Modular Integration of Connectionist and Symbolic Processing in Knowledge-Based Systems Melanie Hilario*, Christian Pellegrini* & Frédéric Alexandre** *CUI
- University of Geneva, CH-1211 Geneva 4, Switzerland **INRIA-Lorraine, BP 239 F-54506 Vandoeuvre-les-Nancy Cedex, France Email:
[email protected],
[email protected],
[email protected]
Abstract: MIX is an ESPRIT project aimed at developing strategies and tools for integrating symbolic and neural methods in hybrid systems. The project arose from the observation that current hybrid systems are generally small-scale experimental systems which couple one symbolic and one connectionist model, often in an ad hoc fashion. Hence the objective of building a versatile testbed for the design, prototyping and assessment of a variety of hybrid models or architectures, in particular those which combine diverse neural network models with rule/model-based, cased-based, and fuzzy reasoning. A multiagent approach has been chosen to facilitate modular implementation of these hybrid models, which will be tested in the context of real-world applications in the steel and automobile industries.
1.
Introduction
Current efforts at integrating symbolic and neural processing can be divided into two major approaches. In the unified approach, better known as connectionist symbol processing, neural networks are used as building blocks to create a functional symbolic architecture (e.g., a connectionist expert system). The unified approach has been actively investigated since the renaissance of neural networks in the 80’s, and is illustrated by systems such as DCPS (Touretzky & Hinton, 1985, 1988), TPPS (Smolensky, 1990), CAP2 (Schneider & Oliver, 1991), MACIE (Gallant, 1993), CONSYDERR (Sun, 1991), RUBICON (Samad, 1988, 1992) and NPS3 (Kasabov & Shishkov, 1993). In the hybrid approach, symbolic components and neural nets are combined to ally the advantages of both the symbolic and the connectionist paradigms. Integration via hybridization is the main focus of MIX (Modular Integration of Connectionist and Symbolic Processing in Knowledge-Based Systems), an ESPRIT project aimed at investigating strategies and developing tools for the integration of symbolic and neural components in knowledge-based systems. Although a number of connectionist symbol processing systems are hybrid in the sense that they combine different representation schemes (local/distributed hybrids) or different neural network models (e.g., backpropagation/competitive hybrids), these pure connectionist hybrids are in a different class from hybrid symbolic-connectionist systems, which comprise distinct but interacting symbolic and connectionist subsystems. For the purposes of this paper, we will use the term “hybrid systems” to refer to hybrid symbolicconnectionist (HSC) systems, unless specified otherwise. This paper is organized as follows. The following section discusses the state of the art in HSC integration. Section 3 describes the MIX project, its research objectives as well as ongoing and future work. Section 4 concludes. International Symposium on Integrating Knowledge and Neural Heuristics (pp. 123-132). Pensacola, Florida. May, 1994.
2.
Overview of hybrid symbolic-connectionist systems
Understandably, early hybrid systems tend to be experimental, small-scale models aimed at elucidating some of the many theoretical and technical issues involved in HSC integration. These systems can be classified along three different dimensions: (1) the mode of integration of their symbolic and connectionist subsystems, (2) their degree of coupling, and (3) the functional role they play in relation to each other and to the overall system. Instead of discussing these dimensions separately, we shall present hybrid systems according to their integration mode and situate each on the fly along the other dimensions. We distinguish four main integration modes: chainprocessing, subprocessing, metaprocessing and coprocessing (see Figure 1).
