Learning and Problem Solving with Multilayer Connectionist Systems September 1986 Charles William Anderson B.S., University of Nebraska M.S., University of Massachusetts Ph.D., University of Massachusetts Directed by: Professor Andrew G. Barto The di#culties of learning in multilayered networks of computational units has limited the use of connectionist systems in complex domains. This dissertation elucidates the issues of learning in a network's hidden units, and reviews methods for addressing these issues that have been developed through the years. Issues of learning in hidden units are shown to be analogous to learning issues for multilayer systems employing symbolic representations.

We consider the requirements of online learning|learning which must be done incrementally and in realtime, with the results of learning available soon after each new example is acquired. Despite the abundance of methods for learning from examples, there are few that can be used e ectively for online learning, e.g., as components of reinforcement learning systems. Most of these few, including radial basis functions, CMACs, Kohonen's self-organizing maps, and those developed in this paper, share the same structure. All expand the original input representation into a higher dimensional representation in an unsupervised way, and then map that representation to the nal answer using a relatively simple supervised learner, such as a perceptron or LMS rule. Such structures learn very rapidly and reliably, but have been thought either to scale poorly or to require extensive domain knowledge. To the contrary, some researchers (Rosenblatt,

This paper extends the currently accepted model of inductive bias by identifying six categories of bias and separates inductive bias from the policy for its selection (the inductive policy). We analyze existing "blas selection " systems, examining the similarities and differences in their inductive policies, and idemify three techniques useful for building inductive policies. We then present a framework for representing and automaticaIly selecting a wide variety of biases and describe experiments with an instantiation of the framework addressing various pragmatic tradeoffs of time, space, accuracy, and the cost oferrors. The experiments show that a common framework can be used to implement policies for a variety of different types of blas selection, such as parameter selection, term selection, and example selection, using similar techniques. The experiments also show that different tradeoffs can be made by the implementation of different policies; for example, from the same data different rule sets can be learned based on different tradeoffs of accuracy versus the cost of erroneous predictions.

A model for learning systems is presented, and representative AI, pattern recognition, and control systems are discussed in terms of its framework. The model details the functional components felt to be essential for any learning system, independent of the techniques used for its construction, and the specific environment in which it operates. These components are erformance element, instance selector, critic, P earning element, blackboard, and world model. Consideration of learning system design leads naturally to the concept of a layered system, each layer operating at a different level of abstraction. Descriptive Terms: adaptation, learning, conceptformatIon, induct ion, performance element, instance selector, critic, learning element, blackboard, world model, multi-layered systems. 1

Many authors have suggested that SP (symbol processing) and CN (connectionist network) models offer radically, or even fundamentally, different paradigms for modeling intelligent behavior (see Schneider, 1987) and the design of intelligent systems. Others have argued that CN models have little to contribute to our efforts to understand intelligence (Fodor & Pylyshyn, 1988). A critical examination of the popular characterizations of SP and CN models suggests that neither of these extreme positions is justified. There are many advantages to be gained by a synthesis of the best of both SP and CN approaches in the design of intelligent systems. The Generalized connectionist networks (GCN) (alternately called generalized neuromorphic systems (GNS)) introduced in this paper provide a framework for such a synthesis.

This memory can take several forms based on the time scales at which such modifications are allowed. Some symbol structures might have the property of determining choice and the order of application of transformations to be applied on other symbol structures. These are essentially the programs. Programs when executed --- typically through the conventional process of compilation and interpretation and eventually --- when they operate on symbols that are linked through grounding to particular effectors --- produce behavior. Working memory holds symbol structures as they are being processed. Long-term memory, generally speaking, is the repository of programs and can be changed by addition, deletion, or modification of symbol structures that it holds. Such a system can compute any Turing-computable function provided it has sufficiently large memory and its primitive set of transformations are adequate for the composition of arbitrarily symbol structures (programs) and the interpreter is capable of interpreting any possible symbol structure. This also means that any particular set of symbolic processes can be carried out by an NANN --- provided it has potentially infinite memory, or finds a way to use its transducers and effectors to use the external physical environment to serve as its memory). 14 Chapter 12 Knowledge in SAI systems is typically embedded in complex symbol structures such as lists (Norvig, 1992), logical databases (Genesereth and Nilsson, 1987), semantic networks (Quillian, 1968), frames (Minsky, 1975), schemas (Arbib, 1972; 1994), and manipulated by (often serial) procedures or inferences (e.g., list processing, application of production rules (Waterman, 1985), or execution of logic programs (Kowalski, 1977) carried out by a central processor that accesse...

