Representation and Applications of Evolutionary Computation to Bioinformatics, Optimization, and Games.

在进化计算中，表示是将问题的解存储起来以供算法操作的一种方式。计算智能的一个常见研究类型是调整参数，如人口规模、变异算子的应用率和选择方法。最近在生物信息学、数学建模、博弈论、优化和自动内容生成方面的工作表明，通过修改表示所取得的改进通常比通过参数调整所获得的改进要好。本提案试图在应用问题的背景下发明新的表示，比较不同表示之间的性能，从而检查由不同表示找到的解决方案的特征。最后一个目标是最新奇的。副作用机器是我的团队为DNA分类发明的一个代表。虽然基于副作用机的特征通常表现良好，但将它们与无法由副作用机编码的特征结合使用可以获得更好的性能。我们发明的另一种可进化的表示，称为编织强内核，定位与副作用机器中产生的特征不同的特征，并与之互补。在游戏研究中，我们小组的早期成果之一是，通过改变参与者的表现，可以在单一实验环境中展示几乎任何期望的行为。在自动内容生成（电子游戏）中，我们发现不同的表示生成具有不同美学角色的关卡地图。这意味着拥有多种表现形式的游戏设计师将拥有更大的设计空间去探索。该小组还发明了实际参数优化的新表示，以提高性能或允许枚举最优值。这些不同的主题被探索再现的主题统一起来。这一建议的结果应该是进化计算的设计和选择表示的原则，以及解决问题的新技术，包括从工业合作者那里获得的应用问题。

Representation in evolutionary computation is the way that solutions to a problem are stored for manipulation by the algorithm. A common type of research in computational intelligence is to tune parameters such as population size, rate of application of variation operators, and selection method. Recent work in bioinformatics, mathematical modeling, game theory, optimization, and automatic content generation has shown that the improvement made possibly by modifying representation is often compares well with that obtained by parameter tuning. This proposal seeks to invent new representations in the context of applied problems, compare performance between representations, and so examine the character of solution found by different representations. This last goal is the most novel. Side effect machines are an example of a representation invented by my group for DNA classification. While side-effect machine based features often perform well, even better performance was obtained by using them in combination with features that cannot be encoded by a side effect machine. Another evolvable representation we invented, called a woven strong kernel locates features that are different from, and complimentary to, those arising in side effect machines. In games research, one of our group's early results is that experiments with evolving prisoner's dilemma agents can be made to exhibit almost any desired behavior within a single experimental context by changing the agent's representation. In automatic content generation (for video games) we found that different representations generated level maps with very different aesthetic characters. This suggests that a game designer with a palette of representations available will have a larger design space to explore. The group has also invented new representations for real parameter optimization which improve performance or permit the enumeration of optima. These diverse topics are unified by the theme of exploring representation. The result of this proposal should be principles for the design and selection of representation for evolutionary computation as well as novel techniques for problem solving, including applied problems obtained from industrial collaborators.