x x
y
x
x SM
y
CM
y Environment
(a) Chainprocessing
Environment
(b) Subprocessing
Environment
(c) Metaprocessing
CM = Connectionist module SM = Symbolic module
Environment
(d) Coprocessing x, y = CM or SM, with the constraint that x≠y
Figure 1. Hybrid SC integration modes
2.1
Chainprocessing
In chainprocessing mode, the two modules operate in sequence and are loosely coupled in the sense that interaction is limited to a single transfer of control and information from one to the other. Two main configurations can be distinguished: (1) the symbolic module (SM) acts as the main problem solver and is assisted by a connectionist preprocessor and/or postprocessor; (2) the connectionist module (CM) is the main processor and is assisted by a symbolic preprocessor and/or postprocessor. An example of a hybrid system with a symbolic preprocessor assisting a connectionist main processor is described in (Goodman et al., 1989). The preprocessor (ITRULE) automatically generates probabilistic rules from input financial data. Learned rules are 0-order rules of the form IF attribute-1 value-1 THEN attribute2 value-2 {with probability} p. All attributes are binary-valued so they can be mapped directly to binary units. These rules are then compiled into a neural net: attributes of rule conditions are mapped into input units, attributes of rule conclusions into output units, and information metrics (such as p) into connection weights. The result can then be loaded into the main processor, a neural network simulator. The reverse case—that of a connectionist preprocessor giving support to a symbolic inference system—is illustrated by SETHEO (Ertel et al., 1989), a first-order logic SEquential THEOrem prover that uses a depthfirst strategy. The preprocessor’s task is to compute the optimal depth bound for proving a given formula (ideally, the length of the shortest proof for that formula). Since there is certainly a relation between the syntactic structure of a formula and the optimal depth, a three-layer neural network (NN) was trained by backpropagation on 100 randomly generated formulas and their corresponding optimal proof depth. The trained neural net can then act as a preprocessor for SETHEO: it receives as input a feature vector represent-
ing the theorem to be proved and outputs the recommended depth bound. Tested on previously unseen formulas, the NN delivered a near-optimal depth bound (a deviation of at most 1 from the optimal proof depth) in 76% of the cases. Another form of this integration mode is postprocessing: once the main processor has done its job, a postprocessor of a different type may be needed to convert the result into a form useful to the application environment. A well-known example is that of a connectionist system which has to explain its reasoning strategies to a human user. Since one problem of connectionist processing is its lack of perspicuity, a symbolic postprocessor may be used to build comprehensible explanations of NN behavior.
2.2
Subprocessing
In this integration mode, one of the two subsystems is embedded in and subordinated to the other which is the main problem solver. In INNATE/QUALMS (Becraft et al., 1991), for instance, the SM is the main processor; the embedded CM executes specific tasks subcontracted to it by the SM. Neural net simulation is a subroutine called by the action part of the symbolic rules. A previously trained neural net is called to diagnose process faults in a distillation plant. Once a given fault scenario is classified, the symbolic knowledge base analyses the CM's diagnostic using a deep model which takes into account symbolic information poorly handled by NNs, after which it either confirms thethe diagnostic or proposes an alternative. In Giambasi et al.’s (1989) model, the CM is not called explicitly by SM rules but comes into action as a subroutine of the rule interpreter. When a rule is selected for execution, its associated neural net is activated to compute uncertainty factors for each of the facts added by the rule. Subprocessing differs from chainprocessing on two points. First, while a pre/postprocessor usually has direct contact with the environment, a subprocessor does not; it interacts with the environment via the processor in which it is embedded. Second, chainprocessing implies loose coupling whereas in subprocessing, the main and the auxiliary processors are more strongly coupled in that they interact iteratively. INNATE/QUALMS and Giambasi’s system are examples of moderate coupling: communication mode is similar to that of chainprocessors, i.e., a simple function call with argument passing, but interaction takes place at each recognize/ act cycle of the symbolic rule-based system. An example of a tightly coupled system in subprocessing mode is the hybrid composed of SCRAPS and an NN that learns by backpropagation (Hendler, 1989). The main processor is a marker-passing semantic net which performs complex reasoning tasks like planning. A shortcoming of this symbolic problem solver is its inability to recognize useful perceptual similarities between objects; for example, it cannot infer that a letter-opener can be used as a crime weapon on the basis of perceptual features similar to those of a knife. To remedy this shortcoming, the semantic net is connected to a neural network which has been trained to recognize microfeatures, defined as features which have little meaning on their own but which taken as a set represent a concept. As a result, activation spreads back from perceptual features identified by the NN to objects in the semantic net which share these features. The particularity of this system is that communication between the two systems is smoother than a simple parameterized function call, since there is a direct connection between the leaf nodes of the semantic net (those representing individual instances) and the output nodes of the neural network which encodes the microfeatures as a result of prior training.
