofsequentially matching productions against WMEs by manipulating the productions and the WMEs to form a software tree structure to increase the speed of the production interpreter. Since its introduction, other Rete based algorithms which attempt to increase the speed of the production interpreter have been introduced [3-IO]. To further reduce the amount of time consumed by the interpreter in the match phase, special hardwares have been developed using *This work was partially supported by the Jet Propulsion Laboratory, Pasadena, California under contract 957856. parallelism and multiprocessor architectures [6,11-181. Almost all of these attempts are based, at least in part, on the Rete Match Algorithm, which is assumed to vield the most efficient match phase of the production system interpreter. These architectures smve to decrease the time required in the match phase by attempting to match as many rules as possible in parallel and by attempting to fire as many rules as possible in parallel. In addition, because these proposed architectures intend to use a multiprocessor implementation, they will consume a significant amount of physical space when realized. A new method is proposed here which simultaneously matches all of the productions to all of the WMEs in parallel via neural computing. In this paper, the potential for using neural computing for the match process of the production system interpreter is investigated. The use of neural computing to aid expert systems was addressed in [I91 which proposed to increase the speed of expert systems by using neural computing for the select phase. The use of neural computing in this paper attempts to achieve increased speeds and reduced space requirements by using neural computing for the match phase. Using a new neural computing model, the ProNet, which resembles the single layer perceptron of [23 ] , to perform the computations in the match phase of the production interpreter, the amount of time required in the match phase can be reduced. Compared to the Rete Match Algorithm, the proposed method achieves a significant increase in speed. In addition, with the recent advances in the realization of very large scale integration and electro-optical techniques for analog and parallel computation, the possibility exists to reduce the physical space required by the production system. In Section 2, a new model for neural computing, the ProNet, is introduced, examined in detail, and, in Section 3, used to perform the match phase for the interpreter of a developed production system. Next, in Section 4, a simulation of a small production system using the ProNet is presented. Finally, in Section 5 , some concluding remarks are made. 2.0 T H E PRONET The ProNet is a new neural computing model based on another neural computing model, the Hamming net (HN). A derivation and explanation of the HN can be found in [20,21]. The HN is a combination of a feedforward net and a feedback net and is used to classify patterns for speech recognition. The HN is given an input vector and specifies which stored exemplar pattern it most closely matches as a function of the Hamming distance. The exemplar patterns are previously stored patterns which exemplify all of the possible patterns to be passed through the system. Due to the fact that only one exemplar patten can be identified with the HN, the "ProNet" is created here and used to identify more than one exemplar pattem. The ProNet uses the HN's feedforward perceptron-like net with slight alterations to the weights and replaces the HN's feedback maxnet by changing the biases of the feedforward net. The ProNet's purpose is for pattern identification in production systems. The ProNet, shown in Figure 2.1, is a feedforward net closely resembling and mimicking the operation of the single layer perceptron. An input vector is presented to the ProNet through the input vector U = [ul U;! ... UN]. The ProNet is expected to identify which exemplar patterns x E X appropriately match it, where X is the set of all exemplar patterns. In the ProNet, the exemplar patterns are stored via the weights wij, where the subscript i denotes which input element and the subscriptj denotes which exemplar pattem. The output vector y = L y l y2 ... y ~ ] identifies which exemplar patterns have been matched. Basically, the ProNet simultaneously compares in parallel an input vector U to every exemplar pattern x E X . One advantage in using the ProNet is the ease with which the weights and the biases are found, compared to the single layer perceptron which requires multiple passes of the input data to train the weights. The weights wi, and the biases c, are determined based on the exemplar pattems. Assume all of the input vectors to the ProNet 0-8186-2012-9/89/0000/0665$01 .OO