进化计算研究生班: 开发了TopCoder马拉松比赛49:megpartty的解决方案。我的小组正在测试不同的域表示法，以及不同的表示法如何影响ga找到正确答案的能力。我们为这个问题编写了自己的代码。

Neuroevolution and Generative and Developmental Systems, Graduate Class: Developed an Othello game board evaluator that was used in the min-max tree of a computer player. The player was set to evaluate one-deep into the game, and trained to play against a greedy computer player that considered corners of vital importance. The training player saw either 3 or 4 deep (I'll need to look at my config files to answer, and they're on a different computer). The goal of the experiment was to compare Novelty Search to traditional, fitness-based search in the Game Board Evaluation domain. Results were relatively inconclusive, unfortunately. While both the novelty search and fitness-based search methods came to a solution (showing that Novelty Search can be used in the Othello domain), it was possible to have a solution to this domain with no hidden nodes. Apparently I didn't create a sufficiently competent trainer if a linear solution was available (and it was possible to have a solution right out of the gates). I believe my implementation of Fitness-based search produced solutions more quickly than my implementation of Novelty search, this time. (this isn't always the case). Either way, I used ANJI, "Another NEAT Java Implementation" for the neural network code, with various modifications. The Othello game I wrote myself.

I used a simple genetic algorithm to optimize the signal to noise ratio of a wave that was represented as a binary string. By flipping the the bits certain ways over several million generations I was able to produce a transform that resulted in a higher signal to noise ratio of that wave. The algorithm could have also been "Simulated Annealing" but was not used in this case. At their core, genetic algorithms are simple, and this was about as simple of a use case that I have seen, so I didn't use a framework for generation creation and selection - only a random seed and the Signal-to-Noise Ratio function at hand.

2004年1月，飞利浦新显示技术公司(Philips New Display Technologies)联系了我，他们正在为有史以来第一款商业电子墨水——索尼Librie——制造电子产品。索尼Librie只在日本上市，比亚马逊Kindle和其他电子墨水在美国和欧洲上市早了好几年。

飞利浦的工程师遇到了一个大问题。在产品上市的几个月前，他们在换页面时仍然会出现重影。问题是产生静电场的200个驱动器。每个驱动器都有一个特定的电压，必须设置在0到1000mv之间。但如果你改变其中一个，就会改变一切。

因此，单独优化每个驱动器的电压是不可能的。可能的值组合的数量以数十亿计，一个特殊的相机大约需要1分钟来评估一个组合。工程师们尝试了许多标准的优化技术，但都没有达到预期的效果。

首席工程师联系了我，因为我之前已经向开源社区发布了一个遗传编程库。他问全科医生/全科医生是否会帮忙，以及我是否能参与其中。我这样做了，在大约一个月的时间里，我们一起工作，我在合成数据上编写和调整GA库，他则将其集成到他们的系统中。然后，有一个周末，他们让它和真人一起直播。

接下来的周一，我收到了他和他们的硬件设计师发来的溢美之词，说没人会相信GA发现的惊人结果。就是这样。同年晚些时候，该产品上市了。

我没有为此得到一分钱，但我有“吹嘘”的权利。他们从一开始就说他们已经超出预算了，所以我在开始工作之前就知道是什么交易。这对于气体的应用是一个很好的例子。：）

几周前，我提出了一个关于SO的解决方案，使用遗传算法来解决图布局的问题。这是一个约束优化问题的例子。

同样在机器学习领域，我用c/c++从头开始实现了一个基于ga的分类规则框架。 我还在一个示例项目中使用了GA来训练人工神经网络(ANN)，而不是使用著名的反向传播算法。

此外，作为我研究生研究的一部分，我已经使用GA来训练隐马尔可夫模型，作为基于em的Baum-Welch算法的额外方法(还是在c/c++中)。

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。

As part of my undergraduate CompSci degree, we were assigned the problem of finding optimal jvm flags for the Jikes research virtual machine. This was evaluated using the Dicappo benchmark suite which returns a time to the console. I wrote a distributed gentic alogirthm that switched these flags to improve the runtime of the benchmark suite, although it took days to run to compensate for hardware jitter affecting the results. The only problem was I didn't properly learn about the compiler theory (which was the intent of the assignment).

我本可以用现有的默认标志来播种初始种群，但有趣的是，算法发现了一个与O3优化级别非常相似的配置(但实际上在许多测试中更快)。

编辑:我还用Python写了我自己的遗传算法框架，只是使用popen命令来运行各种基准测试，尽管如果不是评估作业，我会看看pyEvolve。