2.3
Metaprocessing
One of the two subsystems is the base-level problem solver and the other plays a metalevel role (such as monitoring, control, or performance improvement) vis-à-vis the first. One well-known case is where the connectionist component enforces search control over the symbolic main processor. For example, in a system which solve high school physics problems (Gutknecht & Pfeifer, 1990), the NN receives as input the known
variables and the final goal variable; it outputs the next subgoal variable to solve for, in effect guiding the symbolic problem solver in the solution process. Search control via connectionist metaprocessing is another form of proof system support experimented on using SETHEO (cf. 2.1). In proving first-order logic formulas, heuristic methods must be used to curb combinatorial explosion by selecting the most promising alternative at each choice point. An evaluation function assigns a numeric rating to the candidate nodes, which are then explored in order of decreasing ratings until a branch leads to success or all candidates have been tried. This evaluation function is computed by a neural network which has been trained, using the backpropagation algorithm, on example proofs produced by the same system without heuristics. (Suttner & Ertel, 1990). Metaprocessing, like subprocessing, involves tighter interaction of the two subsystems than preprocessing: whereas in static proof support, the NN comes into action just once (ex. the depth bound is computed once and for all before the proof process), in dynamic proof support, the connectionist metaprocessor is activated at each branching point of the search process.
2.4
Coprocessing
The symbolic and the connectionist subsystems are equal partners in the problem solving process. Each can interact directly with environment, the SM can transmit information to the CM by knowledge compilation and the CM can do the same for the SM by knowledge extraction. Each subsystem can play a diversity of roles within the overall system. An example of this integration mode is SYNHESYS (Giacometti, 1992), a hybrid system shell which has led to the development of hybrid expert systems like SATAN (aeronaval tactics, Amy et al., 1990) and SHADE (medical diagnosis, Iordanova et al., 1992). Here both the connectionist and the symbolic components can be used for knowledge acquisition. Similarly, problem solving can be done by both the SM (through the rule interpreter) and the CM (through network activation). However, when a problem is presented, the CM is run first because it is more efficient than the SM in solving problems it has been correctly trained on (speedup role of the CM); if it arrives at a decision, the SM backward-chains to check the decision (validation role of the SM). If the decision is approved, the SM rule trace can be used to explain the decision (explanation role of the SM); if not, the CM is adapted to undo the association between the situation and the spurious decision (“forgetting” role of the CM). If no decision has been produced by the net, the SM forward-chains to take a decision. If it succeeds in doing so, the situation-decision pair is fed into the NN so it can learn the association. Another example of the coprocessing integration mode is EBL/ANN (Shavlik & Towell, 1991), in which a symbolic explanation-based learner uses a roughly correct domain theory to explain why an example belongs to the target concept. The explanation tree produced by the EBL module is generalized and mapped into a neural network: its root node is mapped into the NN’s output unit, leaf nodes into the input units, intermediate nodes into hidden units, and arcs representing dependencies into net links. Instead of being set randomly, the connection weights between units are calculated according to whether the domain theory asserts necessary or prohibitory dependencies among the concepts represented by the units. The resulting net is then trained on a set of classified examples using backpropagation. Here, the value added by hybridization is that learning is more effective than if it had been done using EBL alone or backpropagation alone. Connectionist inductive generalization allows the system to refine an explanation structure that is deficient due to an imperfect domain theory. On the other hand, the EBL phase enhances NN performance by setting initial connection weights on the basis of the domain theory rather than at random, thus accelerating convergence.
3.
The MIX Project
Project MIX is a cooperative venture of AI research teams belonging to six European organizations: INRIALorraine, France; CUI-University of Geneva, Switzerland; IMAG-Grenoble, France; Kratzer Automation, Munich, Germany; Technical University of Munich, Germany; and Technical University of Madrid, Spain. It is funded jointly by the Commission of the European Communities, the Swiss Federal Government and the project participants themselves.
