Introduction Genetic programming is a domain-independent problem-solving approach in which computer programs are evolved to solve, or approximately solve, problems. Genetic programming is based on the Darwinian principle of reproduction and survival of the fittest and analogs of naturally occurring genetic operations such as crossover (sexual recombination) and mutation. John Holland's pioneering Adaptation in Natural and Artificial Systems (1975) described how an analog of the evolutionary process can be applied to solving mathematical problems and engineering optimization problems using what is now called the genetic algorithm (GA). The genetic algorithm attempts to find a good (or best) solution to the problem by genetically breeding a population of individuals over a series of generations. In the genetic algorithm, each individual in the population represents a candidate solut

Specification refinement is part of formal program derivation, a method by which software is directly constructed from a provably correct specification. Because program derivation is an intensive manual exercise used for critical software systems, an automated approach would allow it to be viable for many other types of software systems. The goal of this research is to determine if genetic programming (GP) can be used to automate the specification refinement process. The initial steps toward this goal are to show that a well--known proof logic for program derivation can be encoded such that a GP--based system can infer sentences in the logic for proof of a particular sentence. The results are promising and indicate that GP can be useful in aiding program derivation.

Genetic Programming(GP) is a technique that can be used to automatically program computers to perform some required task. The technique is a kind of genetic algorithm in which the rep-resentation is a program parse tree instead of a bit-string and the fitness of each parse trees is evaluated by executing the computer program that it represents. The subject of this thesis is to investigate the use of GP to automatically program multiagent systems. To achieve this goal, we consider the general problems in creating multiagent systems, and show how GP can be used to provide solutions to many of them. Our key contributions are as follows: We show that it possible to evolve multi-agent systems using GP that: exhibit coordinated, coherent behaviour communicate explicitly, and in doing so decide what to communicate and how can resolve conflicts can be integrated into an existing society of agents We also consider the scalability problems involved in the use of GP, both generally and in par-

Genetic programming (GP) is a successful machine learning technique that pro- vides powerful parameterized program primitive constructs using evolution as its search mechanism. However, unlike some machine learning techniques, such as Artificial Neu- ral Networks (ANNs), GP does not have a principled procedure for changing parts of a learned structure based on that structure's past performance. GP is missing a clear, locally optimal update procedure, the equivalent of gradient-descent backpropagation for ANNs. In this article, we introduce a new- mechanism for defining and using performance feedback on program evolution. This "internal reinforcement" principled algorithm is implemented within a new- connectionist representation for evolving parameterized programs, namely "neural programming." We present the algorithms for the generation of credit and blame assignment in the process of learning programs using neural programming and internal reinforcement. The article includes a brief overviewof genetic programming and empirical experiments that demonstrate the increased learning rate obtained by using our principled program evolution approach.

Genetic programming is a relatively new domain-independent method for evolving computer programs to solve problems. This chapter suggests avenues for possible future research on genetic programming, opportunities to extend the technique, and areas for possible practical applications.

There is a fundamental problem with genetic programming as it is currently practiced, the genetic recombination operators that drive the learning process act at random, without regard to how the internal components of the programs to be recombined behaved during training. This research introduces a method of program transformations that is principled, based on the program’s internal behavior, and significantly more likely than random local sampling to improve the transformed programs ’ fitness values. The contribution of our research is a detailed approach by which principled credit-blame assignment can be brought to GP and that credit-blame assignment can be focused to improve that same evolutionary process. This principled credit-blame assignment is done through a new program representation called neural programming and applied through a set of principled processes called, collectively, internal reinforcement in neural programming. This internal reinforcement of evolving programs is presented here as a first step toward the desired gradient descent in program space. There is a fundamental problem with evolutionary computation, and particularly with genetic programming, as it is currently practiced. The problem is that in the space of programs, even if it has been carefully defined so that most or all examined programs

This thesis describes a novel method for representing and automatically generating computer programs in an evolutionary computation context. Abstraction-Based Genetic Programming (ABGP) is a typed Genetic Programming representation system that uses System F, an expressive λ-calculus, to represent the computational components from which the evolved programs are assembled. ABGP is based on the manipulation of closed, independent modules expressing computations with effects that have the ability to affect the whole genotype. These modules are plugged into other modules according to precisely defined rules to form complete computer programs. The use of System F allows the straightforward representation and use of many typical computational structures and behaviors (such as iteration, recursion, lists and trees) in modular form. This is done without introducing additional external symbols in the set of predefined functions and terminals of the system. In fact, programming structures typically included in GP terminal sets, such as if then else, may be removed and represented as abstractions in ABGP for the same problems. ABGP also provides a search space partitioning system based on the structure of the genotypes, similar to the species partitioning system of living organisms and derived from the Curry-Howard isomorphism. This thesis also presents the results obtained by applying this method to a set of problems. Acknowledgments I would not have been able to complete this work without the encouragements, trust and support of my