Rules of engagement: competitive coevolutionary dynamics in computational systems

Given that evolutionary biologists have considered coevolutionary interactions since the dawn of Darwinism, it is perhaps surprising that coevolution was largely overlooked during the formative years of evolutionary computing. It was not until the early 1990s that Hillis' seminal work thrust coevolution into the spotlight. Upon attempting to evolve fixed-length sorting networks, a problem with a long and competitive history, Hillis found that his standard evolutionary algorithm was producing sub-standard networks. In response, he decided to reciprocally evolve a population of testlists against the sorting network population; thus producing a coevolutionary system. The result was impressive; coevolution not only outperformed evolution, but the best network it discovered was only one comparison longer than the best-known solution. For the first time, a coevolutionary algorithm had been successfully applied to problem-solving. Pre-Hillis, the shortcomings of standard evolutionary algorithms had been understood for some time: whilst defining an adequate fitness function can be as challenging as the problem one is hoping to solve, once achieved, the accumulation of fitness-improving mutations can push a population towards local optima that are difficult to escape. Coevolution offers a solution. By allowing the fitness of each evolving individual to vary (through competition) with other reciprocally evolving individuals, coevolution removes the requirement of a fitness yardstick. In conjunction, the reciprocal adaptations of each individual begin to erode local optima as soon as they appear. However, coevolution is no panacea. As a problem-solving tool, coevolutionary algorithms suffer from some debilitating dynamics, each a result of the relative fitness assessment of individuals. In a single-, or multi-, population competitive system, coevolution may stabilize at a suboptimal equilibrium, or mediocre stable state; analogous to the traditional problem of local optima. Populations may become highly specialized in an unanticipated (and undesirable) manner; potentially resulting in brittle solutions that are fragile to perturbation. The system may cycle; producing dynamics similar to the children's game rock-paper-scissors. Disengagement may occur, whereby one population out-performs another to the extent that individuals cannot be discriminated on the basis of fitness alone; thus removing selection pressure and allowing populations to drift. Finally, coevolution's relative fitness assessment renders traditional visualization techniques (such as the graph of fitness over time) obsolete; thus exacerbating each of the above problems. This thesis attempts to better understand and address the problems of coevolution through the design and analysis of simple coevolutionary models. 'Reduced virulence' - a novel technique specifically designed to tackle disengagement - is developed. Empirical results demonstrate the ability of reduced virulence to combat disengagement both in simple and complex domains, whilst outperforming the only known competitors. Combining reduced virulence with diversity maintenance techniques is also shown to counteract mediocre stability and over-specialization. A critique of the CIAO plot - a visualization technique developed to detect coevolutionary cycling - highlights previously undocumented ambiguities; experimental evidence demonstrates the need for complementary visualizations. Extending the scope of visualization, a first exploration into coevolutionary steering is performed; a technique allowing the user to interact with a coevolutionary system during run-time. Using a simple model incorporating reduced virulence, the coevolutionary steering demonstration highlights the future potential of such tools for both research and education. The role of neutrality in coevolution is discussed in detail. Whilst much emphasis is placed upon neutral networks in the evolutionary computation literature, the nature of coevolutionary neutrality is generally overlooked. Preliminary ideas for modelling coevolutionary neutrality are presented. Finally, whilst this thesis is primarily aimed at a computing audience, strong reference to evolutionary biology is made throughout. Exemplifying potential crossover, the CIAO plot, a tool previously unused in biology, is applied to a simulation of E. Coli, with results con rming empirical observations of real bacteria.