3.1
Research Issues and Objectives
The goal of project MIX is to advance the state of the art in hybrid SC systems; its objectives are to: (1) enhance the versatility of hybrid systems by combining a variety of representation, inferencing and learning schemes from the symbolic and the connectionist paradigms; (2) augment the reasoning power of hybrid systems by overcoming their traditional limitation to propositional logic; (3) clarify the theoretical impact of symbolic-connectionist integration on fundamental machine learning issues such as the choice and combination of learning methods, credit assignment and knowledge assimilation; (4) ensure the scalability and real-world applicability of the resulting hybrid models by testing them on prediction, optimization and classification problems in industrial and other areas. Below we show how these objectives are grounded on significant problems and issues left open in previous work. Diversifying hybrid models of symbolic-connectionist integration In general, hybrid SC systems have been experimental implementations limited to one or two of the integration modes described in Section 2. Moreover, the symbolic module is often reduced to one reasoning scheme (most often rule-based reasoning) and the connectionist module to one NN model (usually a multilayer feedforward network which learns by backpropagation). On the connectionist side, preponderant use of backpropagation learning poses an important subissue which is the lack of incremental learning (a notable exception is SYNHESYS whose CM is a prototype-based NN that learns incrementally). Typically, the network is trained on a training set and its performance is measured with respect to a test set; it is then used in its application environment without further modification. If the training set does not correctly represent the full range of the performance set, or if the environment changes, the network will have to be retrained to reflect the new conditions. But since learning is non incremental, the NN forgets past knowledge when it is retrained; the only solution is to retrain it all over again on the set union of all previous and new instances. Repeated one-sweep learning, though highly inefficient, is feasible when the difference between the old and the new training sets is a simple matter of statistical distribution. However, when it concerns the problem characteristics themselves, standard connectionist learning in the form of connection weight adaptation proves inadequate. In this case, modifications may have to be made to network topology and input/output encoding. This highlights another research issue in connectionist learning, the need for a principled approach to network configuration and data encoding which will pave the way for the automatic execution of these tasks. As for the symbolic component, it is generally monolithic (e.g., a single RBS or semantic net) and non modular. Lessons learned from symbolic multiagent systems have yet to be applied in the development of hybrid systems. Moreover, most previous work on SC integration has been based on the premise that connectionist processing adds learning and generalization abilities to inherently nonlearning symbolic knowledge-based systems. Actually, symbolic learning has been an area of intensive research since the birth of AI, but very little work has been done to integrate both symbolic and subsymbolic learning in a hybrid learning system. However, current research in symbolic multistrategy learning (Tecuci 1993) and connectionist hybrid learning (de Francesco, 1994), together with seminal attempts to combine symbolic and connectionist learning
(e.g., EBL/ANN’s integration of EBL and backpropagation learning) could pave the way towards a general theory and methodology of multistrategy, multiparadigm learning. Knowledge representation and reasoning power The vast majority of existing hybrid systems lack representational and reasoning power. All the rule-based subsystems described in Section 2 are restricted to propositional rules. Aside from the inability to process variables in rules, many implementations impose drastic constraints on symbolic representation: for instance, the rules used in EBL/ANN and in Goodman et al.'s system take only one conclusion and use only binary-valued attributes. This situates hybrid SC systems approximately at the stage where symbolic production systems were in the early 1970s. Paradoxically, these limitations within the SM are artefacts of coupling with NNs whose knowledge representation restrictions are well known. First-order representation schemes—whether using frames, objects or rules—are straightforward to implement in state-of-the-art symbolic knowledge-based systems. Constraints on symbolic knowledge representation are generally added to ease knowledge transfer between the symbolic and the connectionist sybsystems. Thus the needed improvement has to be effected either within the connectionist subsystem itself or at the interface of the two subsystems. Current research in unified symbolic-connectionist integration, particularly on the problem of variable binding (Sun, 1992; Mani & Shastri, 1993; Kasabov & Shishkov, 1993) may provide useful clues as to possible solutions. Impact of hybrid SC integration on basic machine learning issues The combination of symbolic and connectionist learning strategies in one system adds new complex dimensions to basic issues in machine learning, such as choice of learning method, credit assignment and knowledge assimilation. The choice of learning method is a problem entailed by the combination of different learning strategies within each paradigm. Clear criteria should be established to decide when to use, for instance, inductive logic programming or case-based reasoning in a symbolic system, or when to use backpropagation or simulated annealing in a connectionist system. The problem becomes more complex, of course, when both paradigms are combined in a single system. The credit assignment problem is that of deciding which system element must get the credit or blame for performance improvement or degradation. Within a rule-based system, the problem is deciding which rule to delete or modify. This is the credit assignment problem in its classic form, and previous research has yielded working solutions in symbolic singlestrategy learning systems. However, the problem is, here again, much more complex in mutlistrategy hybrid systems. The knowledge assimilation problem is that of incorporating learned knowledge into the system’s knowledge base so that it becomes a fully exploitable part of the cumulative system theory. One of the specific challenges faced by researchers in HSC integration is assimilating the results of connectionist learning into a system’s (symbolic) domain model or theory. Scalability and real-world applicability of hybrid SC systems Among the hybrid systems described in Section 2, several pursue research or academic objectives: those based on SETHEO are aimed at providing support for theorem-proving; Gutknecht & Pfeifer’s system solves high school physics problems; Giambasi et al.’s fuzzy reasoner works on datasets generated from mathematical formulas. Hendler’s semantic/neural network hybrid and EBL/ANN could potentially be applied to complex tasks but have been tested on rather simple examples. Among those which address real-world applications, ITRULE is illustrated on a small financial database with binary-valued attributes. More demanding applications are INNATE/QUALMS’ diagnosis of process faults in a distillation plant and SYNHESYS’ medical diagnosis and aeronaval tactics decision support. On the whole, it still remains to be shown that hybrid systems can scale up to complex real-world problems such as those faced in the industrial sector.
