Constructivist Artificial Intelligence With Genetic Programming

Learning is an essential attribute of an intelligent system. A proper understanding of the process of learning in terms of knowledge-acquisition, processing and its effective use has been one of the main goals of artificial intelligence (AI). AI, in order to achieve the desired flexibility, performance levels and wide applicability should explore and exploit a variety of learning techniques and representations. Evolutionary algorithms, in recent years, have emerged as powerful learning methods employing task-independent approaches to problem solving and are potential candidates for implementing adaptive computational models. These algorithms, due to their attractive features such as implicit and explicit parallelism, can also be powerful meta-leaming tools for other learning systems such as connectionist networks. These networks, also known as artificial neural networks, offer a paradigm for learning at an individual level that provide an extremely rich landscape of learning mechanisms which AI should exploit. The research proposed in this thesis investigates the role of genetic programming (GP) in connectionism, a learning paradigm that, despite being extremely powerful has a number of limitations. The thesis, by systematically identifying the reasons for these limitations has argued as to why connectionism should be approached with a new perspective in order to realize its true potentialities. With genetic-based designs the key issue has been the encoding strategy. That is, how to encode a neural network within a genotype so as to achieve an optimum network structure and/ or an efficient learning that can best solve a given problem. This in turn raises a number of key questions such as: 1. Is the representation (that is the genotype) that the algorithms employ sufficient to express and explore the vast space of network architectures and learning mechanisms? 2. Is the representation capable of capturing the concepts of hierarchy and modularity that are vital and so naturally employed by humans in problem-solving? 3. Are some representations better in expressing these? If so, how to exploit the strengths that are inherent to these representations? 4. If the aim is really to automate the design process what strategies should be employed so as to minimize the involvement of a designer in the design loop? 5. Is the methodology or the approach able to overcome at least some of the limitations that are commonly seen in connectionist networks? 6. Most importantly, how effective is the approach in problem-solving? These issues are investigated through a novel approach that combines genetic programming and a self-organizing neural network which provides a framework for the simulations. Through the powerful notions of constructivism and micro-macro dynamics the approach provides a way of exploiting the potential features (such as the hierarchy and modularity) that are inherent to the representation that GP employs. By providing a general definition for learning and by imposing a single potential constraint within the representation the approach demonstrates that genetic programming, if used for construction and optimization, could be extremely creative. The method also combines the bottom-up and top-down strategies that are key to evolve ALife-like systems. A comparison with earlier methods is drawn to identify the merits of the proposed approach. A pattern recognition task is considered for illustration. Simulations suggest that genetic- programming can be a powerful meta-leaming tool for implementing useful network architectures and flexible learning mechanisms for self-organizing neural networks while interacting with a given task environment. It appears that it is possible to extend the novel approach further to other types of networks. Finally the role of flexible learning in implementing adaptive AI systems is discussed. A number of potential applications domain is identified