1 An understanding of learning -- the process by which a learner acquires and refines a broad range of knowledge and skills -- is central to the enterprise of building truly adaptive, flexible, robust, and creative intelligent systems. Significant theoretical and empirical contributions to the characterization of learning in computational terms have emerged from research in a number of disparate research paradigms. The limitations of individual paradigms and of particular classes of techniques within each paradigm are beginning to be recognized. Converging lines of evidence from multiple sources, both theoretical as well as empirical, suggest that artificial intelligence systems, in order to be able to deal with complex tasks such as recognizing and describing 3-dimensional objects, or communicating in natural language, must be able to effectively utilize a range of learning algorithms operating with an adequate repertoire of representational structures. This paper draws on a broad ran...

Abstract — In this paper, we use Genetic Programming (GP) to define a set of transforms on the space of greyscale images. The motivation is to allow an evolutionary algorithm means of transforming a set of image patterns into a more classifiable form. To this end, we introduce the notion of a Transformbased Evolvable Feature (TEF), a moment value extracted from a GP-transformed image, used in a classification task. Unlike many previous approaches, the TEF allows the whole image space to be searched and augmented. TEFs are instantiated through Cartesian Genetic Programming, and applied to a medical image classification task, that of detecting muscular dystrophy-indicating inclusions in cell images. It is shown that the inclusion of a single TEF allows for significantly superior classification relative to predefined features alone. I.

This memory can take several forms based on the time scales at which such modifications are allowed. Some symbol structures might have the property of determining choice and the order of application of transformations to be applied on other symbol structures. These are essentially the programs. Programs when executed -- typically through the conventional process of compilation and interpretation and eventually -- when they operate on symbols that are linked through grounding to particular effectors -- produce behavior. Working memory holds symbol structures as they are being processed. Long--term memory, generally speaking, is the repository of programs and can be changed by addition, deletion, or modification of symbol structures that it holds. The reader is refered to (Newell, 1990) for a detailed treatment of symbol systems of this sort. Such a symbol system can compute any Turing--computable function provided it has sufficiently large memory and its primitive set of transformations are Beyond Symbolic AI and Connectionist Networks 7 adequate for the composition of arbitrarily symbol structures (programs) and the interpreter is capable of interpreting any possible symbol structure. This also means that any particular set of symbolic processes can be carried out by a CN -- provided it has potentially infinite memory, or finds a way to use its transducers and effectors to use the external physical environment to augment its memory (just as humans have in their use of stone tablets, papyrus, and books through the ages). Knowledge in SAI systems is typically embedded in complex symbol structures such as lists (Norvig, 1992), logical databases (Genesereth and Nilsson, 1987), semantic networks (Quillian, 1968), frames (Minsky, 1975), schemas (Arbib, 1972; 1994), and mani...

This paper presents and describes a pattern recognition program with a relatively simple and general basic structure upon which has been superimposed a rather wide variety of techniques for learning, or self-organization. The program attempts to generalize n-tuple approaches to pattern recognition, in which an n-tuple is a set of individual cells or small pieces of patterns, and each n-tuple is said to characterize an input pattern when these pieces match it, as specified. The program allows n-tuples to match when only some of their parts match, and it allows these parts to match even though they are not precisely positioned (See Uhr, 1969b, for some simple example programs). It further learns, in a variety of ways: It searches for good weights on its characterizers ' implications, byre-weighting as a function of feedback. It generates and discovers new characterizers (and can therefore begin with no characterizers at all), and discards characterizers that prove to be poor (See Uhr and Vossler, 1961, and Prather and Uhr, 1964). It also uses a set of characterizers of characterizers, to search for good parameter values that newlygenerated characterizers should have. A detailed flow-chart-like "precis " description of the program is given, along with an actual listing. It is thus possible to examine exactly what the program does, and how it does it, and therefore to see how a wide variety of learning mechanisms have been implemented in a single pattern recognition program. But because it was coded in a "high-level " patternmatching and list-processing language the program runs too slowly for extensive tests to be practicable. Therefore only a brief listing of output is given, to show that the program, works and begins to learn. Descriptors: Learning, self-organization, induction, discovery, pattern recognition, learning to