3.2
A testbed for hybrid symbolic-connectionist integration
Given the current state of the art, we feel it is premature to design a generic hybrid architecture that will meet the objectives outlined above. As a consequence, the project’s concrete engineering objective is, more realistically, the development of a software environment for hybrid SC integration—a testbed that will support the design, implementation and comparative assessment of different hybrid models and architectures. Building the testbed: a distributed approach The testbed for hybrid symbolic-connectionist integration will be built using a distributed framework for the cooperation of multiple heterogeneous agents. Work has started on the specification and implementation of a multiagent toolkit with generic agent structures and services as well as communication protocols. More specific agent classes have been created for selected symbolic and connectionist inferencing or learning tasks (e.g., case-based reasoner, fuzzy inferencer, multilayer perceptron). These prototype tools are expected to facilitate modular construction of diverse hybrid models. An application agent can a priori draw from a variety of symbolic and connectionist reasoning templates; but with experience and learning, it will develop highly specialized mechanisms that can be symbolic, connectionist, or novel combinations of both. In other words, for a given application problem, the final integration mode of the symbolic and connectionist components will emerge from the problem-solving activity of the different agents, in function of the application context and task characteristics. Testing the testbed: experimental hybrid models The vocation of the testbed is to support experimentation with different hybridization styles and models. To test the versatility of this integration platform, we have chosen a variety of computational models from both the symbolic (fuzzy logic, case-based reasoning, rule/model-based reasoning) and the connectionist (backpropagation, prototype-based, and Kohonen-type networks) paradigms. Partial hybrids will be designed using these symbolic and neural templates implemented as agents. The first will blend fuzzy logic and connectionist reasoning by using the knowledge acquired by several neural networks (NNs) to learn new fuzzy rules or to adjust existing ones. Another partial hybrid will couple symbolic case-based reasoning (CBR) with connectionist processing, in particular with prototype-based networks. CBR and prototype-based nets can be seen as complementary mechanisms for constructive induction: while the common goal is evaluating a novel case or example on the basis of its similarity to known cases, the distinction lies in the representation of what is already known. In CBR background knowledge is embodied in a body of real cases whereas the “prototypes” of prototypebased nets correspond to generalized abstractions of known cases rather than to any real case. Moreover, the case-based reasoner can provide explanation facilities which are wanting in a prototype-based network. On the other hand, neural processing can provide functional support to a case-based reasoner. One of the major problems of CBR is finding, among the many cases known to the system, the best partial match for the new case being considered; at the same time, partial matching is one of the tasks in which neural networks shine. Thus, one promising point of SC integration consists in embedding NNs in a case-based reasoner to serve as a partial case matcher. A third hybrid model will combine neural networks with a symbolic processor integrating both rule-based and model-based reasoning. The domain theory or model will be stored in an object-oriented formalism accessible to the rules, which will themselves be represented as objects to allow reflection and declarative metalevel reasoning. This will constitute the framework for symbolic model-based learning which will find its natural complement in statistics-based inductive learning provided by neural networks. Moreover, NNs will also be used as embedded processors to execute certain functions of the symbolic component’s inference
