Several years ago I used ga's to optimize asr (automatic speech recognition) grammars for better recognition rates. I started with fairly simple lists of choices (where the ga was testing combinations of possible terms for each slot) and worked my way up to more open and complex grammars. Fitness was determined by measuring separation between terms/sequences under a kind of phonetic distance function. I also experimented with making weakly equivalent variations on a grammar to find one that compiled to a more compact representation (in the end I went with a direct algorithm, and it drastically increased the size of the "language" that we could use in applications).

最近，我将它们用作默认假设，以此来测试由各种算法生成的解决方案的质量。这主要涉及分类和不同类型的拟合问题(即创建一个“规则”，解释审查员对数据集所做的一组选择)。

在工作中，我遇到了这样一个问题:给定M个任务和N个dsp，如何将任务分配给dsp是最好的?“最佳”定义为“最大负载DSP的负载最小化”。有不同类型的任务，不同的任务类型有不同的性能分支，这取决于它们被分配到哪里，所以我将一组工作到dsp的分配编码为“DNA字符串”，然后使用遗传算法来“培育”我所能“培育”的最佳分配字符串。

它运行得相当好(比我之前的方法好得多，之前的方法是评估每个可能的组合……对于非平凡问题的大小，它将需要数年才能完成!)，唯一的问题是无法判断是否已经达到了最优解。你只能决定当前的“最大努力”是否足够好，或者让它运行更长时间，看看它是否可以做得更好。

In 2007-9 I developed some software for reading datamatrix patterns. Often these patterns were difficult to read, being indented into scratched surfaces with all kinds of reflectance properties, fuzzy chemically etched markings and so on. I used a GA to fine tune various parameters of the vision algorithms to give the best results on a database of 300 images having known properties. Parameters were things like downsampling resolution, RANSAC parameters, amount of erosion and dilation, low pass filtering radius, and a few others. Running the optimisation over several days this produced results which were about 20% better than naive values on a test set of images unseen during the optimisation phase.

这个系统完全是从零开始编写的，我没有使用任何其他库。我并不反对使用这些东西，只要它们能提供可靠的结果，但是您必须注意许可兼容性和代码可移植性问题。

我做了一个完整的GA框架，命名为“GALAB”，解决了很多问题:

定位GSM ANTs (BTS)以减少重叠和空白位置。 资源约束项目调度。 进化图景的创造。(Evopic) 旅行推销员问题。 n -皇后和n -颜色问题。 骑士之旅和背包问题。 魔方和数独谜题。 字符串压缩，基于超字符串问题。 二维包装问题。 微型人工生命APP。 鲁比克难题。

我使用遗传算法(以及一些相关技术)来确定风险管理系统的最佳设置，该系统试图阻止淘金者使用偷来的信用卡来购买mmo游戏。该系统将接收数千笔具有“已知”值的交易(欺诈与否)，并找出最佳设置组合，以正确识别欺诈交易，而不会产生太多误报。

We had data on several dozen (boolean) characteristics of a transaction, each of which was given a value and totalled up. If the total was higher than a threshold, the transaction was fraud. The GA would create a large number of random sets of values, evaluate them against a corpus of known data, select the ones that scored the best (on both fraud detection and limiting the number of false positives), then cross breed the best few from each generation to produce a new generation of candidates. After a certain number of generations the best scoring set of values was deemed the winner.

创建用于测试的已知数据语料库是该系统的阿喀琉斯之踵。如果你等待退款，你在试图回应欺诈者时就会落后几个月，所以有人必须手动审查大量交易，以建立数据库，而不必等待太长时间。

这最终确定了绝大多数的欺诈行为，但在最容易欺诈的项目上，这一比例无法低于1%(考虑到90%的交易可能是欺诈，这已经相当不错了)。

我用perl完成了所有这些。在一个相当旧的linux机器上运行一次软件需要1-2个小时(20分钟通过WAN链路加载数据，其余时间用于处理)。任何给定代的大小都受到可用RAM的限制。我会一遍又一遍地运行它，稍微改变参数，寻找一个特别好的结果集。

总而言之，它避免了手动调整数十个欺诈指标的相对值所带来的一些失误，并且始终能够提出比我手动创建的更好的解决方案。AFAIK，它仍然在使用(大约3年后我写了它)。

当你打算粉刷你的房子时，通常很难得到一个确切的颜色组合。通常，你脑海中有一些颜色，但它不是其中一种颜色，供应商向你展示。

昨天，我的GA研究员教授提到了一个发生在德国的真实故事(对不起，我没有更多的参考资料，是的，如果有人要求我可以找到它)。这个家伙(让我们称他为配色员)曾经挨家挨户地帮助人们找到确切的颜色代码(RGB)，这将是客户心目中的衣柜。下面是他的做法:

The color guy used to carry with him a software program which used GA. He used to start with 4 different colors- each coded as a coded Chromosome (whose decoded value would be a RGB value). The consumer picks 1 of the 4 colors (Which is the closest to which he/she has in mind). The program would then assign the maximum fitness to that individual and move onto the next generation using mutation/crossover. The above steps would be repeated till the consumer had found the exact color and then color guy used to tell him the RGB combination!

通过将最大适应度分配给接近消费者想法的颜色，配色员的程序增加了收敛到消费者想法的颜色的机会。我发现它很有趣!

现在我已经得到了一个-1，如果你计划更多的-1，请说明这样做的原因!