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

The problem of programming an artificial ant to follow the Santa Fe trail is used as an example program search space. Analysis of shorter solutions shows they have many of the characteristics often ascribed to manually coded programs. Enumeration of a small fraction of the total search space and random sampling characterise it as rugged with many multiple plateaus split by deep valleys and many local and global optima. This suggests it is difficult for hill climbing algorithms. Analysis of the program search space in terms of fixed length schema suggests it is highly deceptive and that for the simplest solutions large building blocks must be assembled before they have above average fitness. In some cases we show solutions cannot be assembled using a fixed representation from small building blocks of above average fitness. These suggest the Ant problem is difficult for Genetic Algorithms. Random sampling of the program search space suggests on average the density of global optima change...

In this paper a new, general and exact schema theory for genetic programming is presented. The theory includes a microscopic schema theorem applicable to crossover operators which replace a subtree in one parent with a subtree from the other parent to produce the offspring. A more macroscopic schema theorem is also provided which is valid for crossover operators in which the probability of selecting any two crossover points in the parents depends only on their size and shape. The theory is based on the notions of Cartesian node reference systems and variable-arity hyperschemata both introduced here for the first time. In the paper we provide examples which show how the theory can be specialised to specific crossover operators and how it can be used to derive an exact definition of effective fitness and a size-evolution equation for GP. 1

In [Luke and Spector 1997] we presented a comprehensive suite of data comparing GP crossover and point mutation over four domains and a wide range of parameter settings. Unfortunately, the results were marred by statistical flaws. This revision of the study eliminates these flaws, with three times as much the data as the original experiments had. Our results again show that crossover does have some advantageover mutation given the right parameter settings (primarily larger population sizes), though the difference between the two surprisingly small. Further, the results are complex, suggesting that the big picture is more complicated than is commonly believed. 1 Introduction The genetic algorithms and evolutionary programming fields have long been at odds over the proper chief operator for generating new populations from previous ones. Genetic algorithms proponents favor crossover, while evolutionary programming's philosophy emphasizes mutation. Most justification for using crossover ...

We present a detailed analysis of the evo- lution of genetic programming ((P) popu- lations using the problem of finding a program which returns the maximum possible value for a given terminal and function set and a depth limit on the program tree (known as the MAX problem). We confirm the basic message of [Gathercole and Ross, 1996] that crossover together with program size restrictions can be responsible for premature convergence to a suboptimal solution. We show that this can happen even when the population retains a high level of variety and show that in many cases evolution from the sub-optimal solution to the solution is possible if sufficient time is allowed. In both cases theoretical models are presented and compared with actual runs.

A few schema theorems for Genetic Programming (GP) have been proposed in the literature in the last few years. Since they consider schema survival and disruption only, they can only provide a lower bound for the expected value of the number of instances of a given schema at the next generation rather than an exact value. This paper presents theoretical results for GP with one-point crossover which overcome this problem. Firstly, we give an exact formulation for the expected number of instances of a schema at the next generation in terms of microscopic quantities. Thanks to this formulation we are then able to provide an improved version of an earlier GP schema theorem in which some (but not all) schema creation events are accounted for. Then, we extend this result to obtain an exact formulation in terms of macroscopic quantities which makes all the mechanisms of schema creation explicit. This theorem allows the exact formulation of the notion of effective fitness in GP and opens the way to future work on GP convergence, population sizing, operator biases, and bloat, to mention only some of the possibilities.

This paper extends recent results in the GP schema theory by formulating a proper exact schema theorem for GP with one-point crossover. This gives an exact expression for the expected number of instances of a schema at the next generation in terms of macroscopic quantities. This result allows the exact formulation of the notion of effective fitness in GP. 1

In this paper, firstly we specialise the exact GP schema theorem for one-point crossover to the case of linear structures of variable length, for example binary strings or programs with arity-1 primitives only. Secondly, we extend this to an exact schema theorem for GP with standard crossover applicable to the case of linear structures. Then we study, both mathematically and numerically, the schema equations and their fixed points for infinite populations for both a constant and a length-related fitness function. This allows us to characterise the bias induced by standard crossover. This is very peculiar. In the case of a constant fitness function, at the fixed-point, structures of any length are present with non-zero probability. However, shorter structures are sampled exponentially much more frequently than longer ones.

In this paper we use the schema theory presented in [20] to better understand the changes in size distribution when using GP with standard crossover and linear structures. Applications of the theory to problems both with and without fitness suggest that standard crossover induces specific biases in the distributions of sizes, with a strong tendency to over sample small structures, and indicate the existence of strong redistribution effects that may be a major force in the early stages of a GP run. We also present two important theoretical results: An exact theory of bloat, and a general theory of how average size changes on flat landscapes with glitches. The latter implies the surprising result that a single program glitch in an otherwise flat fitness landscape is sufficient to drive the average program size of an infinite population, which may have important implications for the control of code growth.

Two main weaknesses of GA and GP schema theorems axe that they provide only information on the expected value of the number of instances of a given schema at the next generation E[m(H,t + 1)], and they can only give a lower bound for such a quantity. This paper presents new theoretical results on GP and GA schemata which laxgely overcome these weaknesses. Firsfly, unlike previous results which concentrated on schema survival and disruption, our results extend to GP recent work on GA theory by Stephens and Waelbroeck, and make the effects and the mechanisms of schema creation explicit. This allows us to give an exact formulation (rather than a lower bound) for the expected number of instances of a schema at the next generation. Thanks to this formulation we are then able to provide in improved version for an eaxlier GP schema theorem in which some schema creation events axe accounted for, thus obtaining a tighter bound for E[m(H, t + 1)]. This bound is a function of the selection probabilities of the schema itself and of a set of lower-order schemata which one-point crossover uses to build instances of the schema. This result supports the existence of building blocks in GP which, however, axe not necessaxily all short, low-order or highly fit. Building on eaxlier work, we show how Stephens and Waelbroeck 's GA results and the new GP results described in the paper can be used to evaluate schema vaxiance, signal-to-noise ratio and, in general, the probability distribution of re(H, t + 1). In addition, we show how the expectation operator can be removed from the schema theorem so as to predict with a known probability whether re(H, t + 1) (rather than Elm(H, t + 1)]) is going to be above a given threshold.