engine. For instance, previous work by [Suttner & Ertel., 1989] consisted in using NNs to discover heuristic evaluation functions in view of controlling search in theorem proving; this line of research can be extended to control rule evaluation and implement conflict resolution in the hybrid model’s rule interpreter. Whenever the application problem requires that the system interact with a user, the connectionist component can be linked to a symbolic postprocessor which will analyze neural net structures and states in view of providing the user with explanations of NN reasoning and behavior. To summarize, the implementation of three hybrid models will serve to test the versatility of the distributed platform for HSC integration—its capacity to support the combination of different symbolic and neural models according to any of the integration modes and degrees of coupling identified in the state of the art. Validating the models: industrial applications Industrial applications will drive both the integration and the validation of the partial hybrids. The first application, conducted by Kratzer Automation, concerns the optimization of the characteristics of a motor/ gearbox combination for turbocharged engines. This task involves: (1) selection of one of 27 characteristic curves for the ignition control; (2) adjustment of parameters for fuel injection curves for the ignition control; (3) selection of gear ratios for optimal efficiency, depending of the envisaged use of the target vehicle; and (4) adjustment of turbocharger parameters (input and bypass pressure, input/output volume ratio, geometric dimensions, etc.). The usual goal of this process is to reduce fuel consumption and air pollution while achieving “optimal” efficiency, i.e., not maximum power but rather a dynamic adaptation of motor and/or gearbox and/or turbocharger characteristics in order to achieve some informally defined kind of (context-dependent) optimum performance. This adjustment is done intuitively at the moment; achieving a reasonably good fit is a time-consuming process, given the huge size of the parameter space and often inadequate information. The motor/gearbox application has been selected for the following reasons. First, it is sufficiently complex to require a combination of new technologies for its solution, yet simple enough to be tractable in the limited time available. Second, the sensor signals are multimodal with varying precision and significance, thus requiring a very complex model description. Third, its practical value is immediately visible since a solution could be applied in practice at once. Fourth, our industrial partner has access to sufficient expert knowledge to make construction of a knowledge base possible. Finally, enough training and test data are available or can be easily obtained for use in the experimentation phase. The second industrial application concerns a prediction problem in a steel rollmill. The application context is as follows: a new skin-pass mill complex with a new galvanizing line was installed by Usinor-Sacilor in 1991 in Sainte-Agathe (France). The overall goal is to tune this tool in order to guarantee uniform and optimum flatness, roughness and mechanical properties of the coil, i.e., the steel sheet that passes through the rolls. Modular computing structures based on different models of the problem have been installed to predict mill presettings in view of reducing operator intervention. The immediate goal is to predict two kinds of forces—the rolling force (both the initial entering force and the force at the maximal rolling speed of 1200m/ min) and the bending force—on the basis of a set of 12 parameters, either desired (e.g. final elongation of the pressed sheet) or measured (e.g., rolling speed, traction force). For the moment, two kinds of models are in use. On the one hand, 10 years of work contracted by the steel industry to mathematiciens, mechanists, thermists, physicists, have resulted in a symbolic model based on physical equations. This model has been used in their plants but have led to 25% errors. On the other hand, INRIA-Lorraine has been testing different neuronal approaches for three years. Currently, a standard classification NN with 12 input units, 2 hidden layers and 2 output units is used and has reduced the error level to 15%. Hopefully, an integrated symbolic-connectionist approach will yield a still lower error rate. The existence of isolated symbolic and connectionist models with known performance measures provides an ideal
baseline from which to appraise the behavior of the hybrid model.
4.
Conclusion
Based on a critical analysis of the state of the art, we have presented a strategy for pushing the frontiers of hybrid symbolic-connectionist integration. The keystone of this strategy is the implementation of a distributed platform for the design, prototyping and assessment of different hybrid architectures. Three hybrid models will be created and validated on optimization and prediction tasks in the steel and automobile industries. It is hoped that in parallel with the attainment of these short-term engineering objectives, a clarification of outstanding research issues will help to lay the groundwork for a theory and methodology of hybrid SC systems.
References Amy, B., Giacometti, A., & Arnaud, G. (1990). Modeles connexionnistes de l’expertise. Actes des journees internationales Neuro-Nimes, 99-119. Becraft, W.R., Lee, P.L. & Newell, R.B. (1991). Integration of Neural Networks and Expert Systems for Process Fault Diagnosis. Proc. 12th International Joint Conference on Artificial Intelligence (pp. 832-837). Bradshaw, G., Fozzard, R., & Ceci, L. (1989). A Connectionist Expert System That Actually Works.In D.S. Touretzky (ed.), Advances in Neural Information Processing Systems 1, 248-255. de Franceso, M. (1994). Functional Networks: A New Computational Framework for the Specification, Simulation and Algebraic Manipulation of Modular Neural Systems. PhD Dissertation, CUI-University of Geneva. Ertel, W., Schumann, J., & Suttner, C. (1989). Learning Heuristics for a Theorem Prover Using Back Propagation. In J. Retti and K. Leidlmair (Eds.), 5. Österreichische Artificial-Intelligenz-Tagung (pp. 87-95). InformatikFachberichte 208, Berlin: Springer-Verlag. Gallant, S. (1988). Connectionist Expert Systems, Communications of the ACM, 31(2):152-169. Giacometti. A. (1992). Modeles hybrides de l’expertise. Doctorate thesis, Grenoble: LIFIA & Paris: TELECOM. Giambasi, N., Bath, R.L., & Touzet, C. (1989). Une approche connexionniste pour l’implication floue. Neuro-Nîmes ‘89, Editions EC2, 143-158. Goodman, R.M., Miller, J.W. & Smyth, P. (1989). An Information Theoretic Approach to Rule-Based Connectionist Expert Systems. In D.S. Touretzky (ed.), Advances in Neural Information Processing, 1: 256-263. Gutknecht, M., & Pfeifer, R. (1990). An Approach to Integrating Expert Systems with Connectionist Networks. AI Communications, 3 (3), 116-127. Hendler, J. (1989). Marker-passing over Microfeatures: Towards a Hybrid Symbolic/Connectionist Model. Cognitive Science 13, 79-106. Iordanova, I., Giacometti, A., Vila, Amy, B., A., Reymond, F., Abaoud, L., Dahou, M., & Rialle, V. (1992). SHADE A hybrid expert system in electromyography. IX International Congress of Electromyography and Clinical Neurophysiology, Jerusalem, Israel. Kasabov, N.K., & Shishkov, S.I. (1993). A Connectionist Production System with Partial Match and its Use for Approximate Reasoning. Connection Science, Vol. 5, Nos. 3 & 4, 275-306. Mani, D.R., & Shastri, L. (1993). Reflexive Reasoning with Multiple Instantiation in a Connectionist Reasoning System with a Type Hierarchy, Connection Science, Vol. 5, Nos. 3 & 4, 205-242. Saito, K., & Nakano, R. (1988). Medical Diagnostic Expert System Based on PDP Model. IEEE International Conference on Neural Networks (Vol I, pp. 255-262). Samad, T. (1992). Hybrid Distributed/Local Connectionist Architectures. In A. Kandel & G. Langholz, Hybrid Architectures for Intelligent Systems. Boca Raton: CRC Press, 199-220. Schneider, W. & Oliver, W.L. (1991). An Instructable Connectionist/Control Architecture: Using Rule-Based Instructions to Accomplish Connectionist Learning in a Human Time Scale. In K. VanLehn (Ed.), Architectures for Intelligence. Hillsdale, NJ: Lawrence Erlbaum Associates, 113-145.
Shavlik, J.W. & Towell, G.G. (1991). An Approach to Combining Explanation-Based and Neural Learning Algorithms. In N.G. Bourbakis (Ed.), Applications of Learning and Planning Methods. Singapore: World Scientific Pub. Smolensky, P. (1990). Tensor product variable binding and the representation of symbolic structures in connectionist systems. Artificial Intelligence, 46. Also in G.E. Hinton (Ed.), Connectionist Symbol Processing. MIT, 1991, 159-216. Sun, R. (1991). Integrating Rules and Connectionism for Robust Reasoning: A Connectionist Architecture with Dual Representation. Report TR-CS-90-160 (PhD Diss.), Brandeis University. Sun, R. (1992). On Variable Binding in Connectionist Networks, Connection Science, Vol. 4, No. 2, 93-124. Suttner, C., & Ertel, W. (1990). Automatic Acquisition of Search Guiding Heuristics. Proceedings of the10th International Conference on Automated Deduction (pp. 470-484). Springer-Verlag, LNAI 449. Tecuci, G. (1993). Plausible Justification Trees: A Framework for Deep and Dynamic Integration of Learning Strategies. Machine Learning, Vol. 11, Nos. 2/3, 237-261. Touretzky, D.S. & G.E. Hinton (1985). Symbols among the Neurons: Details of a Connectionist Inference Architecture. Proceedings of the Ninth International Joint Conference on Artificial Intelligence (pp. 239-243). Touretzky, D.S. & G.E. Hinton (1988). A distributed connectionist production system. Cognitive Science, 12, 423-466.
